JP6623110B2

JP6623110B2 - Training image selection device, method, and program

Info

Publication number: JP6623110B2
Application number: JP2016076071A
Authority: JP
Inventors: 勇五十嵐; 隆行黒住; 之人渡邉; 豪入江; 周平田良島; 杵渕　哲也; 哲也杵渕
Original assignee: Nippon Telegraph and Telephone Corp; NTT Inc USA
Current assignee: NTT Inc; NTT Inc USA
Priority date: 2016-04-05
Filing date: 2016-04-05
Publication date: 2019-12-18
Anticipated expiration: 2036-04-05
Also published as: JP2017187918A

Description

本発明は、訓練画像選択装置、方法、及びプログラムに係り、特に、被写体を認識することを可能にするための訓練画像選択装置、方法、及びプログラムに関する。 The present invention relates to a training image selection device, method, and program, and more particularly, to a training image selection device, method, and program for enabling recognition of a subject.

カメラ等により空間を撮像したテスト画像に写った被写体の認識や検出を行うための技術として、特許文献１に記載の方法等、テスト画像が予め被写体を撮影した訓練画像群の何れに類似するかを推定する技術が知られている。 As a technique for recognizing and detecting a subject captured in a test image obtained by imaging a space with a camera or the like, which of the training images a test image is similar to, for example, the method described in Patent Literature 1 Techniques for estimating are known.

特開２０１５−２０１１２３号公報JP-A-2015-201123

一方で、特許文献１のような技術では、被写体を様々な方向から撮像した場合でも認識や検出を可能とするために、訓練画像も複数の方向から撮像した物を用意する必要がある。この時、認識させたい撮像角度の範囲、照明条件、時間帯、天候等の撮像条件のバリエーションによっては、用意しなければならない訓練画像の数が膨大になりテスト画像の認識・検出に長い処理時間を要するようになってしまったり、想定したテスト画像の一部を用意した訓練画像では認識できなかったりするという問題がある。 On the other hand, in the technique such as Patent Document 1, it is necessary to prepare training images captured from a plurality of directions in order to enable recognition and detection even when the subject is captured from various directions. At this time, depending on the range of imaging angles to be recognized, lighting conditions, time zones, weather, etc., the number of training images that must be prepared becomes enormous, and the processing time required for recognition and detection of test images is long. Or the training image prepared with a part of the assumed test image cannot be recognized.

本発明は、上記問題点を解決するために成されたものであり、テスト画像に写った被写体の認識に適した訓練画像を選出することができる訓練画像選択装置、方法、及びプログラムを提供することを目的とする。 The present invention has been made to solve the above-described problems, and provides a training image selection device, method, and program capable of selecting a training image suitable for recognition of a subject shown in a test image. For the purpose.

上記目的を達成するために、第１の発明に係る訓練画像選択装置は、空間を撮影したテスト画像からなるテスト画像群と、予め被写体を撮影した訓練画像からなる訓練画像群とに基づいて、前記訓練画像の各々と、前記テスト画像の各々との組み合わせに対し、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表すスコアを算出する被写体認識部と、前記被写体認識部により算出された前記組み合わせの各々の前記スコアに基づいて、前記訓練画像の各々について、前記テスト画像群の認識に寄与する度合いを表す寄与率を算出する寄与率算出部と、前記寄与率算出部により算出された前記訓練画像の各々の前記寄与率に基づいて、前記訓練画像群から前記寄与率が高い訓練画像群を選択する訓練画像選択部と、を含んで構成されている。 In order to achieve the above object, a training image selection device according to a first aspect of the present invention is based on a test image group composed of test images obtained by photographing a space and a training image group comprised of training images obtained by photographing a subject in advance. A subject recognizing unit that calculates a score representing a probability that a common subject is included in the training image and the test image for each combination of the training image and the test image; and the subject recognizing unit A contribution rate calculation unit that calculates a contribution rate that represents the degree of contribution to recognition of the test image group for each of the training images based on the score of each of the combinations calculated by: and the contribution rate calculation unit A training image selection unit that selects, from the training image group, a training image group having a high contribution rate based on the contribution rate of each of the training images calculated by: It has been made.

また、第１の発明に係る訓練画像選択装置において、前記訓練画像の前記寄与率は、前記訓練画像を含む前記組み合わせのうち、算出された前記スコアが予め定めた閾値以上である前記組み合わせの数とするようにしてもよい。 Further, in the training image selection device according to the first invention, the contribution ratio of the training image is the number of the combinations whose calculated score is equal to or greater than a predetermined threshold among the combinations including the training image. You may make it.

また、第１の発明に係る訓練画像選択装置において、前記訓練画像選択部は、前記寄与率算出部により算出された前記訓練画像の各々の前記寄与率に基づいて、前記訓練画像群から未選択の訓練画像を順次選択すると共に、前記テスト画像の各々について、前記選択された訓練画像のうち、前記テスト画像を認識できる訓練画像の数を表す認識回数をカウントし、前記テスト画像の各々の前記認識回数のうち、最小となる前記認識回数が、予め定めた最小認識回数の閾値に達した場合に、前記訓練画像の選択を終了するようにしてもよい。 In the training image selection device according to the first invention, the training image selection unit is not selected from the training image group based on the contribution rate of each of the training images calculated by the contribution rate calculation unit. Each of the test images, and for each of the test images, the number of recognitions representing the number of training images that can recognize the test image among the selected training images is counted, The selection of the training image may be terminated when the minimum number of times of recognition reaches a predetermined minimum number of times of recognition threshold.

また、第１の発明に係る訓練画像選択装置において、前記訓練画像選択部は、前記寄与率算出部により算出された前記訓練画像の各々の前記寄与率に基づいて、前記訓練画像群から未選択の訓練画像を順次選択すると共に、前記テスト画像の各々について、前記選択された訓練画像のうち、前記テスト画像を認識できる訓練画像の数を表す認識回数をカウントし、前記認識回数が最小のテスト画像において、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表す前記スコアが所定の閾値以上である未選択の訓練画像が存在しない場合に、前記訓練画像の選択を終了するようにしてもよい。 In the training image selection device according to the first invention, the training image selection unit is not selected from the training image group based on the contribution rate of each of the training images calculated by the contribution rate calculation unit. The training images are sequentially selected, and for each of the test images, the number of recognitions representing the number of training images that can recognize the test image among the selected training images is counted, and the test with the smallest number of recognitions is performed. In the image, when there is no unselected training image whose score indicating a probability that a common subject is included in the training image and the test image is greater than or equal to a predetermined threshold, the selection of the training image is terminated. You may do it.

また、第１の発明に係る訓練画像選択装置において、前記選択した訓練画像群と、前記テスト画像の各々についての、前記選択された訓練画像群のうち、前記テスト画像を認識できる訓練画像の数を表す認識回数と、前記認識回数が閾値未満となる前記テスト画像と、前記選択した訓練画像群のうち、前記認識回数が閾値未満となる前記テスト画像を認識できる訓練画像の数とのうちの少なくとも一つを提示する結果提示部を更に含むようにしてもよい。 Further, in the training image selection device according to the first invention, the number of training images that can recognize the test image in the selected training image group and the selected training image group for each of the selected training image group. And the number of training images capable of recognizing the test image with the number of recognitions less than the threshold among the selected training image group. You may make it further contain the result presentation part which presents at least one.

また、第１の発明に係る訓練画像選択装置において、前記訓練画像選択部は、前記寄与率算出部により算出された前記訓練画像の各々の前記寄与率に基づいて、前記訓練画像群から未選択の訓練画像を順次選択すると共に、前記テスト画像の各々について、前記選択された訓練画像のうち、前記テスト画像を認識できる訓練画像の数を表す認識回数をカウントし、前記寄与率算出部は、前記訓練画像が順次選択される毎に、前記テスト画像の各々についての、前記選択された訓練画像群のうち、前記テスト画像を認識できる訓練画像の数を表す認識回数に基づいて、前記訓練画像の前記寄与率として、前記訓練画像と、前記認識回数が最小となる前記テスト画像とを組み合わせた前記組み合わせのうち、算出された前記スコアが予め定めた閾値以上である前記組み合わせの数を算出するようにしてもよい。 In the training image selection device according to the first invention, the training image selection unit is not selected from the training image group based on the contribution rate of each of the training images calculated by the contribution rate calculation unit. The training images are sequentially selected, and for each of the test images, among the selected training images, the number of recognitions representing the number of training images capable of recognizing the test image is counted, and the contribution rate calculation unit includes: Each time the training images are sequentially selected, the training images are based on the number of recognitions representing the number of training images capable of recognizing the test images in the selected training image group for each of the test images. As the contribution rate, the calculated score of the combination of the training image and the test image with the smallest number of recognitions is a predetermined threshold. It may be calculated the number of the combinations is greater than or equal.

第２の発明に係る訓練画像選択方法は、被写体認識部が、空間を撮影したテスト画像からなるテスト画像群と、予め被写体を撮影した訓練画像からなる訓練画像群とに基づいて、前記訓練画像の各々と、前記テスト画像の各々との組み合わせに対し、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表すスコアを算出するステップと、寄与率算出部が、前記被写体認識部により算出された前記組み合わせの各々の前記スコアに基づいて、前記訓練画像の各々について、前記テスト画像群の認識に寄与する度合いを表す寄与率を算出するステップと、訓練画像選択部が、前記寄与率算出部により算出された前記訓練画像の各々の前記寄与率に基づいて、前記訓練画像群から前記寄与率が高い訓練画像群を選択するステップと、を含んで実行することを特徴とする。 In the training image selection method according to the second aspect of the present invention, the subject recognition unit uses the training image based on the test image group including the test image obtained by photographing the space and the training image group including the training image obtained by photographing the subject in advance. For each combination of the test image and each of the test images, a step of calculating a score indicating the probability that the subject common to the training image and the test image is included, and a contribution rate calculating unit includes the subject recognition A step of calculating a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images based on the score of each of the combinations calculated by the unit, and a training image selection unit, Selecting a training image group having a high contribution rate from the training image group based on the contribution rate of each of the training images calculated by a contribution rate calculation unit; And executes contain.

第３の発明に係るプログラムは、コンピュータを、上記第１の発明に係る訓練画像選択装置の各部として機能させるためのプログラムである。 A program according to a third invention is a program for causing a computer to function as each part of the training image selection device according to the first invention.

本発明の訓練画像選択装置、方法、及びプログラムによれば、訓練画像の各々と、テスト画像の各々との組み合わせに対し、共通の被写体が含まれる確からしさを表すスコアを算出し、組み合わせの各々のスコアに基づいて、訓練画像の各々について、テスト画像群の認識に寄与する度合いを表す寄与率を算出し、算出された訓練画像の各々の寄与率に基づいて、訓練画像群から寄与率が高い訓練画像群を選択することにより、テスト画像に写った被写体の認識に適した訓練画像を選出することができる、という効果が得られる。 According to the training image selection device, method, and program of the present invention, for each combination of the training image and each of the test images, a score representing the probability that a common subject is included is calculated, and each of the combinations For each of the training images, a contribution rate indicating the degree of contribution to the recognition of the test image group is calculated based on the score of the training image, and the contribution rate is calculated from the training image group based on the calculated contribution rate of each training image. By selecting a high training image group, it is possible to select a training image suitable for recognizing the subject shown in the test image.

本発明の第１の実施の形態に係る訓練画像選択装置の構成を示すブロック図である。It is a block diagram which shows the structure of the training image selection apparatus which concerns on the 1st Embodiment of this invention. 本発明の第１の実施の形態に係る訓練画像選択装置における訓練画像選択処理ルーチンを示すフローチャートである。It is a flowchart which shows the training image selection processing routine in the training image selection apparatus which concerns on the 1st Embodiment of this invention. 本発明の第２の実施の形態に係る訓練画像選択装置の構成を示すブロック図である。It is a block diagram which shows the structure of the training image selection apparatus which concerns on the 2nd Embodiment of this invention. 本発明の第２の実施の形態に係る訓練画像選択装置における訓練画像選択処理ルーチンを示すフローチャートである。It is a flowchart which shows the training image selection process routine in the training image selection apparatus which concerns on the 2nd Embodiment of this invention.

以下、図面を参照して本発明の実施の形態を詳細に説明する。 Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

＜本発明の第１の実施の形態に係る訓練画像選択装置の構成＞ <Configuration of training image selection apparatus according to first embodiment of the present invention>

まず、本発明の第１の実施の形態に係る訓練画像選択装置の構成について説明する。本発明の実施の形態に係る手法によって、複数の被写体毎に予め学習した、当該被写体を撮影した訓練画像群から、訓練画像を選択する。訓練画像群は、別時刻、別の位置・角度で撮影されたテスト画像が、複数の訓練画像群のいずれの被写体を撮影したかを推定するために用いる。なお、下記手順を実施するにあたって、被写体認識部は、訓練画像の選択候補となる画像群を訓練画像としてテスト画像を認識し、各訓練画像に含まれる被写体が各テスト画像にも含まれる確からしさをスコアとして数値化できる状態になっていることを前提とする。これは、例えば特許文献１に記載された方法で実現することができる。また、テスト画像群は、それぞれ同一の被写体を撮影した画像、または同一の被写体として認識したい画像で構成されるものとする。複数の被写体を認識する訓練画像群を構成したい場合は、被写体ごとにテスト画像群を作成し、そのそれぞれについて下記の手法を適用すればよい。 First, the configuration of the training image selection device according to the first embodiment of the present invention will be described. A training image is selected from a group of training images obtained by photographing each subject in advance by using the method according to the embodiment of the present invention. The training image group is used to estimate which subject in the plurality of training image groups is captured by the test images captured at different times and at different positions and angles. In performing the following procedure, the subject recognizing unit recognizes the test image using the image group as the training image selection candidate as the training image, and the probability that the subject included in each training image is also included in each test image. Is assumed to be in a state that can be quantified as a score. This can be realized by the method described in Patent Document 1, for example. In addition, the test image group is configured by an image obtained by photographing the same subject or an image desired to be recognized as the same subject. When it is desired to configure a training image group for recognizing a plurality of subjects, a test image group is created for each subject, and the following method may be applied to each.

図１に示すように、本発明の実施の形態に係る訓練画像選択装置１００は、ＣＰＵと、ＲＡＭと、後述する訓練画像選択処理ルーチンを実行するためのプログラムや各種データを記憶したＲＯＭと、を含むコンピュータで構成することが出来る。この訓練画像選択装置１００は、機能的には図１に示すように入力部１０と、演算部２０と、出力部５０とを備えている。 As shown in FIG. 1, a training image selection device 100 according to an embodiment of the present invention includes a CPU, a RAM, a ROM that stores a program and various data for executing a training image selection processing routine described later, Can be configured with a computer including Functionally, the training image selection apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 50 as shown in FIG.

入力部１０は、空間を撮影したテスト画像からなるテスト画像群と、予め被写体を撮影した訓練画像からなる訓練画像群とを受け付ける。 The input unit 10 receives a test image group composed of test images obtained by photographing a space and a training image group comprised of training images obtained by photographing a subject in advance.

演算部２０は、記憶部２８と、被写体認識部３０と、寄与率算出部３２と、終了条件判定部３６と、訓練画像選択部３４と、結果提示部３８と、を含んで構成されている。なお、各処理部の具体的な処理フローについては後述する訓練画像選択装置の作用の説明において述べる。 The calculation unit 20 includes a storage unit 28, a subject recognition unit 30, a contribution rate calculation unit 32, an end condition determination unit 36, a training image selection unit 34, and a result presentation unit 38. . The specific processing flow of each processing unit will be described in the description of the operation of the training image selection device described later.

被写体認識部３０は、入力部１０で受け付けたテスト画像群と、訓練画像群とに基づいて、訓練画像群における訓練画像の各々と、テスト画像群におけるテスト画像の各々との組み合わせに対し、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表すスコアを算出し、認識結果として記憶部２８に記憶する。 Based on the test image group and the training image group received by the input unit 10, the subject recognition unit 30 performs the combination of each of the training images in the training image group and each of the test images in the test image group. A score representing the probability that a common subject is included in the training image and the test image is calculated and stored in the storage unit 28 as a recognition result.

記憶部２８には、被写体認識部３０で算出された訓練画像の各々と、テスト画像の各々との組み合わせに対するスコアが格納されている。 The storage unit 28 stores a score for a combination of each of the training images calculated by the subject recognition unit 30 and each of the test images.

寄与率算出部３２は、被写体認識部３０により算出された組み合わせの各々のスコアに基づいて、訓練画像の各々について、テスト画像群の認識に寄与する度合いを表す寄与率を算出する。ここで、寄与率は、訓練画像を含む組み合わせのうち、算出されたスコアが予め定めた閾値以上である組み合わせの数とする。また、寄与率算出部３２は、後述する訓練画像選択部３４によって訓練画像が順次選択される毎に、テスト画像の各々についての、選択された訓練画像群のうち、テスト画像を認識できる訓練画像の数を表す認識回数に基づいて、訓練画像の寄与率として、訓練画像と、認識回数が最小となるテスト画像との組み合わせた組み合わせのうち、算出されたスコアが予め定めた閾値以上である組み合わせの数を算出する。 The contribution rate calculation unit 32 calculates a contribution rate representing the degree of contribution to the recognition of the test image group for each of the training images based on the scores of the combinations calculated by the subject recognition unit 30. Here, the contribution rate is the number of combinations whose calculated score is equal to or greater than a predetermined threshold among the combinations including the training image. In addition, the contribution rate calculation unit 32 can recognize a test image in the selected training image group for each test image each time a training image is sequentially selected by a training image selection unit 34 described later. Of the combinations of the training image and the test image that minimizes the number of recognitions, the calculated score is equal to or greater than a predetermined threshold as the contribution rate of the training image based on the number of recognitions representing the number of recognitions The number of is calculated.

訓練画像選択部３４は、寄与率算出部３２により算出された訓練画像の各々の寄与率に基づいて、訓練画像群から寄与率が高い訓練画像群を選択する。本実施の形態では、未選択の訓練画像のうち寄与率が最大のものを選択する操作を反復することにより、訓練画像群から寄与率が高い訓練画像群を選択する。また、訓練画像選択部３４は、寄与率算出部３２により算出された訓練画像の各々の寄与率に基づいて、訓練画像群から未選択の訓練画像を順次選択すると共に、テスト画像の各々について、選択された訓練画像のうち、テスト画像を認識できる訓練画像の数を表す認識回数をカウントする。 The training image selection unit 34 selects a training image group having a high contribution rate from the training image group based on the respective contribution rates of the training images calculated by the contribution rate calculation unit 32. In the present embodiment, a training image group having a high contribution ratio is selected from the training image group by repeating an operation of selecting an unselected training image having the largest contribution ratio. The training image selection unit 34 sequentially selects unselected training images from the training image group based on the contribution rates of the training images calculated by the contribution rate calculation unit 32, and for each of the test images, Of the selected training images, the number of recognitions representing the number of training images that can recognize the test image is counted.

終了条件判定部３６は、所定の条件を満たす場合に、訓練画像選択部３４の訓練画像の選択を終了させる。例えば、テスト画像の各々の認識回数のうち、最小となる認識回数が、予め定めた最小認識回数の閾値に達した場合に、訓練画像の選択を終了させる。また、認識回数が最小のテスト画像において、該テスト画像に共通の被写体が含まれる確からしさを表すスコアが所定の閾値以上である未選択の訓練画像が存在しない場合に、訓練画像の選択を終了させる。 When the predetermined condition is satisfied, the end condition determination unit 36 ends the selection of the training image by the training image selection unit 34. For example, when the minimum number of times of recognition of each test image reaches a predetermined minimum number of times of recognition, the selection of the training image is terminated. In addition, in the test image with the smallest number of recognitions, when there is no unselected training image whose score indicating the probability that the common subject is included in the test image is equal to or greater than a predetermined threshold, the selection of the training image is terminated. Let

結果提示部３８は、訓練画像選択部３４で選択した訓練画像群と、テスト画像の各々についての、選択された訓練画像群のうち、テスト画像を認識できる訓練画像の数を表す認識回数と、認識回数が閾値未満となるテスト画像と、選択した訓練画像群のうち、認識回数が閾値未満となるテスト画像を認識できる訓練画像の数とのうちの少なくとも一つを提示する。 The result presentation unit 38, the training image group selected by the training image selection unit 34, and the number of recognitions representing the number of training images that can recognize the test image among the selected training image group for each of the test images, At least one of the test image with the number of recognitions less than the threshold and the number of training images that can recognize the test image with the number of recognitions less than the threshold among the selected training image group is presented.

＜本発明の実施の第１の形態に係る訓練画像選択装置の作用＞ <Operation of the training image selection device according to the first embodiment of the present invention>

次に、本発明の第１の実施の形態に係る訓練画像選択装置１００の作用について説明する。入力部１０において空間を撮影したテスト画像からなるテスト画像群と、予め被写体を撮影した訓練画像からなる訓練画像群とを受け付けると、訓練画像選択装置１００は、図２に示す訓練画像選択処理ルーチンを実行する。 Next, the operation of the training image selection device 100 according to the first embodiment of the present invention will be described. When the input unit 10 receives a test image group consisting of test images obtained by photographing a space and a training image group consisting of training images obtained by photographing a subject in advance, the training image selection device 100 performs a training image selection processing routine shown in FIG. Execute.

まず、ステップＳ１００では、被写体認識部３０は、入力部１０で受け付けたテスト画像群と、訓練画像群とに基づいて、訓練画像群における訓練画像の各々と、テスト画像群におけるテスト画像の各々との組み合わせに対し、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表すスコアを算出し、認識結果として記憶部２８に保存する。この時、テスト画像群は、被写体認識装置の運用時に入力されることが想定される画像のバリエーションを網羅するように構成されたものであることが望ましい。また、テスト画像群には訓練画像群にも含まれる画像が含まれていてもよい。その場合、認識結果の中に同一の画像同士を比較したスコアが含まれることになるが、そのスコアは削除してもよい。また、本ステップを実行してスコアを算出するのではなく、認識結果は予め生成し保存しておいたものを後続のステップにおいて利用してもよい。 First, in step S100, the subject recognition unit 30 determines each of the training images in the training image group and each of the test images in the test image group based on the test image group received by the input unit 10 and the training image group. A score representing the likelihood that a common subject is included in the training image and the test image is calculated and stored in the storage unit 28 as a recognition result. At this time, it is desirable that the test image group is configured to cover variations of images that are assumed to be input during operation of the subject recognition apparatus. Further, the test image group may include an image included in the training image group. In that case, although the score which compared the same images is contained in the recognition result, you may delete the score. In addition, instead of executing this step and calculating the score, the recognition result generated and stored in advance may be used in the subsequent steps.

次に、ステップＳ１０２では、記憶部２８に保存した認識結果を参照し、被写体認識部３０により算出された組み合わせの各々のスコアに基づいて、訓練画像の各々について、テスト画像群の認識に寄与する度合いを表す寄与率を算出する。各訓練画像が何枚のテスト画像において閾値以上のスコアを出しているか確認する。ここで求める寄与率は、各訓練画像が何枚のテスト画像の認識に寄与しているかを意味する。訓練画像ｊの寄与率をＣ_ｊとすると、寄与率Ｃ_ｊは以下の（１）式で算出することができる、当該訓練画像ｊを含む組み合わせのうち、算出されたスコアが予め定めた閾値以上である組み合わせの数とする。 Next, in step S102, the recognition result stored in the storage unit 28 is referred to, and each of the training images contributes to the recognition of the test image group based on the score of each combination calculated by the subject recognition unit 30. A contribution rate representing the degree is calculated. It is confirmed how many test images each training image has a score equal to or higher than a threshold value. The contribution rate obtained here means how many test images each training image contributes to recognition. When the contribution rate of the training image j is C _j , the contribution rate C _j can be calculated by the following equation (1). Among the combinations including the training image j, the calculated score is equal to or greater than a predetermined threshold value. Is the number of combinations.

ここで、ｓ_ｉｊはテスト画像ｉと訓練画像ｊとの組み合わせに対するスコア、ｔは訓練画像がテスト画像を認識できたと見做すスコアの閾値である。 Here, s _ij is a score for a combination of the test image i and the training image j, and t is a threshold value of a score that assumes that the training image has recognized the test image.

ステップＳ１０４では、終了条件判定部３６は、現時点で選択されていない訓練画像群を対象に、ステップＳ１０２で計算された訓練画像群の寄与率のうちの最大値を確認し、寄与率の最大値が０であった場合には、ステップＳ１１４へ移行し、寄与率の最大値が０以外であった場合にはステップＳ１０６へ移行する。寄与率の最大値が０の場合とは、すなわち、認識回数が最小のテスト画像において、該テスト画像に共通の被写体が含まれる確からしさを表すスコアが所定の閾値以上である未選択の訓練画像が存在しない場合である。 In step S104, the end condition determination unit 36 checks the maximum value of the contribution ratio of the training image group calculated in step S102 for the training image group not selected at the current time, and determines the maximum value of the contribution ratio. Is 0, the process proceeds to step S114, and when the maximum contribution rate is other than 0, the process proceeds to step S106. The case where the maximum value of the contribution rate is 0, that is, in the test image with the smallest number of recognitions, an unselected training image whose score representing the probability of including a common subject in the test image is equal to or greater than a predetermined threshold Is not present.

ステップＳ１０６では、訓練画像選択部３４は、選択されていない訓練画像群のうち寄与率が最大となった訓練画像を選択するとともに、選択した訓練画像で認識できたテスト画像の認識回数を＋１とカウントアップする。ここで、各テスト画像の認識回数の初期値は０である。また、寄与率が最大となった訓練画像が複数存在する場合は、その中から１枚の訓練画像を選択する。選択方法についてはここでは限定しないが、具体的には、訓練画像の各々について、当該訓練画像で認識できたテスト画像群における当該訓練画像に対応するスコアの合計を算出し、合計値が最大の訓練画像を選択する方法や、ｊが最小ないし最大の訓練画像を選択する方法、無作為に選択する方法等が考えられる。 In step S <b> 106, the training image selection unit 34 selects the training image having the maximum contribution rate from the training image group that has not been selected, and sets the number of recognitions of the test image that can be recognized by the selected training image to +1. Count up. Here, the initial value of the number of times each test image is recognized is zero. In addition, when there are a plurality of training images having the maximum contribution rate, one training image is selected from them. Although the selection method is not limited here, specifically, for each training image, the sum of the scores corresponding to the training image in the test image group recognized by the training image is calculated, and the total value is the maximum. A method of selecting a training image, a method of selecting a training image having the smallest or largest j, a method of selecting at random, and the like are conceivable.

ステップＳ１０８では、終了条件判定部３６は、各テスト画像の認識回数の最小値を確認し、最小認識回数が予め定めた最小認識回数の閾値ｒ_ｍａｘ以上であれば、訓練画像の選択を終了してステップＳ１１４へ移行し、閾値ｒ_ｍａｘ未満であればステップＳ１１０へ移行するする。ｒ_ｍａｘは基本的には１であるが、テスト画像のバリエーションの想定が不十分であった場合を考慮し、訓練画像のバリエーションに余裕を持たせたい場合等には２以上としてもよい。 In step S108, the termination condition determination unit 36 confirms the minimum value of the recognition number for each test image, if minimal recognition number minimal recognition count threshold r _max over a predetermined, terminates the selection of the training images Te proceeds to step S114, the process proceeds to step S110 if it is less than the threshold value _{r max.} The r _max is basically 1, but it may be 2 or more when it is desired to give a sufficient margin to the variation of the training image in consideration of the case where the assumption of the variation of the test image is insufficient.

ステップＳ１１０では、終了条件判定部３６は、ここまでで選択された訓練画像の総数を確認し、選択された訓練画像の総数が、予め決定した総数の上限値ｎ_ｍａｘ以上であれば、訓練画像の選択を終了し、ステップＳ１１４へ移行し、上限値ｎ_ｍａｘ未満であればステップＳ１１２へ移行する。上限値ｎ_ｍａｘの値は、被写体認識の運用時に許容される処理時間を充足できる訓練画像数とすればよい。 In step S110, the end condition determination unit 36 confirms the total number of training images selected so far, and if the total number of selected training images is equal to or greater than a predetermined total upper limit value _nmax , the training images are determined. The process proceeds to step S114, and if it is less than the upper limit n _max , the process proceeds to step S112. The value of the upper limit value n _max may be the number of training images that can satisfy the processing time allowed at the time of subject recognition operation.

ステップＳ１１２では、寄与率算出部３２は、認識回数が最小値であるテスト画像群のみに対する訓練画像の各々の寄与率を更新する。この時点での認識回数が最小値であるテスト画像群のみに対する各訓練画像の寄与率を、ステップＳ１０２と同様の方法で算出する。すなわち、以下の（２）式を用いて、訓練画像の寄与率として、訓練画像と、認識回数が最小となるテスト画像とを組み合わせた組み合わせのうち、算出されたスコアが予め定めた閾値ｒ_ｍｉｎ以上である組み合わせの数を算出する。 In step S112, the contribution rate calculation unit 32 updates the contribution rate of each training image to only the test image group having the minimum number of recognitions. The contribution rate of each training image to only the test image group having the minimum number of recognitions at this time is calculated by the same method as in step S102. That is, using the following equation (2), as a contribution ratio of the training image, a calculated score out of a combination of the training image and the test image that minimizes the number of recognitions has a predetermined threshold value r _min. The number of combinations as described above is calculated.

ここで、ｒ_ｉはテスト画像ｉの認識回数、ｒ_ｍｉｎは各テスト画像の認識回数の最小値である。本ステップの後は、ステップＳ１０４に戻る。なお、本ステップの実行を省略してもよい。 Here, r _i is the number of times of recognition of the test image i, and r _min is the minimum value of the number of times of recognition of each test image. After this step, the process returns to step S104. Note that the execution of this step may be omitted.

最後に、ステップＳ１１４では、結果提示部３８は、最終的に選択された訓練画像群と、テスト画像の各々についての、選択された訓練画像群のうち、テスト画像を認識できる訓練画像の数を表す認識回数と、認識回数が所定の予め定めた閾値ｒ_ｑ未満であるテスト画像及びその認識回数の組とを出力部５０に提示し、訓練画像選択処理ルーチンを終了する。また、出力のバリエーションとして、選択した訓練画像群のうち、認識回数が閾値ｒ_ｑ未満となるテスト画像を認識できる訓練画像の数を出力するようにしてもよい。ここで、閾値ｒ_ｑの決定方法はここでは限定しないが、例えばｒ_ｍａｘとすれば、選択された訓練画像群で所望の認識回数を満足できないテスト画像のリストを得る事ができる。また、出力の際に、選択された各訓練画像の寄与率を併せて提示してもよい。どのようなテスト画像の組み合わせに対する寄与率を提示するかは限定しないが、具体的には、全テスト画像に対する寄与率、選択された訓練画像が選択された時点で認識回数が最小であったテスト画像に対する寄与率、テスト画像１枚１枚に対する寄与率、それらの組み合わせ等が考えられる。また、提示方法はディスプレイへの表示、ファイルへの書き出し等様々な方法が考えられるが、ここでは限定しない。提示に用いる項目についても、訓練画像の通し番号、訓練画像に付与した名称、それらの組み合わせ等、様々なものが考えるが、ここでは限定しない。 Finally, in step S114, the result presentation unit 38 determines the number of training images that can recognize the test image from the selected training image group for each of the finally selected training image group and the test image. The number of recognitions to be represented, a test image whose number of recognitions is less than a predetermined threshold r _q and a set of the number of recognitions are presented to the output unit 50, and the training image selection processing routine is terminated. As a variation of the output, the number of training images that can recognize a test image whose number of recognitions is less than the threshold r _q among the selected training image group may be output. Here, the method for determining the threshold r _q is not limited here. For example, if r _max is set, a list of test images that cannot satisfy the desired number of recognitions in the selected training image group can be obtained. Moreover, you may show together the contribution rate of each selected training image in the case of an output. There is no limitation on the test image combination to which the contribution rate is presented, but specifically, the contribution rate to all test images and the test with the smallest number of recognitions when the selected training image was selected A contribution rate to the image, a contribution rate to each test image, a combination thereof, and the like are conceivable. Various presentation methods such as display on a display and writing to a file can be considered, but the presenting method is not limited here. Various items such as a serial number of the training image, a name assigned to the training image, a combination thereof, and the like are considered for the items used for presentation, but are not limited here.

以上説明したように、本発明の第１の実施の形態に係る訓練画像選択装置によれば、訓練画像の各々と、テスト画像の各々との組み合わせに対し、該訓練画像に含まれる被写体が該テスト画像にも含まれる確からしさを表すスコアを算出し、組み合わせの各々のスコアに基づいて、訓練画像の各々について、テスト画像群の認識に寄与する度合いを表す寄与率を算出し、算出された訓練画像の各々の寄与率に基づいて、未選択の訓練画像のうち寄与率が最大のものを選択する操作を反復することにより、訓練画像群から寄与率が高い訓練画像群を選択することにより、テスト画像に写った被写体の認識に適した訓練画像を選出することができる。 As described above, according to the training image selection device according to the first embodiment of the present invention, the subject included in the training image corresponds to the combination of each training image and each test image. A score representing the probability included in the test image is calculated, and a contribution rate representing the degree of contribution to the recognition of the test image group is calculated for each of the training images based on the score of each combination. By selecting a training image group with a high contribution rate from the training image group by repeating the operation of selecting the one with the largest contribution rate among unselected training images based on the respective contribution rates of the training images. Therefore, it is possible to select a training image suitable for recognizing the subject shown in the test image.

＜本発明の第２の実施の形態に係る訓練画像選択装置の構成＞ <Configuration of Training Image Selection Device According to Second Embodiment of the Present Invention>

まず、本発明の第２の実施の形態に係る訓練画像選択装置の構成について説明する。なお、第１の実施の形態と同様となる箇所については同一符号を付して説明を省略する。 First, the configuration of the training image selection device according to the second embodiment of the present invention will be described. In addition, the same code | symbol is attached | subjected about the location similar to 1st Embodiment, and description is abbreviate | omitted.

第２の実施の形態では、テスト画像を認識できない訓練画像を除外する操作を反復することにより、訓練画像群から寄与率が高い訓練画像群を選択する。 In the second embodiment, a training image group having a high contribution rate is selected from the training image group by repeating the operation of excluding the training image that cannot recognize the test image.

図３に示すように、本発明の第２の実施の形態に係る訓練画像選択装置２００は、ＣＰＵと、ＲＡＭと、後述する訓練画像選択処理ルーチンを実行するためのプログラムや各種データを記憶したＲＯＭと、を含むコンピュータで構成することが出来る。この訓練画像選択装置２００は、機能的には図３に示すように入力部１０と、演算部２２０と、出力部５０とを備えている。 As shown in FIG. 3, the training image selection device 200 according to the second embodiment of the present invention stores a CPU, a RAM, a program for executing a training image selection processing routine described later, and various data. It can be composed of a computer including a ROM. Functionally, the training image selection apparatus 200 includes an input unit 10, a calculation unit 220, and an output unit 50 as shown in FIG.

演算部２２０は、記憶部２２８と、被写体認識部２３０と、寄与率算出部３２と、訓練画像選択部２３４と、終了条件判定部２３６と、結果提示部３８と、を含んで構成されている。 The calculation unit 220 includes a storage unit 228, a subject recognition unit 230, a contribution rate calculation unit 32, a training image selection unit 234, an end condition determination unit 236, and a result presentation unit 38. .

被写体認識部２３０は、第１の実施の形態と同様に、入力部１０で受け付けたテスト画像群と、訓練画像群とに基づいて、訓練画像群における訓練画像の各々と、テスト画像群におけるテスト画像の各々との組み合わせに対し、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表すスコアを算出し、認識結果として記憶部２２８に記憶する。また、認識結果を用いて、各テスト画像について、当該テスト画像が訓練画像のうち何枚で認識できるかを表す認識回数を初期化する。 Similar to the first embodiment, the subject recognition unit 230 determines each of the training images in the training image group and the test in the test image group based on the test image group received by the input unit 10 and the training image group. For each combination with the image, a score representing the probability that a common subject is included in the training image and the test image is calculated and stored in the storage unit 228 as a recognition result. In addition, for each test image, the number of times of recognition representing how many of the test images can be recognized among the training images is initialized using the recognition result.

記憶部２２８には、被写体認識部２３０で算出された訓練画像の各々と、テスト画像の各々との組み合わせに対するスコア、及び認識回数が格納されている。 The storage unit 228 stores a score for the combination of each of the training images calculated by the subject recognition unit 230 and each of the test images, and the number of times of recognition.

訓練画像選択部２３４は、寄与率算出部３２により算出された訓練画像の各々の寄与率に基づいて、未除外の訓練画像のうち寄与率が最小のものを除外する操作を反復することにより、訓練画像群から寄与率が高い訓練画像群を選択する。 The training image selection unit 234 repeats the operation of excluding the one with the smallest contribution rate among the unexcluded training images based on the contribution rate of each training image calculated by the contribution rate calculation unit 32. A training image group having a high contribution rate is selected from the training image group.

終了条件判定部２３６は、所定の条件を満たす場合に、訓練画像選択部２３４の訓練画像の除外を終了させる。例えば、テスト画像の各々の認識回数のうち、最小となる認識回数が、予め定めた最小認識回数の閾値未満になった場合に、訓練画像の除外を終了させる。 When the predetermined condition is satisfied, the end condition determination unit 236 ends the training image selection unit 234 excluding the training image. For example, when the minimum number of times of recognition of each test image is less than a predetermined minimum number of times of recognition, the training image exclusion is terminated.

＜本発明の実施の第２の形態に係る訓練画像選択装置の作用＞ <Operation of the training image selection device according to the second embodiment of the present invention>

次に、本発明の第２の実施の形態に係る訓練画像選択装置２００の作用について説明する。入力部１０において空間を撮影したテスト画像からなるテスト画像群と、予め被写体を撮影した訓練画像からなる訓練画像群とを受け付けると、訓練画像選択装置２００は、図４に示す訓練画像選択処理ルーチンを実行する。 Next, the operation of the training image selection device 200 according to the second embodiment of the present invention will be described. When the input unit 10 receives a test image group consisting of test images obtained by photographing a space and a training image group consisting of training images obtained by photographing a subject in advance, the training image selection device 200 performs a training image selection processing routine shown in FIG. Execute.

まず、ステップＳ２００では、被写体認識部２３０は、入力部１０で受け付けたテスト画像群と、訓練画像群とに基づいて、訓練画像群における訓練画像の各々と、テスト画像群におけるテスト画像の各々との組み合わせに対し、該訓練画像と該テスト画像とに共通の被写体が含まれる確からしさを表すスコアを算出し、認識結果として記憶部２２８に保存する。この時、テスト画像群は、被写体認識装置の運用時に入力されることが想定される画像のバリエーションを網羅するように構成されたものであることが望ましい。また、テスト画像群には訓練画像群にも含まれる画像が含まれていてもよい。その場合、認識結果の中に同一の画像同士を比較したスコアが含まれることになるが、そのスコアは削除してもよい。また、認識結果は予め生成し保存しておいた物を利用してもよい。以上で準備された認識結果を用いて、被写体認識部２３０は、各テスト画像について、当該テスト画像が訓練画像のうち何枚で認識できるかを表す認識回数を初期化する。すなわち、以下（３）式を用いて各テスト画像ｉの認識回数ｒ_ｉを初期化して認識回数を求める。 First, in step S200, the subject recognition unit 230 determines each of the training images in the training image group and each of the test images in the test image group based on the test image group received by the input unit 10 and the training image group. A score representing the likelihood that a common subject is included in the training image and the test image is calculated and stored in the storage unit 228 as a recognition result. At this time, it is desirable that the test image group is configured to cover variations of images that are assumed to be input during operation of the subject recognition apparatus. Further, the test image group may include an image included in the training image group. In that case, although the score which compared the same images is contained in the recognition result, you may delete the score. The recognition result may be generated and stored in advance. Using the recognition result prepared as described above, the subject recognition unit 230 initializes the number of times of recognition indicating how many test images the test image can be recognized for each test image. That is, the following (3) obtaining the recognition number to initialize the recognition number r _i of each test image i using equation.

次に、ステップＳ２０２では、寄与率算出部３２は、記憶部２２８に保存した認識結果及び認識回数を参照し、各訓練画像が何枚のテスト画像において閾値以上のスコアを出しているか確認し、各訓練画像が何枚のテスト画像の認識に寄与しているかを意味する寄与率を求める。訓練画像ｊの寄与率をＣ_ｊとすると、Ｃ_ｊは上記（１）式で算出することができる。この時、寄与率の算出に上記（１）式の代わりに上記（２）式を用い、この時点での認識回数が最小であるテスト画像群のみに対する各訓練画像の寄与率を求めてもよい。ここで、ｒ_ｍｉｎは各テスト画像の認識回数の最小値である。その後、寄与率を上記（１）式で算出した場合は、寄与率が０である訓練画像を訓練画像群から除外する処理を行ってもよい。寄与率が０である訓練画像は想定されるテスト画像の認識に寄与しないため、通常は除外する方が良いが、テスト画像のバリエーションの想定が不十分で寄与率が０である訓練画像によってのみ認識できるテスト画像が運用時に入力される可能性を考慮し、除外せずに残してもよい。 Next, in step S202, the contribution rate calculation unit 32 refers to the recognition result and the number of recognitions stored in the storage unit 228, and confirms how many test images each score has a score equal to or higher than the threshold value. A contribution rate indicating how many test images contribute to the recognition of each test image is obtained. Assuming that the contribution rate of the training image j is C _j , C _j can be calculated by the above equation (1). At this time, instead of the above equation (1), the above equation (2) may be used to calculate the contribution rate, and the contribution rate of each training image for only the test image group with the smallest number of recognitions at this time may be obtained. . Here, r _min is the minimum value of the number of times each test image is recognized. Thereafter, when the contribution rate is calculated by the above equation (1), a process of excluding a training image having a contribution rate of 0 from the training image group may be performed. Training images with a contribution rate of 0 do not contribute to the recognition of the assumed test image, so it is usually better to exclude them, but only with training images with a poor contribution assumption of the test image and a contribution rate of 0 Considering the possibility that a recognizable test image is input during operation, it may be left without being excluded.

ステップＳ２０４では、終了条件判定部２３６は、ステップＳ２０２で求められた各テスト画像の認識回数の最小値を確認し、認識回数の最小値が予め決定した値ｒ_ｍａｘ未満であれば、訓練画像の除外を終了してステップＳ２１２に移行し、認識回数の最小値がｒ_ｍａｘ以上であればステップＳ２０６へ移行する。ｒ_ｍａｘは基本的には１であるが、テスト画像のバリエーションの想定が不十分であった場合を考慮し、訓練画像のバリエーションに余裕を持たせたい場合等には２以上としてもよい。 In step S204, the termination condition determination unit 236 checks the minimum value of the recognition number for each test image obtained in step S202, the minimum value of the recognition number is less than the value r _max in advance determined, the training images the process proceeds to step S212 to end the exclusion, the minimum value of the recognition number is shifted to step S206 if more _{r max.} The r _max is basically 1, but it may be 2 or more when it is desired to give a sufficient margin to the variation of the training image in consideration of the case where the assumption of the variation of the test image is insufficient.

ステップＳ２０６では、終了条件判定部２３６は、現時点で除外されていない訓練画像の総数を確認し、除外されていない訓練画像の総数が予め決定した値ｎ_ｍａｘ以下であれば、訓練画像の除外を終了してステップＳ２１２に移行し、除外されていない訓練画像の総数がｎ_ｍａｘを超えていればステップＳ２０８に移行する。ｎ_ｍａｘの値は、被写体認識装置の運用時に許容される処理時間を充足できる訓練画像数とすればよい。 In step S206, the end condition determination unit 236 checks the total number of training images that are not excluded at the current time. If the total number of training images that are not excluded is equal to or less than a predetermined value n _max , the end of the training images is excluded. When finished, the process proceeds to step S212. If the total number of training images not excluded exceeds n _max , the process proceeds to step S208. The value of n _max may be the number of training images that can satisfy the processing time allowed when the subject recognition apparatus is operated.

ステップＳ２０８では、訓練画像選択部２３４は、除外されていない訓練画像のうち寄与率が最小である訓練画像を除外するとともに、除外した訓練画像で認識できたテスト画像の認識回数を−１とカウントダウンする。また、寄与率が最小となった訓練画像が複数存在する場合は、その中から１枚の訓練画像を除外する。除外する訓練画像の選択方法についてはここでは限定しないが、具体的には、各訓練画像について各テスト画像群におけるその訓練画像に対応するスコアの合計を算出し、合計値が最小の訓練画像を選択する方法や、訓練画像ｊのスコアが最小ないし最大の訓練画像を選択する方法、無作為に選択する方法等が考えられる。 In step S208, the training image selection unit 234 excludes the training image having the smallest contribution rate from the training images that are not excluded, and counts down the number of times of recognition of the test image that can be recognized by the excluded training image as -1. To do. In addition, when there are a plurality of training images having the minimum contribution rate, one training image is excluded from them. Although the method for selecting the training image to be excluded is not limited here, specifically, for each training image, the total score corresponding to the training image in each test image group is calculated, and the training image having the minimum total value is calculated. A method of selecting, a method of selecting a training image having the minimum or maximum score of the training image j, a method of selecting at random, and the like can be considered.

ステップＳ２１０では、寄与率算出部３２は、この時点での認識回数が最小であるテスト画像群のみに対する各訓練画像の寄与率を、ステップＳ２０２と同様の方法で算出する。すなわち、上記（２）式を用いて各訓練画像の寄与率を更新する。本ステップの後は、ステップＳ２０４に戻る。なお、本ステップの実行を省略するようにしてもよい。 In step S210, the contribution rate calculation unit 32 calculates the contribution rate of each training image only to the test image group with the smallest number of recognitions at this time by the same method as in step S202. That is, the contribution rate of each training image is updated using the above equation (2). After this step, the process returns to step S204. Note that the execution of this step may be omitted.

最後に、ステップＳ２１２では、結果提示部３８は、最終的に除外されなかった訓練画像群と、テスト画像の各々についての、除外されなかった訓練画像群のうち、テスト画像を認識できる訓練画像の数を表す認識回数と、認識回数が所定の予定めた閾値ｒ_ｑ未満であるテスト画像とその認識回数の組を提示し、訓練画像選択処理ルーチンを終了する。この時、ステップＳ２０４において訓練画像の選択を終了した場合であって、ステップＳ２０８を一度でも既に実行している場合には、最後に除外した訓練画像の除外を取り消す。ｒ_ｑの決定方法はここでは限定しないが、例えばｒ_ｍａｘとすれば、選択された訓練画像群で所望の認識回数を満足できないテスト画像のリストを得る事ができる。また、この時、選択された各訓練画像の寄与率を併せて提示してもよい。どのようなテスト画像の組み合わせに対する寄与率を提示するかは限定しないが、具体的には、全テスト画像に対する訓練画像の寄与率、その訓練画像が選択された時点で認識回数が最小であったテスト画像に対する訓練画像の寄与率、テスト画像１枚１枚に対する訓練画像の寄与率、それらの組み合わせ等が考えられる。また、提示方法はディスプレイへの表示、ファイルへの書き出し等様々な方法が考えられるが、ここでは限定しない。提示に用いる項目についても、訓練画像の通し番号、訓練画像に付与した名称、それらの組み合わせ等、様々なものが考えるが、ここでは限定しない。 Finally, in step S212, the result presentation unit 38 selects a training image that can recognize the test image from the training image group that has not been excluded and the training image group that has not been excluded from each of the test images. A set of the number of recognitions representing the number, a test image whose number of recognitions is less than a predetermined threshold r _q and the number of recognitions is presented, and the training image selection processing routine is terminated. At this time, if the selection of the training image is completed in step S204, and step S208 has already been executed even once, the last excluded training image is canceled. The method for determining r _q is not limited here. For example, if r _max is set, a list of test images that cannot satisfy the desired number of recognitions in the selected training image group can be obtained. At this time, the contribution rate of each selected training image may be presented together. There is no limitation on the test image combination contribution rate to be presented, but specifically, the contribution rate of the training image to all the test images, and the number of recognitions was minimal when the training image was selected. The contribution rate of the training image to the test image, the contribution rate of the training image to each test image, a combination thereof, and the like are conceivable. Various presentation methods such as display on a display and writing to a file can be considered, but the presenting method is not limited here. Various items such as a serial number of the training image, a name assigned to the training image, a combination thereof, and the like are considered for the items used for presentation, but are not limited here.

以上説明したように、本発明の第２の実施の形態に係る訓練画像選択装置によれば、訓練画像の各々と、テスト画像の各々との組み合わせに対し、該訓練画像に含まれる被写体が該テスト画像にも含まれる確からしさを表すスコアを算出し、組み合わせの各々のスコアに基づいて、訓練画像の各々について、テスト画像群の認識に寄与する度合いを表す寄与率を算出し、算出された訓練画像の各々の寄与率に基づいて、未除外の訓練画像のうち寄与率が最小のものを除外する操作を反復することにより、訓練画像群から寄与率が高い訓練画像群を選択することにより、テスト画像に写った被写体の認識に適した訓練画像を選出することができる。 As described above, according to the training image selection device according to the second embodiment of the present invention, the subject included in the training image is the combination of each training image and each test image. A score representing the probability included in the test image is calculated, and a contribution rate representing the degree of contribution to the recognition of the test image group is calculated for each of the training images based on the score of each combination. By selecting a training image group with a high contribution rate from the training image group by repeating the operation of excluding the one with the smallest contribution rate among the unexcluded training images based on the contribution rate of each training image. Therefore, it is possible to select a training image suitable for recognizing the subject shown in the test image.

また上述した第１及び第２の実施の形態によれば、被写体認識装置の運用時に入力が想定されるテスト画像群を認識・検出するために最適な訓練画像を選出するとともに、選出した訓練画像で認識できないテスト画像を提示し認識率向上のために必要な訓練画像追加の指針を示すことができる。 Further, according to the first and second embodiments described above, an optimal training image is selected for recognizing and detecting a test image group assumed to be input when the subject recognition apparatus is operated, and the selected training image is selected. It is possible to present test images that cannot be recognized by, and to provide guidelines for adding training images necessary to improve the recognition rate.

また、撮像条件のバリエーションが多い場合でも、被写体認識に貢献が小さい訓練画像を排除し、貢献が大きい訓練画像に絞ることで、被写体認識率を下げることなく効率的な訓練画像の集合を選出し、認識・検出に要する処理時間を短く抑える事が可能となる。また、既存の訓練画像では認識できない、あるいは認識が困難なテスト画像を特定するとともに、それらを認識できる訓練画像を追加する事で、認識率を高める事が可能となる。 In addition, even when there are many variations in imaging conditions, an effective set of training images can be selected without reducing the subject recognition rate by eliminating training images that have a small contribution to subject recognition and focusing on training images that have a large contribution. Therefore, the processing time required for recognition / detection can be shortened. Further, by identifying test images that cannot be recognized by existing training images or difficult to recognize, and adding training images that can recognize them, the recognition rate can be increased.

なお、本発明は、上述した実施の形態に限定されるものではなく、この発明の要旨を逸脱しない範囲内で様々な変形や応用が可能であり、構成要素の追加、省略、置換、その他の変更を行ってもよい。 The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the gist of the present invention. Changes may be made.

例えば、上記において第１及び第２の実施の形態を説明したが、これらの実施の形態における訓練画像選択装置をコンピュータで実現するようにしてもよい。その場合、この機能を実現するためのプログラムをコンピュータ読み取り可能な記録媒体に記録して、この記録媒体に記録されたプログラムをコンピュータシステムに読み込ませ、実行することによって実現してもよい。なお、ここでいう「コンピュータシステム」とは、ＯＳや周辺機器等のハードウェアを含むものとする。また、「コンピュータ読み取り可能な記録媒体」とは、フレキシブルディスク、光磁気ディスク、ＲＯＭ、ＣＤ−ＲＯＭ等の可搬媒体、コンピュータシステムに内蔵されるハードディスク等の記憶装置のことをいう。さらに「コンピュータ読み取り可能な記録媒体」とは、インターネット等のネットワークや電話回線等の通信回線を介してプログラムを送信する場合の通信線のように、短時間の間、動的にプログラムを保持するもの、その場合のサーバやクライアントとなるコンピュータシステム内部の揮発性メモリのように、一定時間プログラムを保持しているものも含んでもよい。また上記プログラムは、前述した機能の一部を実現するためのものであっても良く、さらに前述した機能をコンピュータシステムにすでに記録されているプログラムとの組み合わせで実現できるものであってもよく、ＰＬＤ（Programmable Logic Device）やＦＰＧＡ（Field Programmable Gate Array）等のハードウェアを用いて実現されるものであってもよい。 For example, although the first and second embodiments have been described above, the training image selection device in these embodiments may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included and a program held for a certain period of time. Further, the program may be for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in the computer system. It may be realized using hardware such as PLD (Programmable Logic Device) or FPGA (Field Programmable Gate Array).

また、形態は本発明の例示に過ぎず、本発明が上記実施の形態に限定されるものではないことは明らかである。したがって、本発明の技術思想及び範囲を逸脱しない範囲で構成要素の追加、省略、置換、その他の変更を行ってもよい。 Further, the form is merely an example of the present invention, and it is obvious that the present invention is not limited to the above-described embodiment. Therefore, additions, omissions, substitutions, and other modifications of the components may be made without departing from the technical idea and scope of the present invention.

１０入力部
２０、２２０演算部
２８、２２８記憶部
３０、２３０被写体認識部
３２寄与率算出部
３４、２３４訓練画像選択部
３６、２３６終了条件判定部
３８結果提示部
５０出力部
１００、２００訓練画像選択装置 10 Input unit 20, 220 Calculation unit 28, 228 Storage unit 30, 230 Subject recognition unit 32 Contribution rate calculation unit 34, 234 Training image selection unit 36, 236 End condition determination unit 38 Result presentation unit 50 Output unit 100, 200 Training image Selection device

Claims

Based on a test image group composed of test images obtained by photographing a space and a training image group comprised of training images obtained by photographing a subject in advance, the training is performed for each combination of the training images and each of the test images. A subject recognizing unit that calculates a score representing the probability that a common subject is included in the image and the test image;
A contribution rate calculation unit that calculates a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images, based on the score of each of the combinations calculated by the subject recognition unit;
Based on the contribution rate of each of the training images calculated by the contribution rate calculation unit, an unselected training image is sequentially selected from the training image group, and the selected training is performed for each of the test images. Among the images, the number of recognitions representing the number of training images that can recognize the test image is counted, and the minimum number of times of recognition among the number of times of recognition of each of the test images is a threshold of a predetermined minimum number of times of recognition. A training image selection unit that terminates the selection of the training image when reaching
A training image selection device including:

Based on a test image group composed of test images obtained by photographing a space and a training image group comprised of training images obtained by photographing a subject in advance, the training is performed for each combination of the training images and each of the test images. A subject recognizing unit that calculates a score representing the probability that a common subject is included in the image and the test image;
A contribution rate calculation unit that calculates a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images, based on the score of each of the combinations calculated by the subject recognition unit;
Based on the contribution rate of each of the training images calculated by the contribution rate calculation unit, an unselected training image is sequentially selected from the training image group, and the selected training is performed for each of the test images. Among the images, the number of recognitions representing the number of training images that can recognize the test image is counted, and in the test image with the smallest number of recognitions, the probability that the subject common to the training image and the test image is included is determined. A training image selection unit that terminates the selection of the training image when there is no unselected training image in which the score to be expressed is equal to or greater than a predetermined threshold ;
A training image selection device including:

Based on a test image group composed of test images obtained by photographing a space and a training image group comprised of training images obtained by photographing a subject in advance, the training is performed for each combination of the training images and each of the test images. A subject recognizing unit that calculates a score representing the probability that a common subject is included in the image and the test image;
A contribution rate calculation unit that calculates a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images, based on the score of each of the combinations calculated by the subject recognition unit;
Based on the contribution rate of each of the training images calculated by the contribution rate calculation unit, an unselected training image is sequentially selected from the training image group, and the selected training is performed for each of the test images. A training image selection unit that counts the number of recognitions representing the number of training images that can recognize the test image among the images ,
Each time the training images are sequentially selected, the contribution ratio calculation unit recognizes the number of training images representing the number of training images that can recognize the test image in the selected training image group for each of the test images. Based on the above, as the contribution ratio of the training image, the calculated score of the combination of the training image and the test image that minimizes the number of recognitions is equal to or greater than a predetermined threshold. Calculating the number of said combinations;
Training image selection device.

For each of the selected training image group and the test image, the number of recognitions representing the number of training images capable of recognizing the test image in the selected training image group, and the number of recognitions is less than a threshold. The result presentation unit further presenting at least one of the test image and the number of training images capable of recognizing the test image in which the number of recognition is less than a threshold among the selected training image group. The training image selection device according to any one of claims 1 to 3 .

The training image selection unit repeats an operation of selecting the largest contribution rate among the unselected training images or an operation of removing the smallest contribution rate of the unexcluded training images. training image selection apparatus according to any one of claims 1 to 4, wherein the contribution from the training images to select the high training images.

A combination of each of the training images and each of the test images based on a test image group including a test image obtained by photographing a space and a training image group including a training image obtained by photographing a subject in advance. On the other hand, calculating a score representing the probability that a common subject is included in the training image and the test image;
A step of calculating a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images based on the score of each of the combinations calculated by the subject recognition unit; When,
A training image selection unit sequentially selects an unselected training image from the training image group based on the contribution rate of each of the training images calculated by the contribution rate calculation unit, and for each of the test images The number of recognitions representing the number of training images that can recognize the test image among the selected training images is counted, and the minimum number of recognitions among the number of recognitions of each of the test images is determined in advance. Ending the selection of the training image when the threshold for the minimum number of times of recognition has been reached ,
A training image selection method including:

A combination of each of the training images and each of the test images based on a test image group including a test image obtained by photographing a space and a training image group including a training image obtained by photographing a subject in advance. On the other hand, calculating a score representing the probability that a common subject is included in the training image and the test image;
A step of calculating a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images based on the score of each of the combinations calculated by the subject recognition unit; When,
A training image selection unit sequentially selects an unselected training image from the training image group based on the contribution rate of each of the training images calculated by the contribution rate calculation unit, and for each of the test images The number of recognitions representing the number of training images capable of recognizing the test image among the selected training images is counted, and in the test image with the smallest number of recognitions, the subject common to the training image and the test image Ending the selection of the training image when there is no unselected training image in which the score representing the probability of being included is greater than or equal to a predetermined threshold ; and
A training image selection method including:

A combination of each of the training images and each of the test images based on a test image group including a test image obtained by photographing a space and a training image group including a training image obtained by photographing a subject in advance. On the other hand, calculating a score representing the probability that a common subject is included in the training image and the test image;
A step of calculating a contribution rate representing a degree of contribution to recognition of the test image group for each of the training images based on the score of each of the combinations calculated by the subject recognition unit; When,
A training image selection unit sequentially selects an unselected training image from the training image group based on the contribution rate of each of the training images calculated by the contribution rate calculation unit, and for each of the test images Counting the number of recognitions representing the number of training images that can recognize the test image among the selected training images ;
Each time the training image is sequentially selected by the contribution rate calculation unit, the number of times of recognition representing the number of training images that can recognize the test image in the selected training image group for each of the test images. Based on the above, as the contribution ratio of the training image, the calculated score of the combination of the training image and the test image that minimizes the number of recognitions is equal to or greater than a predetermined threshold. Calculating the number of combinations;
A training image selection method including:

The program for functioning a computer as each part of the training image selection apparatus of any one of Claims 1-5 .