High-Precision Binary Object Detector Based on a BSF-XNOR Convolutional Layer

Recently, building an efficient and robust model for object detection has attracted the attention of vision community. Although binary networks have a fast inference speed, they cannot be used directly on mobile devices such as unmanned aerial vehicles (UAVs) because their low accuracy. Different from improving accuracy network by adjusting structure or update gradient, we propose improved neural based block scaling factor XNOR (BSF-XNOR) convolutional layer. In addition, two-level densely connected structure, which further enhances layer's feature representation capabilities. Experiments using TensorFlow framework prove effectiveness our algorithm in Compared with original standard network, mean average precision (mAP) detected PASCAL VOC dataset was improved. The experimental results VisDrone2019 UAV confirm that method achieves better balance between speed than previous methods. Our aims to guide deploy high-precision embedded device solves problem low-precision networks.