An end-to-end image compressive sensing algorithm based on attention neural networks

Compressive sensing (CS) algorithm with simple encoder is pretty suitable for small unmanned reconnaissance and strike systems. However, due to the problems of long reconstruction time, large amount of calculation, difficulty to apply hardware acceleration and neglecting optimization of the encoder, the traditional compressed sensing algorithm cannot meet the real-time requirements of unmanned systems. This paper proposes an end-to-end image compressive sensing algorithm based on attention neural network. The result of convolution sliding operating is chosen as block CS measurements to achieve end-to-end optimization and the reconstruction is performed by sub-pixel convolutional layer and visual attention module. Then, the transposed convolution and attention modules are used for supplementary information and the quality of reconstruction is improved by minimizing the L2 loss. Experimental results show that the proposed algorithm has a fast reconstruction speed without sacrificing reconstruction quality. The reconstruction time for an image with a size of 256 * 256 is 100ms, and it's 50ms for an image with a size of 128*128. Compared with the traditional algorithm NLR_CS, the speed of the article algorithm is 5000 times faster on an image with a size of 256*256 and 2000 times faster on an image with a size of 128*128. Moreover, the proposed algorithm has ability to be accelerated by hardware, excellent scalability of the target image resolution. All the modules and loss function used in the article jointly improved compression ratio of entropy coding.