A Compressive Sensing Network For ROI-aware Recovery With Block-based Measurement

Compressed sensing(CS)theory breaks the limitation of Nyquist theory and unifies the sampling and compression of the signal.Because CS technology reduces the consumption of resources greatly in signal acquisition,storage and transmission,it is widely used in many fields and has high value in research and practicality.Block-based CS is widely used in image processing tasks because of low storage burden and complexity in signal reconstruction.However,there are some problems such as low reconstruction speed,discontinuous structure,block effect and low reconstruction quality.Aiming at above problems,we propose an image compression sensing reconstruction deep neural network based on block sampling,which improvs the reconstruction quality greatly.Meanwhile,we futher design a ROIaware compressive sensing network.The designed network detects the region of interest in the scene adaptively and allocates more resources to these regions to improve the quality.The main work contains two aspects as follows:Focus on the problem of block effect,we propose a full image recovery idea and design the corresponding image recovery network based on block compressive sensing.Most traditional CS methods recover images block by block,which ignores the connection of image blocks and leads to block effect.In this paper,we adapt the idea of block-based sampling and full image recovery.The signal is measured by traditional block-based method.While in the recovery stage,to reconstruct the structure information of the image in the reconstruction process and avoid block effect,we build a fully convolutional network to recover the full image from all measurements.This method improves the image quality and avoids the stitching of image blocks.In addition,we obtain adaptive measurement matrix from large data by training.Compared with the traditional random measurement method,the adaptive measurement matrix improve the sampling efficiency greatly and further improve the reconstructed quality.Then,we further propose ROI-enhanced CS method based on full image recovery network,which optimizes the distribution strategy of measurement rate in different regions and improve the reconstruction quality in ROI.The sampling of the scene consists of two phases,the scene is sampled with a low measurement rate and reconstructed firstly.Then ROI location is obtained by using saliency object detection algorithm.More measurement rates are allocated to ROI and implementing the second sampling stage.By combining the measurements from two sampling stages,the multi-resolution image can be reconstructed.In addition,we improve MSE loss function to enhance the performance of network for ROI recovery,and further improve the reconstruction quality.The experimental results show that the proposed method has a significant improvement on the reconstruction quality of ROI.In summary,based on the theory of block-based compressive sensing,we design two kinds of compressed sensing networks by using deep learning technology.Our work improves and supplements the existing algorithms and making some progress in the perception of compressive sensing reconstruction.