Ambiguity-feedback active pairwise enhanced sensing for image-free single-pixel recognition

Image-free single-pixel recognition infers semantic labels directly from spatially compressed measurements without reconstructing images, offering an efficient sensing paradigm that dramatically reduces the acquisition-to-decision overhead. Existing learning-based methods typically rely on globally fixed sensing patterns optimized for average discrimination across all samples, which limits recognition accuracy. We propose an ambiguity-feedback active pairwise (AFAP) enhanced sensing method that uses the initial measurements as feedback to adaptively select subsequent sensing patterns, making the sensing process adaptive and better optimized for each individual sample. In Stage-I, global sensing patterns shared across samples acquire initial measurements, which are jointly tokenized with their associated patterns and processed through token interactions to estimate class confidence and identify the most plausible ambiguous class pair. The predicted pair then guides the selection of pair-specific sensing patterns specialized for distinguishing the two candidates, focusing subsequent measurements on the identified ambiguity. In Stage-II, class–query cross-attention extracts and compares candidate-specific evidence from the pair-conditioned measurement tokens to produce the final prediction. Experiments across four datasets demonstrate that AFAP consistently outperforms existing image-free recognition methods and a matched global-pattern baseline under low measurement budgets.