Research On Image Compression Method Based On Deep Convolutional Neural Network

With the continuous updating of cameras,the resolution of images is getting higher,and the storage is becoming larger,which brings huge challenges to the storage and transmission of images.In addition to considering increasing storage space and upgrading bandwidth,we should also focus on the image itself and develop image compression technology to reduce the storage space as much as possible while ensuring the images' quality.In recent years,image compression methods based deep learning have developed rapidly,and convolutional neural networks have become the mainstream compression method.The entire compression framework can be divided into four parts: encoder,quantizer,entropy encoder and decoder.The encoder uses convolutional layers to extract image features and learns compact representations.After the intermediate features passing through the quantizer and entropy encoder,a bit stream for storage and transmission is formed.The decoder uses a convolution layer to recover the image's information,and the objective function is optimized using rate-distortion function.Hyperspectral image is a new type of remote sensing image,which contains rich spatial,spectral,and radiation information.It is widely used in geological monitoring,environmental protection,marine meteorological detection and resource census.However,the large amount of hyperspectral image data brings great challenges to the transmission and storage.Therefore,compression of hyperspectral images is also very important.Many hyperspectral image compression methods have been proposed,but there is currently no related work on lossy compression of hyperspectral images based on deep learning.In view of the above problems,this paper studies the lossy compression methods of natural images and hyperspectral images based on end-to-end convolutional neural networks.The main work is as follows:1.The Coupled Squeeze-and-Excitation blocks based CNN for image compression is proposed.Considering the large number and strong correlation of feature maps in the natural image deep compression framework,this algorithm combined the channel attention mechanism to design the coupled SE module.The SEblock is used in the encoder to learn different weight responses for different feature maps,which improves the distribution of feature maps and increases the compression rate of entropy coding.The inversed SEblock(ISEblock)is used in the decoder which scaling method is changed to restore lost details.The algorithm is compared with Ballè's,JPEG,JPEG2000,and WebP methods on the public Kodak test set.Our method retains more textures and details at low bit rates,and the visual quality of image is significantly improved.2.A hyperspectral image compression method based on end-to-end residual convolutional network is proposed.The algorithm considers that the hyperspectral image data has local spacespectral correlation and large amount of data which is difficult to train.A full convolution framework combined with residual layers is used to form the encoder and decoder.The hyperspectral image is fedforward into the network as a 3D tensor.The convolution kernels are arranged in order from large to small strides in the encoder.The encoder uses convolutional layers and residual layers alternately,while retaining spectral information and extracting spatial spectral fusion features coincidentally,while reducing the dimension of the feature map.The decoder uses deconvolution layers,residual layers and subpixel layers alternately to recover the spatial and spectral information.The train set and test set of the algorithm are both extracted from the CAVE data set.Two traditional methods,e.g.,JPEG and JPEG2000,are used for comparison.The experimental results show that our algorithm gets higher quality and better spectral information at low bit rates,and has better processing effect on hyperspectral image compression.3.A hyperspectral image compression based on joint spectral-spatial deep convolutional network is proposed.The algorithm considers that not only the spatial redundancy,but also the correlation between the bands,need to be elimiated from the hyperspectral image data.The deep learning framework is used to compress the spectral and spatial information step by step.First,the spectral information is compressed using the 1 × 1 convolution kernels,and then the large convolution kernels are used to extract the spatial features and reduce the spatial resolution of the feature maps.The 1 × 1 deconvolution structure is used symmetrically at the decoder to recover the compressed spectral information.This method is compared with JPEG,JPEG2000 and the hyperspectral image compression method based on end-to-end residual convolutional network proposed in this thesis,respectively.Significant improvements in terms of PSNR and MS-SSIM can be observed.