Research On Image Compression For Salient Regions Protection

The rapid development of sensor technology makes the resolution of images higher and higher,which brings great challenges to the transmission,storage and processing of image data.It will require efficient image compression technology to support,and according to the characteristics of the human visual system,it is always a priority pay attention to the salient regions in the image,therefore,the paper proposed an research on image compression for salient regions protection.There are two problems studied in depth: First,due to the current model using neural networks for salient detection requires a clear boundary between the SR and the background,and requires high data during training,and there is a problem that the multi-targets of image cannot be perceived effectively;Secondly,based on question 1,no clear boundaries are required for the acquired salient map,so how to accurately separate the salient regions and the background region before compression to achieve the protection of the SR during compression is important.Aiming at the above problems,the paper built a Multi-scale Depth Feature Salient Regions(MS-DFSR)detection model,and proposed Image Compression for Salient Regions Protection(SRPIC)algorithm.First,the algorithm constructed a deep convolutional neural network to extract the depth features of the input image,and combined with global average pooling and class activation mapping,it achieved salient regions positioning work.More,used max-avg pooling and multi-scale concepts to detect multiple salient targets improved the model's effective perception,and data only needed image-level labels during model training.Secondly,this paper used the improved OSTU threshold algorithm to effectively separate the SR and the background region;Finally,the constructed MS-DFSR model is combined with coding compression technology to implement near-lossless compression on the SR,and the background used lossy coding compression to complete image compression and reconstruction for the protection of the salient regions.The experiments in this paper are divided into the effective verification of salient map segmentation algorithm,the verification of MS-DFSR detection model's perception of image content,the experiment of image compression at different bit rates,and the experiment of MS-DFSR combined with different encoding algorithms.The experiment was conducted on the Kodak Photo CD and the Pascal Voc dataset.The experimental results showed: the proposed SRPIC algorithm has obtained better results than traditional coding algorithms,which showed that the algorithm in this paper effectively perceives multiple salient regions in the image through the MS-DFSR detection model,and through the improved OSTU threshold algorithm,SR and background region are effectively separated,which reduced the loss of image contentduring the image compression process,thereby improved the quality of the reconstructed image after compression.There are 25 figures,4 tables,and 65 references in the paper.