A Model Of Optimizing Saliency Detection And Its Application In Image Compression

The development of the Internet makes the image data increase rapidly.The mas-sive image data not only means more storage equipment,but also challenges the limited computing resources.It requires the efficient computer vision algorithm for process-ing and storing the image data.When facing a large number of external information,the human visual system filters visual information by selective attention mechanism to reduce the computation of brain,which is beneficial to understand the surrounding environment.Inspired by the characteristics of human visual system,researchers have introduced visual saliency into the field of computer vision and proposed many saliency detection models.With the development of shooting equipment,saliency detection is no longer confined to 2D images,and the saliency detection models which combine depth information and light field data are also increasing.On the basis of summarizing and analyzing the existing work,this thesis does the research on the optimization and application of the saliency detection.Computer vision algorithms can allocate computing resources based on the results of saliency detection,so as to achieve efficient data processing.However,the quality of saliency map affects the performance of the subsequent algorithm.In this paper,a universal framework is proposed to improve the saliency map of existing models.We propose an iterative algorithm which considers the local and global information in the optimization stage,the saliency consistency of foreground regions is enhanced by the iteratively updating strategy.Besides,the adaptive denoising in the preprocessing stage is integrated into the universal framework.The experiment shows that this model can effectively enhance existing saliency detection models based on different data types.And it has better adaptability and performance than the similar work.The image compression algorithm can use saliency map to optimize bit allocation,which can improve the compression efficiency and save storage space.A lossy image compression model is proposed in this paper to avoid distortion existing in traditional codec with low bit rate.The coder and decoder are constructed by neural network,and the number of bits in different image areas is adjusted using saliency map.We propose a multi branch network using the recurrent neural network and stacks the same struc-ture network to implement residual coding,the hidden state of recurrent neural network passes the relevant information between different iterations.In the coding phase,the bits of non-salient regions are reduced,that further improves the compression rate while maintaining the visual quality of reconstructed image's salient regions.The experiments show that images which are reconstructed by the proposed image compression model have better visual quality with low bit rate.