As traditional RGB cameras cannot perform well under weak light in the darkness and poor weather conditions, thermal cameras have become an essential component of edge systems. This paper proposes a lightweight, faster binarized R-CNN network (a state-of-the-art instance segmentation model), called BiThermalNet, for thermal object detection with high detection capabilities and lower memory usage. It designs a new Region Proposal Network(RPN) structure with a binary neural network (BNN) to lower model size by 16%, having higher accuracy performance. BiThermalNet adds novel-designed residual gates to maximum information entropy and offers channel-wise weight and bias to reduce errors from binarization. The extensive experiments on different thermal datasets (such as Dogs&People Thermal Dataset, UNIRI-TID) confirm that BiThermalNet can outperform traditional faster R-CNN by sizable margins with smaller models size. Moreover, a comparative analysis of the proposed methods on thermal images will also be presented.
|