阅读说明：本技术 一种基于区域块级jnd预测的感知图像压缩方法 (Perceptual image compression method based on region block level JND prediction ) 是由 王瀚漓 田涛 于 2020-04-20 设计创作，主要内容包括：本发明涉及一种基于区域块级JND预测的感知图像压缩方法,包括以下步骤：1)根据数据集中的图像和对应的JND信息,利用大津阈值方法,生成区域块级JND值；2)根据生成的区域块级JND值,建立基于CNN的区域块级JND预测模型；3)将测试图像在多个固定的QF值下进行压缩,得到对应的多张失真图像,将全部失真图像分割为多个不重叠的区域块,并预测每个区域块的JND标签,最后采用标签处理方法获取每个区域块最终的JND值；4)根据目标压缩QF值和每个区域块最终的JND值,对测试图像进行预处理操作,选取区域块感知QF值中最大的作为压缩参数,并采用JPEG压缩预处理后的测试图像。与现有技术相比,本发明具有自适应预测、压缩质量好、压缩效率高等优点。(The invention relates to a perceptual image compression method based on region block level JND prediction, which comprises the following steps: 1) generating a JND value of a region block level by using a Otsu threshold method according to the images in the data set and the corresponding JND information; 2) establishing a CNN-based region block level JND prediction model according to the generated region block level JND value; 3) compressing the test image under a plurality of fixed QF values to obtain a plurality of corresponding distorted images, dividing all the distorted images into a plurality of non-overlapping area blocks, predicting a JND label of each area block, and finally acquiring a final JND value of each area block by adopting a label processing method; 4) and preprocessing the test image according to the target compressed QF value and the final JND value of each area block, selecting the largest area block sensing QF value as a compression parameter, and compressing the preprocessed test image by JPEG. Compared with the prior art, the method has the advantages of self-adaptive prediction, good compression quality, high compression efficiency and the like.)

1. A perceptual image compression method based on region block level JND prediction is characterized by comprising the following steps:

1) generating a JND value of a region block level by using a Otsu threshold method according to the images in the data set and the corresponding JND information;

2) establishing a CNN-based region block level JND prediction model according to the generated region block level JND value;

3) compressing the test image under a plurality of fixed QF values to obtain a plurality of corresponding distorted images, dividing all the distorted images into a plurality of non-overlapping area blocks, predicting a JND label of each area block, and finally acquiring a final JND value of each area block by adopting a label processing method;

4) and preprocessing the test image according to the target compressed QF value and the final JND value of each area block, selecting the largest area block sensing QF value as a compression parameter, and compressing the preprocessed test image by JPEG.

2. The method as claimed in claim 1, wherein the step 1) comprises the following steps:

11) for a smooth region, setting the region block level JND value in the smooth region to be consistent with the image level JND value, there are:

wherein S is_ITo test the picture level JND value of picture I,for the ith area block b in the test image I_iThe compression parameters of (1);

12) for the region with complex texture, the SSIM value of each region block is obtained under the image level JND value;

13) taking the quality difference delta SSIM of each region block under the continuous JND value as the intensity of each region block, and adaptively judging a distortion region under the current image level JND value by using the region block as a basic unit by using a greater fluid threshold method;

14) and circularly executing the steps 12) -13) until all image levels JND of each image are completely executed, and generating a final region block level JND value.

3. The method as claimed in claim 2, wherein in step 13), the expression of the quality difference Δ SSIM of each region block at consecutive JND values is:

wherein the content of the first and second substances,when the compression parameter isThen, the ith area block b in the test image I_iSSIM value of.

4. The method as claimed in claim 1, wherein the step 2) comprises the following steps:

21) sorting the generated JND values of the region block level from small to large, and marking training label values after classification to form a data set;

22) 90% of the area blocks in the data set were used for training and 10% for testing;

23) and training a region block level JND prediction model by adopting an AlexNet network.

5. The method as claimed in claim 4, wherein in step 23), in the training process of the JND prediction model, the image block size is set to 64 × 64, the initial learning rate is set to 0.001, the maximum iteration number is set to 250000, and the batch size is set to 64.

6. The method as claimed in claim 1, wherein the step 3) comprises the following steps:

31) compressing the test image under a plurality of fixed QF values to obtain a plurality of distorted images;

32) dividing all distorted images into a plurality of non-overlapping area blocks, and predicting a JND label of each area block by adopting an area block level JND prediction model;

33) when the prediction JND label of the area block at the same position of a plurality of distorted images meets the judgment formula L (b, q)_i)≥L(b,q_j),Step 34) is performed, if not, step 35) is performed, wherein q is performed_i、q_jQF values, b area blocks and L (-) prediction JND labels respectively;

34) the QF value corresponding to the JND label is the JND value of the current area block;

35) sorting the JND label values from small to large to enable the JND label values to meet a judgment formula in 33), and acquiring a corresponding QF value as the JND value of the current area block.

7. The method of claim 6, wherein the fixed QF values are 9 in total and are 15, 20, 25, 30, 35, 40, 45, 50, and 55, and the non-overlapping region blocks are 64 x 64 in size.

8. The method as claimed in claim 1, wherein the step 4) comprises the following steps:

41) obtaining the ith area block b of the test image I_iPredicted JND value of

wherein the content of the first and second substances,for the k-th JND value,

42) the target compressed QF value is preset toThe final adopted perceived QF valueComprises the following steps:

wherein the content of the first and second substances,

43) selecting the largest region block perception QF value as the compression parameter of the image levelThe expression is as follows:

wherein, NB^IThe number of the area blocks in the test image I;

44) if the JND value of the area block is smaller than the compression parameter of the image level, preprocessing the image;

45) after all DCT coefficients are preprocessed, inverse DCT transform operation is carried out to generate a preprocessed test image, and the image-level compression parameters in the step 43) are adoptedAnd (5) compressing by adopting standard JPEG to obtain a compressed image.

9. The method as claimed in claim 8, wherein in step 43), if there is a small JND value for the prediction of the partial region block, the method proceeds as follows:

10. the method as claimed in claim 9, wherein in step 44), the pre-processing of the image is specifically:

wherein the content of the first and second substances,is the DCT coefficient at the quantized position (m, n).

Technical Field

The invention relates to the field of image compression, in particular to a perceptual image compression method based on region block level JND prediction.

Background

With the development of social networks and multimedia technologies, a great deal of picture information is generated on the internet. Based on recent statistics, Instagram users upload approximately 9000 ten thousand pictures per day. Therefore, how to store and transmit the images is a very challenging task, and existing image compression standards such as JPEG, h.264 and HEVC all use PSNR and MSE as standards for measuring distortion, however, PSNR considers each pixel point as important in the calculation process and is inconsistent with the human eye visual system, and therefore, it is very important to research an image compression algorithm oriented to the human eye visual system.

Various approaches have been proposed to address this challenge, including JND-based approaches, attention model-based approaches, and the like. Currently, image/video perceptual compression methods based on JND (Just Noticeable distortion) are the focus of research. The existing JND models are mainly divided into two types: pixel-based domain and dct (discetecosine transform) based domain. The pixel domain-based method mainly considers the brightness masking effect and the contrast masking effect in a human eye visual system; the JND model of the DCT domain is added with a spatial contrast function on the basis of a pixel domain model. Although the existing perceptual coding model can reduce perceptual redundant information in coding to a certain extent, only limited visual characteristics are considered and the perceptual redundant information is not changed along with the change of quantization parameters, and the latest perceptual experiment shows that the perception of a human visual system on image quality presents a staircase shape and is not continuously changed, and each mutation point can be regarded as a JND value. However, for a single image, a large amount of subjective experiments are required to obtain a final JND value, and the final JND value cannot be applied in reality.

Disclosure of Invention

The present invention is directed to overcoming the above-mentioned drawbacks of the prior art and providing a perceptual image compression method based on region block level JND prediction.

The purpose of the invention can be realized by the following technical scheme:

a perceptual image compression method based on region block level JND prediction comprises the following steps:

1) generating a JND value of a region block level by using a Otsu threshold method according to the images in the data set and the corresponding JND information;

2) establishing a CNN-based region block level JND prediction model according to the generated region block level JND value;

3) compressing the test image under a plurality of fixed QF values to obtain a plurality of corresponding distorted images, dividing all the distorted images into a plurality of non-overlapping area blocks, predicting a JND label of each area block, and finally acquiring a final JND value of each area block by adopting a label processing method;

4) and preprocessing the test image according to the target compressed QF value and the final JND value of each area block, selecting the largest area block sensing QF value as a compression parameter, and compressing the preprocessed test image by JPEG.

The step 1) specifically comprises the following steps:

11) for a smooth region, setting the region block level JND value in the smooth region to be consistent with the image level JND value, there are:

wherein S is_ITo test the picture level JND value of picture I,for the ith area block b in the test image I_iThe compression parameters of (1);

12) for the region with complex texture, the SSIM value of each region block is obtained under the image level JND value;

13) taking the quality difference delta SSIM of each region block under the continuous JND value as the intensity of each region block, and adaptively judging a distortion region under the current image level JND value by using the region block as a basic unit by using a greater fluid threshold method;

14) and circularly executing the steps 12) -13) until all image levels JND of each image are completely executed, and generating a final region block level JND value.

In the step 13), the expression of the quality difference Δ SSIM of each region block under the continuous JND values is as follows:

wherein the content of the first and second substances,

when the compression parameter isThen, the ith area block b in the test image I_iSSIM value of.

The step 2) specifically comprises the following steps:

21) sorting the generated JND values of the region block level from small to large, and marking training label values after classification to form a data set;

22) 90% of the area blocks in the data set were used for training and 10% for testing;

23) and training a region block level JND prediction model by adopting an AlexNet network.

In the step 23), in the training process of the region block level JND prediction model, the size of the image block is set to 64 × 64, the initial learning rate is set to 0.001, the maximum iteration number is set to 250000, and the size of the batch size is set to 64.

The step 3) specifically comprises the following steps:

31) compressing the test image under a plurality of fixed QF values to obtain a plurality of distorted images;

32) dividing all distorted images into a plurality of non-overlapping area blocks, and predicting a JND label of each area block by adopting an area block level JND prediction model;

33) when the prediction JND label of the area block at the same position of the multiple distorted images meets the judgment formulaStep 34) is performed, if not, step 35) is performed, wherein q is performed_i、q_jQF values, b area blocks and L (-) prediction JND labels respectively;

34) the QF value corresponding to the JND label is the JND value of the current area block;

35) sorting the JND label values from small to large to enable the JND label values to meet a judgment formula in 33), and acquiring a corresponding QF value as the JND value of the current area block.

The fixed QF values are 9 in total, 15, 20, 25, 30, 35, 40, 45, 50 and 55, and the non-overlapping area blocks are 64 × 64 in size.

The step 4) specifically comprises the following steps:

41) obtaining the ith area block b of the test image I_iPredicted JND value ofThen there are:

wherein the content of the first and second substances,for the k-th JND value,

is the total number of predicted JND values;

42) the target compressed QF value is preset to

The final adopted perceived QF value

Comprises the following steps:

wherein the content of the first and second substances,is the 1 st JND value and is,is as follows

A JND value;

43) selecting the largest region block perception QF value as the compression parameter of the image levelThe expression is as follows:

wherein, NB^IThe number of the area blocks in the test image I;

44) if the JND value of the area block is smaller than the compression parameter of the image level, preprocessing the image;

45) after all DCT coefficients are preprocessed, inverse DCT transform operation is carried out to generate a preprocessed test image, and the image-level compression parameters in the step 43) are adoptedAnd (5) compressing by adopting standard JPEG to obtain a compressed image.

In the step 43), if the predicted JND value of the partial region block is small, the following steps are performed:

in the step 44), the preprocessing the image specifically includes:

wherein the content of the first and second substances,is the DCT coefficient at the quantized position (m, n).

Compared with the prior art, the invention has the following advantages:

firstly, adaptive prediction: the method and the device do not need to carry out subjective experiments, and adaptively predict the JND information of the block level of the region according to the content of the input test image, and are used for perceptual coding.

Secondly, the compression quality is good: the method avoids the condition that subjective quality is reduced due to the fact that the JND value of individual region prediction is low, protects the quality of the region with simple texture, improves compression efficiency, and obtains subjective quality similar to JPEG.

Thirdly, the compression efficiency is high: the method can adaptively predict JND information according to the content of the test image area block and improve the compression efficiency of the image, 10 images are selected from a Kodak data set to be used as tests, 3 QF values are respectively tested, the QF values are 75, 50 and 30 from high to low, compared with a JPEG algorithm, under the condition that subjective perception quality is similar, code rates are respectively saved by 43.91%, 18.76% and 13.11%, and the compression efficiency exceeds that of other similar models.

Drawings

Fig. 1 is a flow chart of model-based training and perceptual coding in the present invention, where fig. 1a is a flow chart of model-based training and fig. 1b is a flow chart of perceptual coding.

FIG. 2 is a schematic diagram of a selected test image.

FIG. 3 is a flow chart of the method of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments.