JP7544000B2 - Quality defect cause extraction method and quality defect cause extraction device - Google Patents

Description

The present invention relates to a quality defect cause extraction method and a quality defect cause extraction device.

Patent Document 1 discloses the following method for predicting the quality of steel products and extracting factors that affect the quality. First, past actual operating condition data is converted into mutually uncorrelated principal component scores, and a regression equation showing the relationship between the quality data and the principal component scores is obtained by performing a regression analysis on the relationship between the past actual quality data and the principal component scores. Then, the principal component scores in the regression equation are returned to the original operating condition data to obtain the degree of influence that each operating condition has on the quality.

Patent Document 2 also discloses the following method for extracting factors that affect the quality of industrial products, not limited to steel products. First, past operating condition data and quality data are collected and compared, and the correlation strength between the two is calculated. Next, potential variation causes in the manufacturing process are listed, and the mutual correlation strength between each operating condition data is calculated for all combinations. Then, by using manufacturing sequence information, a mutual causal association structure model is automatically generated, and the causal model is automatically analyzed to automatically derive the root variation cause from among the candidate quality variation causes.

Patent No. 4613751 Patent No. 4239932

However, the method of Patent Document 1 is unable to express the nonlinear effect of operating conditions on quality, so the reliability of the regression equation is low and it may not be sufficiently effective in extracting and identifying the causes of quality defects.

Similarly, the method of Patent Document 2 is also unable to express the nonlinear effect of operating conditions on quality, so the reliability of the correlation coefficient is low and it may not be sufficiently effective in extracting and identifying the causes of quality defects.

As a result, in the past, verifying and identifying the causes of quality defects in product manufacturing across multiple processes could require a great deal of time and effort, and even during this verification process, quality defects would continue to occur in newly manufactured products, resulting in a problem of reduced product yields.

The present invention has been made in consideration of the above, and aims to provide a quality defect factor extraction method and quality defect factor extraction device that can extract the causes of product quality defects with high accuracy.

In order to solve the above-mentioned problems and achieve the objective, the quality defect factor extraction method of the present invention includes an influence calculation step of calculating, for each quality category, the relative influence of the operating conditions on the defect occurrence probability score, which is the final output for each element of a convolutional layer, in a convolutional neural network trained with input data being operating conditions in multiple processes when manufacturing a product and output data being defect occurrence probability scores for each quality category of the manufactured product, and a factor presentation step of presenting candidates for the operating conditions that are factors of quality defects of the product for each quality category based on the relative influence.

The quality defect factor extraction method according to the present invention is the above invention, in which the convolutional neural network is trained in each process using the operating conditions of parts corresponding to multiple parts in the product manufactured in the final process as input data, and the defect occurrence probability scores of multiple parts in the product manufactured in the final process as output data.

In the quality defect factor extraction method according to the present invention, the input data is data in which the operating conditions in multiple processes are arranged as a one-dimensional vector.

In addition, in the quality defect factor extraction method according to the present invention, the convolutional layer is one or more final convolutional layers.

In addition, in the method for extracting quality defects according to the present invention, the product is a steel product manufactured through a plurality of processes, and the steel product manufactured in the final process is treated as a unit, and the operating conditions in each process are linked to the quality of the steel product inspected after the final process to form the input data.

The quality defect factor extraction method according to the present invention is the above-mentioned invention, in which the product is a steel product manufactured through a plurality of processes, and the plurality of processes includes one or more of ironmaking, steelmaking, hot rolling, cold rolling, annealing, and surface treatment.

In addition, in the quality defect factor extraction method according to the present invention, in the above invention, the influence calculation step evaluates, for each element of the convolution layer, a differential coefficient to the final output for a small change in the pixel in each element to determine a weighting coefficient for the element, and calculates the relative influence of the operating conditions based on each weighting coefficient and each element.

In order to solve the above-mentioned problems and achieve the objective, the quality defect factor extraction device according to the present invention is equipped with an influence calculation means for calculating, for each quality category, the relative influence of the operating conditions on the defect occurrence probability score, which is the final output for each element of the convolutional layer, in a convolutional neural network trained with input data being the operating conditions in multiple processes when manufacturing a product and output data being the defect occurrence probability score for each quality category of the manufactured product, and a factor presentation means for presenting candidates for the operating conditions that are factors of quality defects of the product for each quality category based on the relative influence.

The quality defect factor extraction method and quality defect factor extraction device according to the present invention can extract the quality defect factors of products for each quality category with high accuracy by calculating the relative influence of each operating condition on the defect occurrence probability score. This allows the quality defect factors to be identified earlier than before, and improvements to the operating conditions to be started earlier, thereby suppressing the decline in the long-term yield of manufactured products.

FIG. 1 is a diagram showing a schematic configuration of a quality defect cause extraction device according to an embodiment of the present invention. FIG. 2 is a diagram showing, in a table format, performance data created by linking operation condition data with quality data in the quality defect cause extraction device according to the embodiment of the present invention. FIG. 3 is a flowchart showing the procedure of a quality defect cause extraction method executed by the quality defect cause extraction device according to the embodiment of the present invention. FIG. 4 is a graph showing the relative influence of each operating condition according to an embodiment of the present invention.

The quality defect cause extraction method and quality defect cause extraction device according to the embodiment of the present invention will be described with reference to the drawings.

[Quality defect cause extraction device]
First, the configuration of a quality defect factor extraction device according to an embodiment of the present invention will be described with reference to Figures 1 and 2. The quality defect factor extraction device uses data on actual operating conditions (hereinafter referred to as "operation condition data") and data on the quality of the final product obtained under those operating conditions (hereinafter referred to as "quality data") to build a quality prediction model (mathematical model) that predicts the quality of the product for any operating conditions. Then, based on the input/output results of this quality prediction model, factors that have a large influence on the occurrence of quality defects (hereinafter referred to as "quality defect factors") are extracted, and the extracted quality defect factors are presented to an operator, etc.

As shown in FIG. 1, the quality defect cause extraction device 1 includes an input unit 10, a memory unit 20, a calculation unit 30, and a display unit 40.

The input unit 10 is an input means for the calculation unit 30, and is realized by an input device such as a keyboard, a mouse pointer, or a numeric keypad.

The storage unit 20 is composed of recording media such as an EPROM (Erasable Programmable ROM), a hard disk drive (HDD), and removable media. Examples of removable media include a universal serial bus (USB) memory, a compact disc (CD), a digital versatile disc (DVD), and a Blu-ray (registered trademark) disc (BD). The storage unit 20 can store an operating system (OS), various programs, various tables, various databases, and the like. The storage unit 20 also stores an operation database (DB) 21.

Operation DB21 stores searchable operational data for products that have been manufactured in the past. In this embodiment, the "product" may be a steel product manufactured through multiple processes. For example, in the case of a steel product, the "multiple processes" may include one or more of the following processes: iron-making process, steel-making process, hot rolling process, cold rolling process, annealing process, and surface treatment process.

In addition, examples of "operation data" include operational condition data (manufacturing condition data, sensor data, etc.) from multiple processes when a product was manufactured in the past, and quality data of the product inspected after the final process.

For example, in the case of steel products, the above-mentioned operational condition data includes data on product thickness, product width, product length, steel components, casting speed, etc. in the steelmaking process, data on product thickness, product width, product length, product temperature, rolling speed, etc. in the hot rolling process, data on product thickness, product width, product length, rolling speed, etc. in the cold rolling process, data on product thickness, product width, product length, temperature in the surface treatment furnace, surface treatment speed, etc. in the surface treatment process. In addition, quality data includes data on the results of surface defect inspection at the end of the surface treatment process.

The calculation unit 30 is realized by a processor such as a CPU (Central Processing Unit) and a memory (main storage unit) such as a RAM (Random Access Memory) and a ROM (Read Only Memory).

The calculation unit 30 loads a program into the working area of the main memory and executes it, and controls each component part through the execution of the program to realize functions that meet a predetermined purpose. Through the execution of the program, the calculation unit 30 functions as a data linking unit 31, a model creation unit 32, an influence degree calculation unit 33, and a factor presentation unit 34. Note that while FIG. 1 shows an example in which the functions of each unit are realized by, for example, one computer, the means for realizing the functions of each unit is not particularly limited, and the functions of each unit may be realized by, for example, multiple computers.

The data linking unit 31 links the operation condition data with the quality data. First, the data linking unit 31 collects from the operation DB 21 the operation condition data measured by multiple measuring devices during the multi-process manufacturing process and the quality data of the product actually inspected after the final process.

The operation condition data X during the manufacture of a certain product has, for example, the following elements:
X=[x ₁ , x ₂ ,..., x _m ,..., x _M ] (1≦m≦M)

However, the above "x _m " indicates one actual value of one manufacturing facility or manufacturing device at process k (1≦k≦K) of the K processes when manufacturing a product through K processes from the first process to the Kth process, and indicates the mth actual value (1≦m≦M) out of a total of M types of actual values in all processes.

Moreover, quality data Y of a certain product has, for example, the following elements:
Y = [y]

Here, "y" above indicates the quality inspection result of the product after process K, i.e., the final process, when K processes from the first process to the Kth process are passed in manufacturing the product. Here, the quality may be a plurality of different quality evaluations. In other words, a plurality of outputs may be considered as _yi for each quality category.

The data linking unit 31 links the operational condition data in each process with the quality data of the product inspected after the final process, with the product manufactured in the final process (final product) as one unit, to create performance data in a table format, for example, as shown in Fig. 2. In the performance data in the figure, data for N products are arranged in the row direction, and quality data Y and operational condition data _Xm (1≦m≦M) are arranged in the column direction. The performance data shown in the figure is used as input data when creating a quality prediction model, which will be described later. The data linking unit 31 may store the linked performance data in the storage unit 20.

The data linking unit 31 may link the operating condition data and the quality data for each specific portion of the product. For example, in the case of steel products such as steel plates, the length changes as the product progresses through the processes due to rolling, etc. Therefore, using specific portions within the final product (steel plate) as a reference, it is tracked to determine which portions of the product in the intermediate process correspond to those portions. Then, using the specific portion of the final product as a reference, the operating condition data for specific portions of each process is linked to the quality data for the corresponding portion of the final product. This makes it possible to create input data for the quality prediction model for each specific portion of the final product.

The model creation unit 32 uses the performance data created by the data linking unit 31 to build a quality prediction model that predicts product quality for any operating conditions. In this embodiment, the quality prediction model is built using a Convolutional Neural Network (hereinafter referred to as "CNN"), which is a type of deep learning algorithm. CNN is a machine learning algorithm that mimics the network structure of nerve cells (neurons) called the neural network of the human brain and its learning mechanism. CNN is composed of an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer.

Input data is input to the input layer. In this embodiment, the input data is performance data that links the operating conditions in each process with the product quality. In addition, the input data is data in which the operating conditions in multiple processes are arranged as a one-dimensional vector, as shown in Figure 2.

In the convolutional layer, a filter called a kernel is applied to each partial region of the input data to convolve the data and extract features to construct a feature map (also called an "element"). There may be multiple filters, in which case multiple feature maps will be constructed. In addition, when training a CNN, the filter parameters are learned automatically.

In the pooling layer, the size of each feature map obtained by convolution is reduced. The convolutional layers and pooling layers may be configured with the same number of layers, alternating. In the fully connected layer, classification is performed to classify the combination of feature maps into specific prediction results. For example, in the case of steel products, the aforementioned "specific prediction results" include the presence or absence of quality defects such as surface defects.

By applying CNN, it is possible to extract characteristic patterns from input data and obtain classifications that are identified with high accuracy based on those patterns as output data. In this embodiment, complex patterns of multiple operating conditions that affect quality are identified by a CNN (one-dimensional CNN) that outputs the classification probability of the occurrence or absence of quality defects for each one-dimensional input of each operating condition. In this case, a threshold for the classification probability is set so that if the classification probability of the occurrence of quality defects is less than an arbitrary value p (0≦p≦1), it is classified as "no quality defect," and if the classification probability is equal to or greater than an arbitrary value p, it is classified as "quality defect."

In addition, instead of a one-dimensional CNN, a two-dimensional CNN may be used that outputs the classification probability of the occurrence of a quality defect for a two-dimensional input, for example, with the row direction being the position coordinates of the product in a specific direction (e.g., the length direction of the product) and the column direction being each operating condition.

The model creation unit 32 uses the operating conditions in multiple processes when manufacturing a product as input data and the defect occurrence probability score for each quality category of the manufactured product as output data to train the CNN. Note that the "defect occurrence probability score for each quality category" refers to the classification probability regarding the presence or absence of multiple types of defects, such as "the probability of occurrence of defect A is 70%" and "the probability of occurrence of defect B is 30%."

The model creation unit 32 may perform CNN training in each process using the operating conditions of parts corresponding to multiple parts in the product manufactured in the final process as input data and the defect occurrence probability scores of multiple parts in the product manufactured in the final process as output data. This makes it possible to trace the operating conditions corresponding to a specific part of the final product back through each process and build a quality prediction model.

The influence calculation unit 33 calculates the relative influence (relative influence evaluation value) of each operating condition on the defect occurrence probability score. Specifically, the influence calculation unit 33 calculates the relative influence of the operating condition on the defect occurrence probability score, which is the final output for each element (feature map) of the convolutional layer of the CNN, for each quality category. Note that the "convolutional layer" mentioned above is one or more final convolutional layers in the CNN.

The influence calculation unit 33 applies a mathematical method to the CNN that predicts product quality, which expresses the basis of the prediction result corresponding to the factors that affect the product quality. In this embodiment, the mathematical method used is Grad-CAM (Gradient-weighted Class Activation Mapping) described in Reference 1 below.

Reference 1: Ramprasaath R. Selvaraju, et al. “Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization”, 2017, IEEE International Conference on Computer Vision (ICCV)

The influence calculation unit 33 evaluates the differential coefficient of the final output for each element (feature map) of the convolutional layer of the convolutional neural network in response to a small change in the pixels in each element to determine the weighting coefficient of the element, and calculates the relative influence of the operating conditions based on each weighting coefficient and each element.

Specifically, the influence calculation unit 33 calculates the relative influence of each operational condition on the defect occurrence probability score by the following procedure.
(1) In a trained CNN, the output of a specific classification class judgment (a defect occurrence probability score for a specific quality category) is differentiated with respect to the value of one point (one pixel) of one feature map in the final convolutional layer, and a differential coefficient that represents the rate of change in the defect occurrence probability score when the pixel value is slightly changed is calculated.
(2) A weighting factor for one feature map corresponding to a particular classification class is calculated by averaging the derivatives of all pixels.
(3) All feature maps in the final convolutional layer are multiplied by a weighting factor and averaged, resulting in a superposition that emphasizes feature maps that are important for classifying a particular class.
(4) Calculate the relative influence, which takes real values ranging from 0 to 1, through an activation function ReLU (ramp function).

Here, in the above calculation, the effect on the output of a specific classification class is evaluated in relation to the final convolutional layer, which is obtained from the input layer via multiple convolutional layers and pooling layers. Therefore, the size of the final convolutional layer is a reduced size of the input layer. Therefore, by expanding it by the amount of the reduced size, it is possible to clearly establish a correspondence with the input.

The factor presentation unit 34 presents candidates for operating conditions that are factors that cause product quality defects for each quality category based on the relative influence of each operating condition calculated by the influence calculation unit 33.

For example, when a one-dimensional CNN is used, the factor presentation unit 34 displays a bar graph on the display unit 40 with each item of the input data on the horizontal axis and the relative influence on the vertical axis. Furthermore, the factor presentation unit 34 may preset a color according to the magnitude of the value of the relative influence of each item of the input data, and display the bar graph on the display unit 40 as a one-dimensional color map with the preset color assigned to each item on the horizontal axis. Alternatively, the factor presentation unit 34 may rearrange each item of the input data in descending order of relative influence, extract only items with a relative influence greater than a preset threshold, and display them on the display unit 40 as a list of quality defect factors.

When a two-dimensional CNN is used, a color is preset according to the magnitude of the value for the relative influence of each item of input data, and the position coordinate in a specific direction of the product (e.g., the length direction of the product) is set on the vertical axis, and each operating condition is set on the horizontal axis. A preset color is assigned to each corresponding position, and the map is displayed on the display unit 40 as a two-dimensional color map.

In this way, based on the feature map of the CNN, it captures quality defect patterns from the input operating conditions and presents the factors (operating conditions) that are the basis for predicting the classified output. This makes it possible to present the operating conditions that have the greatest impact on the output (defect occurrence probability score), and to identify the cause of quality abnormalities.

The display unit 40 is realized by a display device such as an LCD display or a CRT display. Based on the display signal input from the calculation unit 30, the display unit 40 displays, for example, the results of the calculation of the relative influence degree by the influence degree calculation unit 33, candidates for operating conditions that are factors causing quality defects by the cause presentation unit 34, etc., in the form of characters, figures, etc.

[Method for extracting quality defect causes]
A quality defect factor extraction method using the quality defect factor extraction device 1 according to the embodiment of the present invention will be described with reference to Fig. 3. The quality defect factor extraction method includes a data collection step (step S1), a data linking step (step S2), a model creation step (step S3), an influence degree calculation step (step S4), and a factor presentation step (step S5). Among the steps in the quality defect factor extraction method, the data collection step, the data linking step, and the model creation step may be performed in advance at a timing different from that of the other steps.

First, the data linking unit 31 collects operation condition data and quality data for multiple processes from the operation DB 21 (step S1). Next, the data linking unit 31 links the operation condition data and the quality data (step S2). Next, the model creation unit 32 trains the CNN using the operation conditions for multiple processes when manufacturing a product as input data and the defect occurrence probability scores for each quality category of the manufactured product as output data (step S3).

Next, the impact calculation unit 33 calculates the relative impact of each operating condition on the defect occurrence probability score (step S4). Finally, the cause presentation unit 34 presents the causes of product quality defects for each quality category based on the relative impact of each operating condition calculated by the impact calculation unit 33 (step S5).

[Example]
An example of the present invention will be described with reference to Fig. 4. In this example, the quality defect factor extraction method according to the embodiment was applied to performance data on steel products. Specifically, in the manufacture of steel sheets, the factors of surface defects occurring on the surface of the steel sheet after passing through the surface treatment process were extracted from performance data on products that had passed through the steelmaking process, hot rolling process, cold rolling process, and surface treatment process.

In the data collection step, the following operational condition data were collected: data on operational conditions in the steelmaking process (product thickness, product width, product length, steel components, casting speed, etc.), data on operational conditions in the hot rolling process (product thickness, product width, product length, product temperature, rolling speed, etc.), data on operational conditions in the cold rolling process (product thickness, product width, product length, rolling speed, etc.), data on operational conditions in the surface treatment process (product thickness, product width, product length, surface treatment furnace temperature, surface treatment speed, etc.), and surface defect inspection result data (quality data) of the steel plate inspected at the end of the surface treatment process.

In the data linking step, the operating condition data for the steelmaking process, hot rolling process, cold rolling process, and surface treatment process, and the surface defect inspection result data at the end of the surface treatment process were linked and stored in the memory unit 20.

In the model creation step, a quality prediction model was created using a one-dimensional CNN that takes the operating conditions of each process as one-dimensional input and outputs the classification probability (defect occurrence probability score) of the presence or absence of surface defects. In this quality prediction model, a classification probability threshold was set so that if the classification probability of surface defect occurrence is less than 0.5, it is classified and predicted as "no surface defect," and if the classification probability is 0.5 or more, it is classified and predicted as "surface defect present."

In the impact calculation step, Grad-CAM was applied to the CNN to calculate the relative impact of each operating condition on the classification probability of surface defect occurrence.

In the factor presentation step, candidates for operating conditions that are factors in the occurrence of surface defects were presented based on the relative influence of each operating condition. In this embodiment, as shown in FIG. 4, a bar graph with the horizontal axis representing each operating condition item, which is the input data, and the vertical axis representing the relative influence was displayed on the display unit 40 with a color preset for each operating condition item. In this embodiment, the operating conditions with a relative influence of 0 or more and less than 0.5 were displayed in blue (shown in FIG. 4 by diagonal hatching slanting downward to the right), 0.5 or more and less than 0.8 in yellow (shown in FIG. 4 by diagonal hatching slanting downward to the left), and 0.8 or more and less than 1 in red (shown in FIG. 4 by dot hatching), and were displayed in different colors according to their relative influence.

In this embodiment, as shown in FIG. 4, the relative influence of all operating conditions is presented, but it is also possible to extract and present, for example, only operating conditions whose relative influence is equal to or greater than a threshold value Th1 (0.5), or only operating conditions whose relative influence is equal to or greater than a threshold value Th2 (0.8).

In this way, in this embodiment, by inputting the actual values of the operating conditions for which it is desired to extract the causes of surface defects, which are a type of quality defect, the relative influence of each operating condition on the occurrence of surface defects is calculated, and the operating conditions with the highest relative influence are presented as the causes. As a result, it was possible to extract the causes of surface defects from the actual data early and with high accuracy.

In the quality defect factor extraction method and quality defect factor extraction device 1 according to the embodiment described above, a highly accurate quality prediction model (CNN) is created that takes the operating conditions of each process as input and outputs a product defect occurrence probability score, based on performance data that links the operating conditions across multiple processes of products manufactured in the past with the quality of the products obtained by manufacturing under those operating conditions. The relative impact of each operating condition on the occurrence of product quality defects is then calculated, and operating conditions with a high relative impact are presented as factors of quality defects.

In this way, according to the quality defect factor extraction method and quality defect factor extraction device 1 of the embodiment, by calculating the relative influence of each operating condition on the defect occurrence probability score, it is possible to extract the quality defect factors of products for each quality category with high accuracy. Therefore, it is possible to investigate the quality defect factors earlier than before and start improving the operating conditions earlier, thereby suppressing the decline in the long-term yield of manufactured products.

The quality defect factor extraction method and quality defect factor extraction device 1 according to the present invention have been specifically described above using the form and examples for carrying out the invention, but the spirit of the present invention is not limited to these descriptions and must be interpreted broadly based on the claims. Furthermore, it goes without saying that various changes, modifications, etc. based on these descriptions are also included in the spirit of the present invention.

Reference Signs List 1 Quality defect cause extraction device 10 Input unit 20 Storage unit 21 Operation DB
30 Calculation unit 31 Data linking unit 32 Model creation unit 33 Influence degree calculation unit 34 Factor presentation unit 40 Display unit

Claims (8)

In a convolutional neural network trained with input data representing operation conditions in a plurality of processes in manufacturing a product and output data representing the quality of each quality category of the manufactured product, an influence degree calculation step of calculating, for each quality category, a relative influence degree of the operation conditions on the quality , which is a final output for each element of a convolutional layer;
a factor presentation step of presenting candidates of the operation conditions that are factors of quality defects of the products for each of the quality categories based on the relative influence degree;
Including,
The product is a steel product manufactured through a plurality of processes,
The steel product manufactured in the final process is regarded as one unit, and the operating conditions in each process are linked to the quality of the steel product inspected after the final process to obtain the input data.
Method for extracting causes of quality defects. 2. The method for extracting causes of quality defects according to claim 1, wherein the convolutional neural network is trained in each process using operating conditions of parts corresponding to multiple parts in a product manufactured in a final process as input data and qualities of the multiple parts in the product manufactured in the final process as output data. The quality defect factor extraction method according to claim 1 or 2, wherein the input data is data in which the operating conditions in multiple processes are arranged as a one-dimensional vector. The quality defect factor extraction method according to any one of claims 1 to 3, wherein the convolutional layer is one or more final convolutional layers. 2. The method for extracting causes of quality defects according to claim 1, wherein the convolutional neural network is trained in each process using operating conditions of a product to be manufactured in a final process as input data and the quality of the product to be manufactured in the final process as output data. The method for extracting cause of quality defects according to claim 1 , wherein the plurality of processes include at least one of ironmaking, steelmaking, hot rolling, cold rolling, annealing, and surface treatment. The quality defect factor extraction method according to any one of claims 1 to 6, wherein the influence calculation step determines a weighting factor for each element of the convolution layer by evaluating a differential coefficient to a final output for a small change in pixels in each element, and calculates a relative influence of the operating conditions based on each weighting factor and each element. In a convolutional neural network trained with input data being operational conditions in a plurality of processes when manufacturing a product and output data being the quality of each quality category of the manufactured product, an influence degree calculation means calculates, for each quality category, a relative influence degree of the operational conditions on the quality , which is a final output for each element of a convolutional layer;
a factor presenting means for presenting candidates of the operation conditions that are factors of quality defects of the products for each of the quality categories based on the relative influence degree;
Equipped with
The product is a steel product manufactured through a plurality of processes,
The steel product manufactured in the final process is regarded as one unit, and the operating conditions in each process are linked to the quality of the steel product inspected after the final process to obtain the input data.
Quality defect cause extraction device.

Images