Research On Image Compression Algorithm Based On Deep Learning

With the rapid development of technology industries such as big data,cloud computing and the Internet of Things,the explosive growth of data leads to increasing pressure of data storage.As images are the main carrier of information,it is significant to design efficient image compression algorithms to reduce the compression storage of images and improve the quality of compressed images.While traditional image compression algorithms such as JPEG,JPEG2000,and BPG have been continuously developed and widely used,the development of hardware equipment has greatly improved the processing speed and storage capacity of computers,benefitting the development and implementation of image compression algorithms research based on neural network.Furthermore,with the development of neural networks,it has been a trend that model structure is getting deeper,wider,and more densely connected.The storage of neural network model is also getting larger and larger,which caused difficulties in the deployment of mobile terminals such as mobile phone.Model compression based on deep learning is a popular research direction in recent years.This thesis proposes a fully convolutional auto encoder deep neural network for lossy image compression applications.This network model can compress images of different sizes through the end-to-end.The network framework includes five parts: encoder,quantizer,entropy coding,entropy decoding,and decoder.In the encoder and decoder module,we build encoder and decoder network based on the residual convolution block to avoid the performance degradation caused by the deepening of the neural network layer.In the decoder network,we use resize convolution instead of universal deconvolution to up-samples the feature vector,which is for avoiding the model capacity reduction caused by the checkerboard effect.In the quantizer,we use soft-margin quantization instead of uniform quantization,because the jump discontinuities caused by round quantization will obstruct gradient descent for optimization.In the entropy encoder,we encode the quantized feature vector symbols with arithmetic coding for further reducing the storage of the feature vectors,thereby improving the compression ratio of the image.For the model optimization,we use different mixed perceptual loss functions to improve the image quality generated by the model.Compared with the traditional image compression standards JPEG and JPEG2000,the experimental results show that our end-to-end fully convolutional auto encoder network has better performance on MSSSIM index.Based on the above framework of image compression neural network,we propose a lightweight image compression neural network model.We use parameter quantization to quantize the model weights from32-bit floating point to 8-bit integer.The quantized model accuracy loss is only 1.68%,and the compressed model storage occupies only a quarter of the original model.