Research On Real-Time Image Compression Algorithm Based On Neural Network

With the development of The Times and the progress of technology,people's appreciation demand for hd picture quality is increasing day by day.At the same time,the large capacity information storage following the "information explosion" puts forward an urgent demand for efficient compression technology.Therefore,improving the image compression efficiency and real-time performance at the same time has become one of the hot spots in today's research.This paper mainly improves the real-time performance from the perspective of improving the image compression ratio,which means less bit consumption and less bandwidth consumption can reduce the network delay,speed up the transmission speed and enhance the real-time performance.This paper not only improves the real-time performance,but also pays attention to the restoration quality of video images,especially for low-bitrate images,which not only improves the compression ratio,but also ensures the compression and recovery effect of video images.Based on the above directions,this paper mainly does the following work:1.The adaptive optimization of Round function based on deep learning image compression framework is completed,and the image compression adaptive training based on image understanding is completed.2.Deep sparse autoencoder is used as the basis of compression framework:image segmentation and image clustering algorithm are used to preprocess the original data before model training.The optimal solution for finding the appropriate model in the solution space domain is selected by using particle swarm optimization algorithm.And use cross-validation method to test the stability and reliability of the improved MSAE model,and also design a new quantization method,propose a new compression feature prior modeling method and introduce rate-distortion optimization to achieve the "rate"-Distortion"trade-offs improve compression efficiency and increase compression ratio.3.The design and optimization of image interest region perception algorithm based on semantic understanding is completed.Aiming at the problems of low efficiency and weak anti-interference ability in the current image compression,a multi-structured region of interest is proposed.On the one hand,the effective conversion from low-level features to semantic features is realized,so that "reading the picture" is achieved;On the other hand,by effectively combining similar classes and sharing the same features across classes,the interference of insignificant"minor branches" of the same generic class is reduced,which effectively enhances the robustness of region perception and improves the recognition of target region perception.Accuracy.4.At the same time,in the feature extraction process,the acquired image directional features are changed to more significant edge features,and the saliency map generation focuses on the fusion of different scales and different dimensional features,introduces texture feature description,and integrates linear fusion methods to form features.In the process of the graph,a focus similar to human vision is added,and the optimization of the image perception mechanism of the image region is completed,so that the subsequent image restoration is "targeted",which effectively improves the image compression and restoration quality.5.Finally,the interest perception experiment proves that the multi-level region of interest sensing mechanism proposed in this paper can effectively resist many types of noise interference,even in the case of low-rate images,it can effectively reduce the "blockiness" and improve the target.Based on the recognition accuracy of the region of interest,the image compression recovery effect is improved.The compression framework based on a variety of standard image sets demonstrates that the compression framework proposed in this paper effectively improves the compression ratio,that is,effectively under the same visual perception.The bit consumption of the image size space is reduced,the real-time performance is improved,and the same visual effect and the same image parameter index value can effectively reduce the picture bit consumption by about 3%,achieving the "rate-distortion" trade-off and ensuring picture restoration.Quality,reaching the research expectation of improving real-time performance by increasing the compression ratio.