Research On Methods Of Image Coding Based On Deep Learning

Image coding has always been a hot topic in the field of image processing.JPEG coding has played an i mportant role in image coding since its appearance.The traditional image coding methods are based on the pixel value.For the purpose of co mpression,they use the pixel value to predict,transfor m,quantify and entropy coding.As a result,these encoding methods do not take full account of the contents of the image.Therefore,we propose an image coding method based on deep learni ng to make classification of the image contents.Then we according to the categories of images to encode,so as to achieve the purpose of images compressing and we finally get a better performance on image coding.Firstly,in order to encode the image base d on the image content,a classification method based on image content is proposed in this paper.In the method,three different d eep learning network structure s,namely VGG-16 network,Res Net network and H-Re Net network structure model s,are used to extract image features.Features extracted by the network are mapped to the corresponding hash code by the iterative quantization method and the image content is classified by the hash code.For the H-Re Net network structure model,we need to mark 200 training images as network supervisi ng for training.In addition,when using the H-Re Net network structure model,we separate the image content into different classification according to mark ed the picture.Secondly,we apply these three classification methods base d on image content to i mage coding.The results of the classification and the image index blocks of each category are used as the dictionary in the image coding process.For the test image to encode,the feature is first extracted from the test image,the feature is mapped into the corresponding hash code,and the index image is generated according to the dictionary.Then,the residual image is calculated fro m the index image and the original i mage.At last,the residual image is compressed.Simulation experi ments show that our proposed deep-learningbased coding method has a better performance than the traditional coding method JPEG.