Single-pixel Perception Method based on Sorting Pattern and Meta-classifier Correction

In the fixed pattern, there exists a limitation in the utilization of information in the encoding patterns. Specifically, the effectiveness of the information acquisition via compressed sensing varies across encoding patterns and different datasets. This not only leads to discrepancies in the feature information conveyed by light intensity values but also negatively impacts the efficiency of the perception system and affects model classification accuracy. To address these issues, this paper proposes an approach for sequential optimization of encoding patterns. The proposed scheme first combines encoding patterns with category representative images of the dataset under a fixed sort, obtaining intra-class significant feature information for sorting. Subsequently, inter-class feature differences are identified within the ranked intra-class significant features, and a new sorting pointer is derived to reorder the encoding patterns. Finally, image feature information is extracted using the reordered encoding patterns and is directly classified after being corrected by the meta-classifier. Simulation results demonstrate that with only 20 samples, the classifier achieves accuracies of 88.20% and 80.00% on the MNIST and FASHION MNIST datasets, respectively. In practical optical experiments, classification accuracies of 87.50% and 80.00% are also achieved for MNIST and FASHION MNIST, respectively, using only 20 samples. The proposed encoded patterns sequence optimization scheme achieves high classification performance under low sampling rates, establishing a foundation for further advancements in single-pixel perception classification.