Error-suppressed image-free single-pixel target recognition via Walsh-Hadamard transform and improved Hu invariant...

Accurate target recognition and robust feature extraction are critical in many applications. Existing image-free single-pixel detection (SPD) systems suffer from severe quantization errors induced by spatial dithering algorithms and poor robustness of the invariant moment to target pose variations and system parameter fluctuations, significantly limiting their applicability. In this paper, we propose an error-suppressed image-free SPD framework for robust target recognition. Walsh-Hadamard transform-based orthogonal modulation is employed to effectively eliminate binarization-induced errors, ensuring high-fidelity measurement acquisition. By exploiting the sparsity of geometric moment kernels in the Hadamard domain, we enable accurate high-order feature extraction with 470 Walsh-Hadamard coefficients, corresponding to an ultralow sampling ratio of 2.87% (for 128 × 128 pixels). An improved Hu invariant moments is developed to enhance robustness against both target pose variations (translation, rotation, scaling) and system parameter fluctuations (modulator size, detector gain). Simulations and experiments validate that the proposed SPD method based on Walsh-Hadamard transform maintains over 93.8% accuracy in global moment extraction. Consequently, the proposed framework achieves a classification accuracy of 92% under complex geometric and system fluctuations, providing a potential solution for practical target recognition and classification applications.