JP2581136B2 - Image quality evaluation device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] Regarding the image quality evaluation of a multi-value halftone image, with the aim of improving the image quality by improving the conventional objective scale, the multi-value halftone image is converted into a band-pass image having different spatial frequencies. Band-pass image generating means for decomposing, error calculating means for calculating an error between images for each band-pass image having the same spatial frequency for the band-pass images of the original image and the processed image, and an error calculated by the error calculating means in advance. A linear sum calculating means for calculating a sum by multiplying the weight by a predetermined weight, determining the weight in relation to a spatial frequency characteristic of human vision, and outputting an output of the linear sum calculating means as an evaluation value.

[Industrial applications]

The present invention relates to image quality evaluation of a multi-value halftone image, and more particularly to an image quality evaluation device in which an image-processed image is improved in an objective scale indicating a degree of deterioration from an original image.

2. Description of the Related Art It is important to objectively measure the degree of image quality degradation from an original image of an image that has been subjected to image processing, and to evaluate the performance of a noise elimination filter, an information preservation-type coding method, and the like. In other words, image quality evaluation in the performance evaluation of image processing methods such as noise removal and coding must ultimately be based on subjective evaluation, but subjective evaluation varies greatly among subjects and is difficult to quantify, and Because of the drawback that it takes time to obtain a result, a simple objective evaluation method is often used.

Objective image quality evaluation methods for digital images include mean square error (RMS) and RMS (Root Mean
Square) value and SN ratio (Signal Noise Ratio) are generally used.

The original image (input image) is an N × N pixel array x, y = 0, 1,..., N
−1, given by f (x, y), the processed image (output image) is given by g (x, y) in the same pixel arrangement, and each pixel is represented by m bits (2 ^m Gradation), the error e (x, y) for each pixel between the original image and the processed image is (2) RMS value: e (rms) = [e ² ] ^1/2 ... As the S / N ratio, the following equation that easily uses the dynamic range of an output image is often used.

(SNR) p = [{peak ^{g (x, y)} 2} /2] 1/2 = {2 2m / 2} 1/2 ...... Further, as the SN ratio, to match the subjective intervals, A logarithmic scale such as the following equation is also used.

[Problem to be Solved by the Invention] The above-mentioned scale conventionally used indicates how far the original image and the processed image are on average apart from each other, and does not necessarily subjectively correspond to the image quality. was there.

For example, according to the SN ratio, when the gradation of the processed image is relatively different from the original image by several levels, a large value is taken, while a small number of pixels having a gradation that is significantly different from the original image. When (noise) is dispersed in the processed image, a small value is taken. However, from a subjective evaluation point of view, although the former is not much of a concern, the latter is more perceived as having a lower image quality.

For subjective image quality, reproduction of gradation is more important than resolution in an image whose gradation changes gradually, whereas resolution is more important than gradation in an image containing many fine parts. It has been known. However, there is a problem in the SN ratio that differences in image quality due to image properties are not taken into account.

An object of the present invention is to provide an image quality evaluation device that solves such a conventional problem.

[Means for solving the problem]

FIG. 1 is a block diagram showing the principle of the image quality evaluation apparatus of the present invention.

In the figure, reference numeral 1 denotes a band-pass image creating means for decomposing a multi-value halftone image into band-pass images having different spatial frequencies.

Reference numeral 2 denotes an error calculation stage, which calculates an error between images for each band-pass image having the same spatial frequency with respect to the band-pass images of the original image and the processed image.

Numeral 3 denotes a linear sum calculating means for multiplying the calculated error by a predetermined weight to calculate the sum.

(Operation)

The image quality evaluation device of the present invention considers that the image quality of a multi-value halftone image is related to visual spatial frequency characteristics, and the bandpass image creating means 1 decomposes the image into bandpass images having different spatial frequencies. The error calculating means 2 calculates an error for each band-pass image having the same spatial frequency, and the linear sum calculating means 3 calculates a weighted sum of the obtained errors and outputs it as an image quality evaluation value. is there.

In this way, an evaluation value of a scale that is more consistent with the subjective evaluation is obtained.

Hereinafter, the reason why the image quality evaluation value according to the present invention matches the subjective image quality evaluation will be described in detail.

The image quality of an image is related to the visual spatial frequency characteristic. For example, in the case of television, the frequency characteristic of noise visibility of random noise (particulate noise) takes the reduction filter characteristic in the case of one-dimensional noise. An expression is used as the evaluation function.

−20log ₁₂ {1+ (f / fc) ² } [dB] Here, fc corresponds to the cutoff frequency of the visual spatial frequency characteristic, and the standard television image is taken from a distance of 4H (H: screen height). When viewed, the horizontal and vertical spatial frequencies are taken to be 2.0 MHz and 2.5 MHz, respectively. Also, in the case of two-dimensional noise, it is known that the visibility of the noise has a band-pass filter type spatial frequency characteristic.

In the present invention, the error between the original image and the processed image is decomposed for each spatial frequency, and the weight of the error is changed for each spatial frequency. In order to decompose the error between the original image and the processed image into spatial frequencies, in the present invention, a method of representing an image as a Laplacian pyramid (for example, EHAdelson, PJBurt: “Image Dat
a Compression with the Laplacian Pyramid ", IEEE 198
1 Pattern Recognition & Image Processing, P.218-223
See). The Laplacian pyramid is a method of representing an image as a set of bandpass filter images of different spatial frequencies (hereinafter simply referred to as bandpass images),
It can be configured as follows.

The correlation between the pixels of an image is represented as a low-pass filtered image of the image (hereinafter simply referred to as a low-pass image),
The band pass image is an image obtained by subtracting the low pass image from the original image. In addition, a low-pass image (an image blurred by applying a localization filter) can be expressed at a lower resolution than the original by the sampling theorem.

Therefore, a low-pass filter is applied to the image to reduce the cutoff frequency by a factor of two, and the resolution is reduced by a factor of two.
The difference between adjacent low-pass images is calculated to create a band-pass image. FIG. 2 is a diagram showing a method of forming a bandpass image. The low-pass image and the band-pass image are pyramid hierarchical data in order to sequentially reduce the resolution.

The original image is reconstructed as shown in FIG. 3 using a pyramid-shaped band-pass image. That is, by successively accumulating the bandpass images, the lowpass image pyramid hierarchy is sequentially restored from the top to the bottom, and an original image is finally obtained.

In the above document, the pyramid hierarchical data of the low-pass image and the pyramid hierarchical data of the band-pass image are called a Gaussian Pyramid and a Laplacian Pyramid, respectively. The operation of creating the i-th hierarchy Gi of the Gaussian pyramid and the i-th hierarchy Li of the Laplacian pyramid can be formulated as follows.

Gi + 1 = REDUCE [Gi] Li + 1 = Gi-EXPAND [Gi + 1] = Gi-EXPAND [REDUCE [Gi]] The REDUCE in the formula is applied with a low-pass filter to reduce the resolution and create an upper layer of the upper Gaussian pyramid. Operation, defined by:

Here, W is a weight function, and performs spatial convolution over the mask size m × n. The reduction in low-pass filtering is as shown in FIG.
× 5 (odd type) or 4 × 4 (even type) is used, and the resolution is successively reduced. A, b, c in the figure are convolution coefficients, which determine the strength of the filter.

EXPAND is an operation for expanding the same size as the lower layer in the Gaussian pyramid, and is defined by the following equation.

The EXPAND operation increases the image size by interpolation.

The Laplacian pyramid as a bandpass image
The higher the lower layer, the higher the frequency component of the spatial frequency. For this reason, in a portion where gradation is gradually changed by the Laplacian pyramid, fine quantization (gradation) of the gradation is required, and an image including many fine parts is finely sampled (resolution). ) Is important.

Next, the meaning of the present invention will be considered with respect to an error between the original image, the bandpass image, and the bandpass image of the processed image after the original image and the processed image are decomposed into bandpass images. Original image f (x,
y) and the processed image g (x, y) are represented by the following row vector.

F = (f (0,0), f (0,1), ..., f (0, N-1), f (1,
0), ..., f (N-1, N-1)) G = (g (0,0), g (0,1), ..., g (0, N-1), g (1,
0),..., G (N−1, N−1)) The original image and the processed image are represented by the following equation as a pyramid of a bandpass image.

F = F ₀ + F ₁ +... + F _M G = G ₀ + G ₁ +... + G _M where N = 2 ^M , Fi = EXPAND ^{[i + 1]} (REDUCE ^{[i + 1]} (F)) Gi = EXPAND ^{[i + 1]} (REDUCE ^{[i + 1]} (G)) EXPAND ^[j] and REDUCE ^[j] indicate that each operation is repeated j times.

 The mean square error is represented by the following equation.

Here, T represents transposition.

If the bandpass layers are close to uncorrelated, the second
The term can be ignored, and the mean square error is represented by the sum of the mean square errors between the layers.

In the present invention, this is further developed and expressed as a weighted sum of mean square errors between layers as follows.

When hierarchical pyramid data is used, the above equation is expressed as follows.

Where Fi '= REDUCE ^{[i + 1]} (F), Gi' = REDU
CE ^{[i + 1]} (G), 'when = ^{2 2i} / N 2, between bandpass hierarchy described above under the uncorrelated condition, ² ≒ ^2' wi becomes.

As described above, it has been shown that the sum of the mean square errors obtained for each bandpass image is associated with the conventional mean square error.

In the present invention, by selecting the weight wi so as to match the visual spatial frequency characteristics for each spatial frequency band, it is possible to obtain an objective evaluation value that is more consistent with the subjective evaluation than in the related art.

In the present invention, the average luminance of the entire image is obtained as the average value of the highest level (DC component). Therefore, when the processed image is shifted from the original image by several levels relative to the original image, the evaluation value is obtained. The weight is determined so as to take a small value, and when a small number of pixels (high-frequency components) that are significantly different from the original image are dispersed in the processed image, a large value can be taken, and the subjective evaluation can be matched.

Further, in the present invention, if the SN ratio that is logarithmically converted for each bandpass image is obtained and the weighted sum is obtained, the conventional SN
Although there is no direct association with the ratio, a more subjective evaluation scale can be obtained.

〔Example〕

FIG. 5 is a diagram showing a configuration of an image quality evaluation device according to one embodiment of the present invention.

 Hereinafter, the operation of the present embodiment will be described.

(1) Two images, an original image and a processed image, are sequentially input from an input terminal 4 as image data, and are decomposed into band-pass images for each frequency band by a band-pass image creating unit 1.
First, when an original image is input, the image input from the input terminal 4 is stored in the memory 111 (G ₀ ) for one screen. The image in the memory 111 is the 0th layer of the Gaussian pyramid. The image in the memory 111 is subjected to the aforementioned REDUCE operation in the circuit 112 (REDUCE), and is stored in the memory 121 (G ₁ ). The image in the memory 121 is 1 the size of the image in the memory 111, and is the first layer of the Gaussian pyramid. Next, the image in the memory 121 is subjected to the above-described EXPAND operation in the circuit 114 (EXPAND), and is input to the subtraction circuit 113 (Σ). The subtraction circuit 113 also receives an input from the memory 111, obtains a bandpass image of the 0th layer, and stores it in the memory 211 (L02) via the demultiplexer 115 (DMUX). The creation of bandpass images in the first and higher layers is almost the same, except that each time the image size becomes higher, the size is reduced to 順次 sequentially. The highest level is the average value of all pixels. The bandpass images on the first and higher layers are stored in the memories 211, 221,. When the processed image is input from the input terminal 4, the processed band image is processed in the same manner as the original image, and the decomposed band-pass image is output from the demagnifier 11.
The memory 212 (L01), 222 (L1
1), ...

(2) Next, the error calculator 2 calculates a square error between the original image and the processed image. In the 0th hierarchy, the difference between the bandpass image 211 of the original image and the bandpass image 212 of the processed image is obtained by the subtraction circuit 213 (Σ), and the square thereof is obtained by the squaring circuit 214 (SQ). Then the squared error is calculated in circuit 215
Accumulated by (Σ). As a result, the sum of the square errors of the 0th layer is obtained. A square error is similarly obtained in the first and higher layers.

(3) Next, in the linear sum calculation means 3, the weighting circuit 311 (w
_{0), 312 (w 1)} , ... , the respective weight w ₀ to the square error of each layer obtained by the error calculation means _{2 ', w 1', ...} , after multiplication by w _M ', the addition circuit 302 ( The sum is obtained in Σ), and the calculation result is output from the output terminal 5.

〔The invention's effect〕

As described above, according to the present invention, an error scale of a processed image in consideration of the spatial frequency characteristic of visual perception is provided, and an objective evaluation more consistent with subjective evaluation can be performed, and its practical effect is extremely large. It is.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the principle of the present invention, FIG. 2 is a diagram showing a method of constructing a pyramid of a bandpass image, FIG. 3 is a diagram showing restoration of an original image from a bandpass image, and FIG. FIG. 5 is a diagram showing a configuration of one embodiment of the present invention. In the drawings, 1 is a band-pass image creating means (band-pass image creating unit), 2 is an error calculating means (part), 3 is a linear sum calculating means (part), 4 is an input terminal, 5 is an output terminal, 111, 121,. , 141 are memories (G ₀ , G ₁ ,…, G _M ), 112, 122,…, 142 are REDUCE circuits (REDUCE), 113, 123,…, 143, 213, 223,…, 243 are subtraction circuits (Σ), 114, 124,. Is an EXPAND circuit (EXPAND), 115,125, ... are demultiplexers (DMUX), 211,212,221,222, ..., 241,242 are memories (L01, L02, ...), 214,224, ..., 244 are squaring circuits (SQ), and 215,225, ..., 245,302 are additions , 341 represent weighting circuits (w ₀ , w ₁ ,..., W _M ), respectively.

Claims (1)

(57) [Claims] In an image quality evaluation for evaluating the degree of image quality deterioration of an image-processed multi-value halftone image from an original image, a band-pass image is generated by decomposing the multi-value halftone image into band-pass images having different spatial frequencies. Means (1), error calculating means (2) for calculating an error between images for each band-pass image having the same spatial frequency with respect to the band-pass images of the original image and the processed image, and error calculating means (2). A linear sum calculation means (3) for calculating a sum by multiplying the error by a predetermined weight, determining the weight in relation to a spatial frequency characteristic of human vision, and evaluating an output of the linear sum calculation output (3) An image quality evaluation device characterized in that it is configured to output as a value.