Image Compression Method Based On Content-Aware And Multiscale Reconstruction

Image compression is an important research direction in the field of computer vision.It refers to a technology that represents image data with fewer bits while maintaining a certain image quality,also known as image coding.Due to the information explosion brought by the information age,which makes the scale of data present a geometric growth trend.In the face of such massive data,it is particularly important to develop a more efficient image compression technology for the storage,transmission and processing of image information.The traditional image compression method relies on a manually-made encoder/decoder framework,which uses fixed transformations to reduce redundant information and doesn't adapt to the image content in any way.Meanwhile,the method will appear fuzzy and artifact in very low bit.Therefore,this thesis adopts the deep learning method,which applies the generative adversarial network to the encoder/decoder framework of image compression and builds a novel deep image compression system.It not only achieve the compression of different content of the image to different degrees,but also preserves more details for the area with complex or important structure as far as possible,so as to improve the perceptual quality of the reconstructed image.The main work and innovative points of this thesis are as follows:(1)Regarding content-aware,we consider that one generally more inclined to be interested in only important areas in an image.The difficulty of compressing different parts of the image should be different.Smooth areas in the image should be easier to compress than areas with obvious objects or rich textures.Therefore,this thesis uses the GAN-based compression model as the baseline,and designs a content-aware importance subnet basing on the convolutional block attention module.This network can generate content-aware mask according to the importance of image information to guide the adaptive space allocation of bits.In addition,the introduction of the atrous-residual network on the basis of CBAM can efficiently extract and compress the deep and typical features in the image.In this way,the model can focus on more important features and suppress unnecessary features.While aggregating multi-scale feature information,the extracted intermediate features can be effectively emphasized or compressed to improve the overall performance of the model.At the same time,aiming at the problem that the target content in many image accounts for a small proportion of image pixels,we propose that training the network with content perception loss,so that the problem of serious imbalance between positive and negative samples in training can be effectively solved.The experimental results show that the compression reconstruction effect and generalization of the model are good,and the two indexes of MS-SSIM and PNSR are significantly better than JPEG 2000 and several CNNbased methods.(2)In the aspect of multiscale reconstruction,in view of the image's unimportant area due to insufficient bit allocation leads to serious distortion at the low bpp(rate)and the problem of the poorer compression and reconstruction for the image of high similarity and the highresolution,this article put forward by training GAN to reconstruct the unimportant regions,and the application of the idea of multiscale pyramid decomposition in the discriminator.Moreover,we use more efficient convolutional networks and minimize the distortion of each scale to capture the feature information of the image simultaneously in the local and global scope.Furthermore,we take advantage of the manner of progressive reconstruction from low resolution to high resolution to reduce the complexity of the image content,so as to achieve the compression and reconstruction of high-resolution images under the low rate.In addition,Adam optimizer is used to optimize and update model in an end-to-end manner,and the model is trained with adversarial loss and feature-matched loss,which not only avoids the problem of gradient disappearance in training,but also restores the texture and structure of the reconstructed images to be more realistic and natural.The experimental results show that the reconstruction quality of this model under low bpp is more favored by users,the image content is clearer,and the user's satisfaction is higher;the two indicators of MS-SSIM and PNSR are 98.5% and 35.4,respectively,and the performance is significantly better than other kinds of GAN-based method.