Image-free inference for 3D object classification via multi-view single-pixel detection

Image-free classification enables direct target recognition from single-pixel measurements without image reconstruction during inference, offering high sensing efficiency and low computational cost. However, existing approaches mainly rely on single-view observations and therefore cannot fully exploit the intrinsic three-dimensional (3D) structural information of real-world objects. Here, we propose a multi-view single-pixel detection framework for image-free inference of 3D object classes. By leveraging optical reciprocity, synchronized detectors acquire complementary view-dependent responses under identical spatial modulation patterns, enabling efficient extraction and fusion of stereoscopic features. A hybrid optical-neural architecture is developed, in which convolution kernels learned offline are converted into binary modulation patterns. A multi-view aircraft dataset containing five categories of 3D targets was established using a synchronized three-detector single-pixel imaging system. Experiments on the dataset demonstrate that multi-view fusion consistently outperforms single-view baselines. Statistical evaluation over 25 independent training runs shows that the proposed framework achieves 99.97% ± 0.06% with seven patterns. This work provides an effective route toward high-speed image-free 3D object perception in resource-constrained sensing applications.