Dual-Task Learning for Long-Range Classification in Single-Pixel Imaging Under Atmospheric Turbulence

Unlike traditional imaging, single-pixel imaging (SPI) exhibits greater resistance to atmospheric turbulence. Therefore, we use SPI for long-range classification, in which atmospheric turbulence often cause significant degradation in performance. We propose a dual-task learning method for SPI classification. Specifically, we design the Long-Range Dual-Task Single-Pixel Network (LR-DTSPNet) to perform object classification and image restoration simultaneously, enhancing the model’s generalization and robustness. Attention mechanisms and residual convolutions are used to strengthen feature modeling and improve classification performance on low-resolution images. To improve the efficiency of SPI, low-resolution objects are used in this work. Experimental results on the DOTA remote sensing dataset demonstrate that our method significantly outperforms conventional object classification approaches. Furthermore, our approach holds promise for delivering high-quality images that are applicable to other computer vision tasks.