Enhancing object recognition for lensless cameras through PSF correction and feature loss

Lensless cameras have significant application prospects in multiple specialized fields due to their unique advantages of small size, low cost, and a naturally optical encryption effect. However, achieving high-accuracy object recognition in complex scenarios remains a challenging task. Firstly, many factors in the actual imaging process can induce the point spread function (PSF) distortion, resulting in a mismatch between the ideal model and the experiment. In this study, these factors are equivalently regarded as a blur kernel. Based on the singular value decomposition (SVD) results of the blur kernel, a lightweight and interpretable PSF correction network is proposed to counteract this blur kernel. Then, inspired by the class activation mapping (CAM) of classification networks, a feature loss function is proposed to indirectly achieve the reconstruction of local regions that play a crucial role in recognition by seeking feature layers alignment. Based on these, a series of incremental network models that integrate physical models and deep learning are designed, which achieve significantly better performance than other advanced methods on the cats_vs_dogs and ImageNet_10 datasets, and also show good generalization in the recognition test of real-world scenes. The proposed method significantly improves the object recognition performance of lensless cameras and has the potential for application in complex scenarios. The codes will be available at https://github.com/fylr/WienerNet_lensless.