FSPI-R&D: Joint Reconstruction and Detection to Enhance the Object Detection Precision of Fourier Single-Pixel Imaging

Abstract

Compared with conventional imaging methods, Fourier single-pixel imaging (FSPI) has efficient noise immunity, wide spectral coverage, non-local imaging ability and long imaging range. Leveraging FSPI for object detection holds promising applications. However, considering the imaging speed of FSPI, it is necessary to obtain the imaging scene information in the under-sampling condition. The quality of FSPI reconstructions with low sampling rate is low and utilizing low quality reconstruction results for object detection will lead to low detection accuracy. To address the challenges, a joint reconstruction-detection framework based on FSPI is proposed. The Spatial-Adaptive Reconstruction Network (SARN) is designed to rapidly reconstruct the low-sampling rate image to improve the image quality. The Mixed Spatial Pyramid Pooling Fast (MSPPF) and Deformable Convolution (DCN) are integrated into the object detection network to improve the detection performance. Through joint training strategy, the synergy between high-level and low-level vision tasks is strengthened, so as to further improve the detection accuracy. Numerical simulations and real-world experiments show that the proposed method not only improves the quality of FSPI reconstruction with low sampling rate, but also significantly improves the performance of object detection tasks.

Keywords

• Fourier single-pixel imaging (FSPI)
• object detection
• image reconstruction
• joint training