Single-pixel image classification via optical compression encoding and all-optical diffraction decoding

Single-pixel sensing is a perception technology that directly decouples advanced semantic inference results from one-dimensional collected data without the need for multi-dimensional image reconstruction. Compared with the traditional imaging-first and then perception technical path, it significantly enhances perception efficiency. However, electrical neural network enabled single-pixel sensing methods introduce latency and impose a significant computational burden, leading to increased power consumption. In this paper, we propose a single-pixel image classification via optical compressive encoding and all-optical diffraction decoding. This architecture optically computes compressed features from the input, which are then processed by an all-optical diffractive decoder to yield classification outcomes. It incorporates compressive sensing into a diffraction optical neural network (DONN), synergizing the benefits of conventional DONNs with those of compressed encoding to surmount the modulation limitations inherent in DONNs. Experimental results demonstrate that the proposed framework consistently outperforms traditional DONN using only one diffractive layer in noisy or low-light conditions. By transmitting compressed features rather than the original images in the optical domain, the proposed framework provides strong privacy protection, making it highly potential for applications such as medical imaging and biometric authentication that have high privacy requirements.