Image compression framework combining super-resolution and residual coding technology

[0018] The present invention will be further explained below in conjunction with the drawings:

[0019] figure 1 , An image compression framework combining super-resolution and residual coding technology, including the following steps:

[0020] (1) Iteratively back-project and down-sample the original image to be compressed to obtain a sampled small image;

[0021] (2) Compress the small image through the JPEG2000 standard at the encoding end, and transmit the code stream and denoising side information to the decoding end;

[0022] (3) Obtain a small decoded image at the decoding end, and reconstruct the small image into a large decoded image through super resolution based on code rate classification;

[0023] (4) Combine the denoising side information to suppress the compression noise in the decoded large image and improve the quality of the decoded large image;

[0024] (5) Calculate the residual error between the decoded large image and the original image to be compressed at the encoding end, and transmit the residual to the decoding end through residual encoding technology;

[0025] (6) Obtain the decoded residual image at the decoding end, and superimpose it with the large decoded image to recover the final decoded image.

[0026] Specifically, in the step (1), we first obtain a small image from the image to be compressed through the iterative back-projection down-sampling method ↓. After the sampled image X obtained by the sampling method ↓ is reconstructed to the original size by the super-resolution method ↑, the error with the original input image Y is as small as possible, and its mathematical model satisfies the formula (1). At the same time, in order to achieve fast convergence, ↓ can be specifically defined as bicubic downsampling, and ↑ can be specifically defined as simple bicubic interpolation.

[0027] X=arg min||Y-↑X|| (1)

[0028] In the step (2), the small image is compressed into a code stream by JPEG2000 at the encoding end and transmitted to the decoding end. At the same time, the same decoding and reconstruction process as step (3) is performed on the encoding end, and then the reconstructed image is used to suppress the compression noise by denoising technology. Since the denoising intensity is adjustable, the optimal denoising coefficient can be calculated in this process, and the coefficient will be transmitted to the decoding end to assist in improving the quality of the decoded image.

[0029] In the step (3), an A+ (Anchored neighborhood regression) algorithm based on code rate classification is used to super-resolution and reconstruct the decoded small image into the original size. The A+ algorithm itself is a pure super-resolution algorithm, and its dictionary learning model is quite similar to dictionary learning based on sparse representation. In essence, it is still learning the correspondence between high and low resolution images. Therefore, it is particularly important to accurately grasp the image degradation model. In this framework, the input image Y has undergone two degradation processes: (1) adaptive downsampling↓; (2) JPEG2000 compression loss J. Moreover, JPEG2000 has different levels of information loss at different bit rates. Therefore, this section proposes to classify the code rate with an interval of 0.1bpp, and train each type of conversion relationship Pi to satisfy formula (2), thereby constructing a "one-to-many combination of iterative back-projection sampling loss and compression loss" "Mapping model.

[0030] P i =arg min|Y-(↓JY)*P i || (2)

[0031] In the step (4), the denoising side information separated from the code stream obtains the optimal denoising coefficient for the current decoded large image. Use this coefficient to control the denoising intensity, thereby suppressing the compression noise of the decoded large image to the greatest extent, and improving the quality of the decoded large image.

[0032] In the step (5), by simulating the decoding, reconstruction and denoising process at the encoding end, the decoded big picture can be calculated, and the decoded big picture is subtracted from the original image to be compressed to construct a closed loop feedback loop. Since the compression process and the sampling process will inevitably cause a lot of loss of image information, this part of the loss information is defined as the residual between the decoded image and the original image. The residual image obtained through the closed-loop feedback loop usually does not have the common characteristics of natural images, so we use targeted residual coding technology to compress it into a bit stream and transmit it to the decoder. The residual image coding adopts a tree-structured coding mode, that is, the image is divided into 16×16 macroblocks and processed separately. According to its complex texture and other characteristics, each macroblock can be divided in 4 ways: one 16×16; two 16×8; two 8×16; four 8×8; for 8×8 sub-macroblocks the same can be done Make similar divisions. This segmentation method fully considers the correlation between each image data block. Such as figure 2 In tree structure coding, the regions with richer details are more detailed, while the smooth regions use larger coding units. This adaptive coding method can compress residual image information efficiently. At the same time, because the blocking can cause blockiness in the decoded image to a certain extent, the loop filter method is used in the residual coding to suppress the blockiness. Finally, the residual code stream is transmitted to the decoding end.

[0033] In the step (6), the residual code stream is separated at the decoding end to obtain a decoded residual image, and the blocking effect can be suppressed by the method of loop filtering. The decoded residual image is superimposed with the large decoded image obtained from the open-loop frame to obtain the final decoded image.

[0034] Randomly select two images ‘leaves’ and ‘butterfly’ in the test image library, test them with the above steps, and compare the rate-distortion performance with JPEG2000. Rate distortion is like image 3 and Figure 4 As shown, the horizontal axis is the code rate, the unit is bpp; the vertical axis is the peak signal-to-noise ratio (PSNR), the unit is dB. At the same bit rate, the higher the PSNR, the better the rate-distortion performance. Figure 5 Indicates the visual effect comparison of the ‘lena’ compression result between JPEG2000 and the present invention when the bit rate is 0.1bpp. Image 6 When the code rate is 0.6pp, the visual effect of JPEG2000 and the present invention on the result of "leaves" compression is compared. The experimental results have universal applicability to other images.