Compressed Sensing Image Reconstruction With Multi-Scale Assisted Attention

With the continuous development of artificial intelligence and big data technology,more data needs to be processed in the process of information transmission.Therefore,the theory of compressed perception is proposed.However,compressive sensing theory still faces two major challenges when processing signals: On the one hand,due to the limitation of computer memory,in the compressed sensing sampling stage,the image needs to be divided into blocks,which destroys the complete structural information of the image,resulting in obvious block effects in the reconstructed image;On the other hand,In the compressed sensing reconstruction stage,the network model of the reconstruction algorithm is relatively simple,the image feature extraction is insufficient,and the quality of the reconstructed image is poor,especially for the compressed sensing reconstruction of areas with rich image detail information and complex texture structure.In view of the above problems,the research content of this paper is as follows:(1)Aiming at the problem of blockiness in reconstructed images,a multi-scale fully convolutional compressed sensing reconstruction algorithm based on channel attention(MSANet)is proposed.The algorithm first,set the convolution kernel,and the overlapping convolution kernel with a step size of 16 performs full convolution measurement on the complete image;then,deconvolves the measured feature map to obtain an initial reconstructed image with poor visual effect;Finally,the initial reconstructed image is sent to the deep reconstruction network to obtain a high-quality reconstructed image.A variety of natural images are simulated,and the results show that the MSANet algorithm avoids the block effect of the reconstructed image,improves the quality of the reconstructed image,and obtains a better-quality reconstructed image at a low sampling rate.(2)In order to further improve the restoration effect of images in areas with rich detailed information and complex texture structure,a deep reconstruction network based on multiscale assisted attention(MFDP-Net)is proposed.The network model includes a dual-path module,an assisted attention module,and a feature fusion module.The dual-path module uses different convolution kernels to extract the deep-level multi-scale features of the image,which enriches the feature information of the original image in the network;the assisting attention module consists of a local attention module and a non-local attention module.Cooperate with each other to restore the detailed structural information and complex texture information of the image;the feature fusion module is under the action of the classification function softmax,and the local attention module and non-local attention module can selectively restore image texture and structural information.Reconstruction simulation experiments are carried out on public datasets,and the results show that the visual effect of reconstructed images has been significantly improved.When the low sampling rate is 1%,the time spent to reconstruct an image is reduced by 2.618 s,and the objective evaluation index peak signal-to-noise ratio and structural similarity are improved by 0.56 d B and 0.0247.To sum up,this paper proposes a compressed sensing reconstruction algorithm based on multi-scale assisted attention to address the two challenges in the field of compressed sensing.In the sampling part,full convolutional sampling is used instead of random Gaussian matrix measurement to avoid the block effect of the image due to block measurement and obtain more complete original image information.In the reconstruction part,firstly,convolution kernels of different sizes are used to form a dual-path module to extract the deep-level multiscale image feature information of the image;then the feature information is input into the assisting attention module to recover the texture structure information of the complete image;finally,the complex texture and repetitive structure information restored by the attention module are fused to obtain the final reconstructed image.The experimental results show that the algorithm proposed in this paper has achieved remarkable results compared with other comparative algorithms,especially at low sampling rates,it shortens the time of image reconstruction and generates a reconstructed image with clearer visual effects.