CN106651829A - Non-reference image objective quality evaluation method based on energy and texture analysis - Google Patents

Abstract

The invention discloses an objective quality evaluation method of no-reference image based on energy and texture analysis, comprising the following steps: step S1, acquiring the distorted image; step S2, performing statistics on the damage characteristics of the distorted image; step S3, calculating the wavelet of the distorted image Energy difference; step S4, calculating the texture feature of the distorted image; step S5, using BP neural network learning to establish the mapping relationship between the extracted damage performance feature, wavelet energy difference and texture feature and the subjective evaluation score of the test image, to obtain the image quality Evaluation. By adopting the technical solution of the invention, the image quality can be tested automatically and comprehensively, and the consistency between the evaluation result and the subjective vision can be improved.

Description

A No-Reference Image Objective Quality Assessment Method Based on Energy and Texture Analysis

technical field

The invention belongs to the field of computer vision, and in particular relates to an objective quality evaluation method of no reference image based on energy and texture analysis.

Background technique

Digital images will inevitably be damaged in the process of acquisition, storage, and transmission. The decline in quality not only affects the visual perception effect, but also reduces the processing efficiency of all links in image communication. In image quality evaluation methods, two methods are mainly used: subjective evaluation and objective evaluation. Subjective experimental methods are the most accurate and reliable, but time-consuming and labor-intensive and unable to achieve embedded processing. The current research mainly focuses on the objective algorithm. Obtaining a simple and accurate method that conforms to the human visual system is the main goal of the current research on objective quality evaluation methods.

In the field of actual image processing, most of the original information of the loss image cannot be obtained. In image acquisition systems, real-time monitoring systems, network transmission systems, etc., the real-time requirements for image evaluation make the non-parametric objective quality evaluation algorithm extremely important. At present, there are no reference image quality evaluation methods can be roughly divided into three categories: 1) Specified distortion type, this type of method evaluates image quality by measuring the unique distortion effect of known type, but the limited distortion type description limits the application of this type of algorithm . 2) A method based on training and learning. This type of method extracts the perceptual features of the image, and then uses the training learning algorithm to evaluate the image quality. Its performance is directly related to the extracted features, so whether the extracted features are accurate and comprehensive determines the pros and cons of this type of method. 3) Methods based on statistical models of natural scenes. This method assumes that an image is only a tiny subset of all image information, and then attempts to evaluate image quality by looking for differences between the two. The distortion features or statistical features extracted by these current algorithms are very limited and cannot comprehensively summarize the image distortion information; secondly, the performance of each feature in human vision is closely related to the image content. If the image content texture and complexity are ignored, the performance of the algorithm can be objectively evaluated. The evaluation results will have certain deviations from the subjective vision.

Contents of the invention

The main purpose of the present invention is to provide a no-reference image quality evaluation method, which can automatically and comprehensively test image quality, and improve the consistency between evaluation results and subjective vision.

To achieve the above object, the present invention adopts the following technical solutions:

A no-reference image objective quality assessment method based on energy and texture analysis includes the following steps:

Step S1, the acquired distorted image;

Step S2, performing damage feature statistics on the distorted image;

Step S3, calculating the wavelet energy difference of the distorted image;

Step S4, calculating texture features of the distorted image;

Step S5, using BP neural network to learn and establish the mapping relationship between the extracted damage performance features, wavelet energy difference and texture features and the subjective evaluation score of the test image, so as to obtain the image quality evaluation.

Preferably, the damage feature statistics include: block effect, blur effect and noise factor.

Preferably, step S3 specifically includes: establishing a wavelet lossless energy model for the distorted image, calculating the damage energy spectrum for the test image, and using its first-level and second-level energies to predict the third-level to eighth-level of the image according to the lossless energy model Non-destructive energy performance, compare the difference between the original energy and the restored energy of the damaged image, and obtain the wavelet energy difference.

Preferably, step S4 specifically includes: statistical image gray level co-occurrence matrix, calculating 8 features that respectively reflect texture uniformity, texture smoothness and texture linear correlation: dissimilar DIS, contrast CON, contrast INV, inverse disparity IDM, Angular second moment UNI, entropy ENT, correlation COR, maximum probability MAX.

The innovation of the present invention lies in: combining the damage representation and the wavelet energy difference to solve the limitations of the traditional feature method, and adding texture analysis to improve the consistency of the image evaluation algorithm and visual perception.

The theoretical basis of the present invention is: solve the limitation of directly mapping damage performance or statistical features and subjective evaluation in traditional methods, add wavelet energy difference, make up for the deficiency of damage performance statistics from a macroscopic point of view; simulate the image background in the human visual system For the influence of damage performance, the image texture analysis is proposed, which comes from the gray level co-occurrence matrix, and the image texture features are extracted from the three angles of texture uniformity, texture smoothness and texture linear correlation.

The present invention uses the statistical wavelet energy difference as a macroscopic feature to supplement the distortion feature detection, expresses the image damage from both the distortion and statistics aspects; then extracts the texture information to describe the relationship between the damage feature performance and the image background according to the visual perception characteristics. After a large number of experiments, the objective evaluation result of the present invention has a high consistency with the subjective evaluation, and consumes less time, which satisfies the universality of image quality evaluation application scenarios.

Description of drawings

Figure 1: System block diagram;

Figure 2: Tracking flowchart;

Figure 3: Principle block diagram of image quality evaluation combined with wavelet energy and texture analysis;

Figure 4: Block diagram of block effect detection principle;

Figure 5: Noise statistical distribution map;

Figure 6: Schematic diagram of wavelet subband energy sequence numbers;

Figure 7: Natural lossless image energy distribution map.

detailed description

The technical solution of the no-reference image quality evaluation method of the present invention is shown in FIG. 1 , FIG. 2 and FIG. 3 . The invention simulates the visual system of human eyes to detect the most intuitive damage performance. In order to solve the limitation of incomplete damage statistics, a wavelet energy model is established from the natural lossless image, and the energy difference of the damaged image is compared with it to make up for the undescribed damage performance as a macroscopic feature. Considering that the performance of damage features does not exist independently and has a great correlation with the image content, texture analysis is introduced to solve the problem of image content covering the above features. Here, the derived features of the gray level co-occurrence matrix are used to represent the texture information of the image and Content activity. Finally, the mapping relationship between the extracted distortion factor, energy, texture information and subjective score is learned by BP neural network, and the image quality evaluation method is obtained. The program includes the following technical content:

1. Wavelet transform: Efficiently extract information from signals through local transformations of space and frequency. It can perform multi-scale detailed analysis of functions or signals through operations such as stretching and translation. It is an ideal tool for time-frequency analysis and processing of signals.

2. Damage performance detection: Generally, for common image damage performance, such as block effect, noise, ringing effect and blur effect, detection is performed from the pixel domain or the whole world.

3. Texture features: represented by a gray-scale co-occurrence matrix, which can reflect pixel changes in the image with respect to direction and interval. According to the distribution of the co-occurrence matrix structure, the pixel arrangement and local features of the image are analyzed, and the derived features are used as the basic data representing the texture information of the analyzed image.

4. BP neural network: BP neural network is a multi-layer feed-forward artificial neural network using error backpropagation algorithm. The weight of the network is determined through learning, and it has strong self-adaptation to the environment and self-learning to external things. ability.

The specific scheme is as follows: use the LIVE image quality evaluation database and various distorted images in the TID2008 database to perform damage performance statistics, wavelet energy difference prediction, and texture analysis, and finally input the BP neural network to obtain an objective quality evaluation algorithm. Specific steps are as follows:

1. Statistics of damage characteristics: Starting from the human visual system, the most common damage manifestations are counted, including block effect, blur effect and noise.

1) Blocking effect: The key steps of blocking effect detection are shown in Figure 4. Traversing the wavelet LH and HL subbands of the test image in units of blocks, retaining block boundaries, eliminating pseudo-blocking effects by judging the chromaticity difference on both sides of the block boundary, reducing interference, and then dividing the blocking effect into the first Class One and Class Two. Finally, the block effect measurement value BLOCK is given by the block boundary wavelet coefficients of the first type of block effect and the second type of block effect, and the difference of the chrominance components of adjacent blocks.

2) The product of wavelet coefficients in different layers can effectively extract useful signals and suppress the interference of noise on blur effect measurement. The sharpness discrimination function constructed using the first and second layer wavelet coefficients, such as formula (1), where W ₁ represents the first layer wavelet subband, W ₂ represents the second layer wavelet subband, I=1,2 ,3 represent LH, HL and HH sub-bands respectively, and Blur is the blur effect detection result.

3) Traverse each pixel of the image and its 3*3 neighborhood, record the center pixel i and the average gray level j of its neighborhood as a feature pair (i, j), and accumulate the pair (i, j) in the whole The frequency of occurrence I(i, j) in an image. With i as the horizontal axis and j as the vertical axis, label all (i, j). As shown in Figure 5, the data distribution of the lossless image falls around y=x; the data distribution of the noise image is messy. The more serious the noise, the more scattered the data distribution. With y=x as the center, delineate the judgment noise standard with y offset from x by five pixels. The noise detection function is as formula (2), where Noise is the image noise detection result.

2. Wavelet energy difference measurement:

Calculate the wavelet subband energy, such as formula (3), where E is the subband energy, T is the total number of subband pixels, W is the subband coefficient, s is the scale number, and I is the direction number. The energy deviation caused when it approaches 0 and guarantees the linear distribution of sub-band energy, Φ is the adjustment factor, which is taken as 0.05. The average value of the LH and HL subband energies of the same scale is output as the primary energy Es. As shown in Figure 6, the sequence numbers of the wavelet sub-bands decomposed multiple times are arranged to obtain a smooth transition of the natural lossless image energy spectrum, and Figure 7 shows the wavelet energy spectrum of 29 lossless images.

The core idea of extracting energy difference is to use a part of the current damaged image to restore the energy distribution of its non-destructive image, so as to compare the difference between the original image and the damaged image. Here, the 1st and 2nd levels of the damage image energy spectrum are used to restore the 3rd to 8th level wavelet energy as the energy distribution of the original image. Specific steps are as follows:

1) Use matrix transformation to simulate the wavelet energy relationship in Figure 7, and use formula (4) to establish a natural image wavelet energy distribution model.

Hs is the prediction matrix of the sth scale; s is the number of scales, with values of 3, 2, and 1, representing the wavelet energy of the 3rd, 2nd, and 1st layers respectively; Is represents the subband energy of the sth layer of the natural lossless image; I ₄ is Subband energy at the 4th scale for natural lossless images.

2) Predict the subband energy Ps of the corresponding ideal image based on the subband energy D4 of the fourth scale of the distorted image and the Hs coefficient matrix obtained in formula (5).

P _s =D ₄ ·H _s (5)

3) When the image distortion is serious, the 1st and 2nd level energies on the 4th scale will also change greatly. If these values with large changes are used for prediction, errors will be transmitted. In order to prevent this situation, the predicted value D4 is adjusted, as shown in formula (6), where u4 is the sub-band mean value of the distorted image D4; U4 and avg are the natural The minimum and average values of the fourth-scale subband coefficients of the lossless image wavelet. .

Ifu4≤U4, _ThenD4 =avg (6)

Then according to the Ps and Es of the distorted image, the wavelet energy difference is obtained, such as formula (7).

D represents the wavelet energy difference of the image, s represents the wavelet energy series, Ps and Es are the restored energy and the original image energy respectively.

3. Texture features: the gray level co-occurrence matrix P _d (i, j) reflects the pixel changes in the image in terms of direction and interval, where P _d (i, j) is a fixed position away from the point of gray level i The relationship d=(Dx, Dy) is the probability of reaching the gray level j, L represents the gray level of the image, d represents the spatial position relationship between two pixels, and θ is the direction in which the gray level co-occurrence matrix is generated. According to the distribution of the co-occurrence matrix structure, the pixel arrangement and local features of the image are analyzed, and the derived features are used as the basic data representing the texture information of the analyzed image. Considering that there is a certain correlation among the derived features, in order to remove redundancy, we selected 8 features that respectively reflect texture uniformity, texture smoothness, and texture linear correlation: dissimilarity DIS, contrast CON, contrast INV, inverse The difference IDM, the second moment of the angle UNI, the entropy ENT, the correlation COR, and the maximum probability MAX are respectively obtained from formula (8) to formula (15).

MAX=max(Pd(i, _j )), (15)

Example 1

Based on the above-mentioned technical solution of the present invention, various features of the distorted image are counted by the evaluation system, and the input neural network is trained and processed by a computer to obtain an image evaluation method. Its method steps are:

1) Using the Itti visual saliency model to calculate the visual saliency map of the image, under a certain threshold condition, select the visual saliency region, and the present invention sets the threshold to 0.0095. And carry out 4-level wavelet transform on the image, and choose sym8 as the wavelet base.

2) Statistical damage performance is performed in the visually significant area of the image, combined with wavelet subband coefficients, image chromaticity and brightness to detect block effects, blur effects and noise factors, and use them as the basic features of the evaluation method.

3) Establish a wavelet lossless energy model through 29 natural images in the LIVE image library, calculate the damage energy spectrum of the test image, and use its first-level and second-level energies to predict the third-level to eighth level lossless of the image according to the lossless energy model Energy performance, compare the difference between the original energy and restored energy of the damaged image, and use this energy difference as a macroscopic feature to make up for the lack of loss feature extraction.

4) Statistical image gray level co-occurrence matrix, calculate 8 features that respectively reflect texture uniformity, texture smoothness and texture linear correlation: dissimilarity DIS, contrast CON, contrast INV, inverse difference IDM, angular second moment UNI, Entropy ENT, correlation COR, maximum probability MAX. Take them as features representing image texture information.

5) Establish a three-layer BP neural network, including an input layer, a hidden layer and an output layer, wherein the hidden layer contains 17 neurons. Three kinds of damage performance features (block effect, blur effect and noise factor), wavelet energy difference and 8 texture features are used as input nodes, and the subjective evaluation scores of test images are used as output nodes, and the mapping relationship is obtained through training. This relationship is the discovery A non-parametric image quality method for .

6) Use the test images of the LIVE image library and TID2008 image library to evaluate the image quality of this method, and use SROCC to represent the correlation between the evaluation score and the subjective score. A value of 0 represents no correlation at all, and a value of 1 represents complete correlation. An average of 1000 test results, and compared with the current excellent evaluation methods, as shown in the following table, it can be seen that the method of the present invention can accurately evaluate the image quality, conforms to the human visual system, and is not limited to a certain type of distortion, and has universal sex.

TABLE I. _{SROCC AMONG} 1000 _{TRIALS ON THE} LIVE IQA _DATABASE

JP2K JPEG noise Blurred FF PSNR 0.8990 0.8484 0.9835 0.8076 0.8986 SSIM 0.9510 0.9173 0.9697 0.9513 0.9555 VIF 0.9515 0.9104 0.9844 0.9722 0.9631 BIQI 0.8551 0.7767 0.9764 0.9258 0.7695 DIIVINE 0.9352 0.8921 0.9828 0.9551 0.9096 BLIINDS-II 0.9462 0.9350 0.9634 0.9336 0.8992 BRISQUE 0.9457 0.9250 0.9892 0.9511 0.9028 SSEQ 0.9420 0.9510 0.9784 0.9483 0.9035 this invention 0.9590 0.9814 0.9908 0.9816 0.9811

TABLE II.LCC _AMONG 1000 _{TRIALS ON THE} LIVE IQA _DATABASE

JP2K JPEG noise Blurred FF PSNR 0.8837 0.8515 0.9817 0.8006 0.8939 SSIM 0.9601 0.9485 0.9861 0.9537 0.9616 VIF 0.9664 0.9478 0.9924 0.9774 0.9698 BIQI 0.8414 0.7603 0.9732 0.9118 0.7342 DIIVINE 0.9409 0.9097 0.9744 0.9393 0.9128 BLIINDS-II 0.9493 0.9505 0.9614 0.9375 0.9079 BRISQUE 0.9472 0.9330 0.9883 0.9463 0.9142 SSEQ 0.9464 0.9702 0.9806 0.9607 0.9198 this invention 0.9540 0.9808 0.9939 0.9808 0.9810

TABLE III. SROCC _ACROSS 1000 _{TRIALS ON THE} TID _DATABASE

JP2K JPEG noise Blurred PSNR 0.8251 0.8754 0.9180 0.9344 SSIM 0.8780 0.9255 0.8120 0.9458 VIF 0.9703 0.9310 0.9131 0.9584 BIQI 0.9742 0.9253 0.8334 0.8472 DIIVINE 0.9070 0.8718 0.8342 0.8592 BLIINDS-II 0.9119 0.8383 0.7150 0.8261 BRISQUE 0.9044 0.9101 0.8237 0.8740 SSEQ 0.8468 0.8665 0.8011 0.8355 this invention 0.9672 0.9813 0.9074 0.9883

TABLE IV.LCC _ACROSS 1000 _{TRIALS ON THE} TID _DATABASE

JP2K JPEG noise Blurred PSNR 0.8855 0.8787 0.9427 0.9304 SSIM 0.8751 0.9372 0.8078 0.9384 VIF 0.9713 0.9733 0.9073 0.9423 BIQI 0.9824 0.9614 0.8192 0.8011 DIIVINE 0.8790 0.8998 0.8101 0.8407 BLIINDS-II 0.9199 0.8890 0.7145 0.8258 BRISQUE 0.9061 0.9503 0.8109 0.8739 SSEQ 0.9854 0.9812 0.8474 0.9406 this invention 0.9646 0.9826 0.9078 0.9870

Claims (4)

1. it is a kind of based on energy and the non-reference picture method for evaluating objective quality of texture analysis, it is characterised in that including following Step：

Step S1, the distorted image for obtaining；

Step S2, damage characteristic statistics is carried out to distorted image；

Step S3, calculated distortion image wavelet energy it is poor；

The textural characteristics of step S4, calculated distortion image；

Step S5, the impaired performance feature that foundation extraction is learnt using BP neural network, wavelet energy difference and textural characteristics and survey Attempt the mapping relations between the subjective assessment fraction of picture, obtain image quality evaluation.

2. as claimed in claim 1 based on energy and the non-reference picture method for evaluating objective quality of texture analysis, its feature It is that the damage characteristic statistics is included：Blocking effect, blurring effect and noise factor.

3. as claimed in claim 1 based on energy and the non-reference picture method for evaluating objective quality of texture analysis, its feature It is that step S3 is specially：The lossless energy model of small echo is set up to distorted image, test image is calculated and is damaged energy spectrum, profit The lossless energy performance of the third level～eight grade of the image is predicted according to lossless energy model with its first order, second level energy, The difference relatively damaged between the proper energy amount of image and reduction energy, obtains wavelet energy poor.

4. as claimed in claim 1 based on energy and the non-reference picture method for evaluating objective quality of texture analysis, its feature It is that step S4 is specially：Statistical picture gray level co-occurrence matrixes, calculate and reflect respectively the texture uniformity, texture smoothness and line 8 features of the linear dependence of reason：Different DIS, contrast C ON, contrast INV, unfavourable balance is away from IDM, angular second moment UNI, entropy ENT, correlation COR, maximum probability MAX.