Multi-task single-pixel sensing integrating imaging and classification with sorted block-measurement matrices

Most single-pixel sensing methods focus on either image reconstruction or target classification. This leads to long training time, limited functionality, and low utilization of sampling information. Conventional measurement matrices also have weak feature-capturing ability and high noise sensitivity, further reducing sampling efficiency and sensing stability. To address these challenges, we propose a multi-task single-pixel sensing method, termed Single-Pixel Sensing Method Based on Sorted Block-Measurement Matrix (SPS-SBM2). SPS-SBM2 integrates image reconstruction and target classification into a unified framework. It employs multi-task-oriented sorted block-measurement matrices to enhance the acquisition of target-related features. In the network model, it shares the encoded features to achieve collaborative computation of both tasks. Experiments are conducted on the modified National Institute of Standards and Technology and Terravic Research infrared datasets, as well as on real tests with mini Clothing and Mini Cup targets. The results show that SPS-SBM2 achieves reconstruction quality and classification accuracy comparable to single-task models under the same sampling times and interference conditions. When the sampling times range from 50 to 200, SPS-SBM2 performs classification and ×2/×4 reconstruction tasks. It reduces total computation time by 25.02%/30.77% for ×2 tasks and 37.50%/35.72% for ×4 tasks. SPS-SBM2 improves computational efficiency while maintaining robust reconstruction quality and classification accuracy. It provides an effective solution for multi-task single-pixel sensing.