Research Of Image Compression Based On Deep Learning

With the rapid development of the multimedia age and the continuous upgrade of smart devices,massive image data has brought great challenges to the transmission bandwidth of the network and the storage space of the device.Therefore,designing efficient and applicable image compression algorithms has become one of the research hotspots in the image field.Recently,deep learning theory has approached rapid progress.In the field of image compression,deep learning methods use neural network convolution operations to learn the internal features of images while reducing the image size.Compared with traditional compression algorithms,it does not need to construct features and coding manually,which is a potential image compression technology.The final object-oriented image is human.The visual quality of the human eye is the key to judge the effect of image reconstruction.The higher the degree of attention of the human eye in the image,the greater the impact on visual quality.However,the current deep learning-based image compression algorithms assign the same processing priority to all regions in the image during compression and reconstruction,and distribute the compression bits evenly for each pixel,lacking consideration of the human eye's visual attention mechanism.This article studies the visual attention mechanism in image compression.The main work is as follows:This paper designs and implements a deep learning image compression method based on semantic analysis.It consists of an image compression network and an image semantic analysis network.Among them,the image compression network uses a self-encoder model based on a neural network to implement image feature extraction and size compression through a convolutional neural network,and uses an upsampling technique to reconstruct the original picture.At the same time,the iterative scheme of the network is designed by using the memory characteristics of the recurrent neural network.The image compression ratio can be controlled by adjusting the number of iterations.The image semantic analysis network uses the network structure based on VGG16 to visualize the classification results of the network through class activation mapping to generate a semantic importance map of the input image.This paper proposes a compression bit allocation algorithm that combines a compression network with a semantic analysis network,calculates the compression level of each region of the image according to the semantic importance map,and compresses and reconstructs each region according to the corresponding level.Experiments show that the compression method proposed in this paper improves the visual quality of the reconstructed image with the same compression overhead.