Optimizing object classification with single-pixel imaging and transformer networks

Traditional computer vision tasks rely on high-quality images for classification, but single-pixel imaging (SPI) uses compressive sensing with a bucket detector to simplify hardware and reduce costs. However, factors like atmospheric turbulence can degrade SPI image quality, lowering classification accuracy. To address this, a block-based single-pixel image-free transformer classification framework is proposed. It uses a learnable binary sampling matrix for flexible sampling during SPI and divides the measurements into blocks, which are processed by a transformer network. Each block is treated as a sequence element, allowing the transformer to capture spatial and contextual relationships. Experiments show strong robustness to atmospheric turbulence, achieving 87% classification accuracy even at a 1% sampling rate, making this method suitable for challenging environments.