Image Compression Technique Based On Convolutional Neural Network

In traditional image compression,the encoding stage generally includes Discrete Cosine Transform(DCT),Quantization,and entropy coding.The corresponding decoding processes are inverse discrete cosine variation,inverse quantization and inverse entropy coding.Compared with the traditional image compression method,the main research content of this thesis is the algorithm of compressing for RGB three-channel 8-bit quantized image using deep learning method.In the coding phase,a convolutional neural network(CNN)is used to make 16 times downsampling of the pictures which have been transformed to a data matrix,and the size of the data matrix of picture is effectively reduced.Then the data matrix obtained by downsampling is rounded.The compressed storage space of picture is effectively reduced.In the network training process,the distribution of feature is learned by constructing a super-parameter network.In the paper,the Laplace function is used to fit the feature distribution,and the ?and ?values of the Laplace distribution function are trained through the network.The probability distribution of each positive and negative 0.5 interval of feature is calculated by the fitted distribution function to adaptively perform context arithmetic coding.Corresponding to the coding operation,in the decoding stage,the method proposed by the paper will do arithmetic decoding and deconvolution operations for entropy-encoded feature.In order to enhance the learning ability of the network during the network training,feature rich operations are introduced.We added MSE values corresponding to features obtained by multiples' different downsampling to the training loss function.This operate constrains the performance of the feature obtained by each downsampling on the convolution operation,and the learning speed of the network is accelerated and improving the performance of the network is improved effectively.Moreover,in order to further enhance the effect of the convolutional network,a residual compensation block is introduced in the convolutional network by borrowing the idea of residual network.The final compression effect can be significantly higher than traditional compression algorithms such as JPEG2000,JPEG,webp,etc.The PSNR indicator can be about 0.5-1dB higher than JPEG2000,and on the MS-SSIM-DB index about 0.2-0.5dB higher than JPEG2000.