CSDet: A Compressed Sensing Object Detection Architecture With Lightweight Networks

Compressive learning (CL) is an emerging framework that enables machine learning inference tasks to be performed directly in the measurements of compressed sensing (CS), which can reduce memory usage and improve compute and transmission efficiency. However, as a typical CL task, CS object detection is still challenging due to severe feature loss during sampling and is hard for deployment of terminal equipment with limited computing resources. In recent years, one-stage object detection models have achieved remarkable detection speed and accuracy in the image domain. In this paper, we introduce a CS object detection architecture (CSDet), which can perform object detection in the compressed domain and reconstruct color images with high quality at arbitrary sampling rate. CSDet is composed of a CS object detection module (CSDM) and a CS reconstruction module (CSRM) with lightweight networks. The CSDM integrates an optimizable joint multi-channel sampling matrix and a lightweight one-stage object detection network, which are trained jointly. The CSRM employs a joint multi-channel global reconstruction network. Experimental validation on the MS-COCO-Person dataset demonstrates that the proposed CS compressed domain object detection method achieves an 80% decrease in floating point of operations with even better accuracy compared to the image domain method. Meanwhile, the method exhibits notable improvements in reconstruction quality and speed compared to recent approaches.