Optimal Reconstruction of Single-Pixel Images through Feature Feedback Mechanism and Attention

Abstract

The single-pixel imaging technique can reconstruct high-quality images using only a bucket detector with no spatial resolution, and the image quality is degraded in order to meet the demands of real-time applications. According to some studies of algorithm performance, the network model performs differently in simulated and real-world experiments. We propose an end-to-end neural network capable of reconstructing 2D images from experimentally obtained 1D signals optimally. In order to improve the image quality of real-time single-pixel imaging, we built a feedback module in the hidden layer of the recurrent neural network to implement feature feedback. The feedback module fuses high-level features of undersampled images with low-level features through dense jump connections and multi-scale balanced attention modules to gradually optimize the feature extraction process and reconstruct high-quality images. In addition, we introduce a learning strategy that combines mean loss with frequency domain loss to improve the network’s ability to reconstruct complex undersampled images. In this paper, the factors that lead to the degradation of single-pixel imaging are analyzed, and a network degradation model suitable for physical imaging systems is designed. The experiment results indicate that the reconstructed images utilizing the proposed method have better quality metrics and visual effects than the excellent methods in the field of single-pixel imaging.