B-PesNet: Smoothly Propagating Semantics for Robust and Reliable Multi-Scale Object Detection for Secure Systems

Abstract

Listen

Translate article

Multi-scale object detection is a research hotspot, and it has critical applications in many secure systems. Although the object detection algorithms have constantly been progressing recently, how to perform highly accurate and reliable multi-class object detection is still a challenging task due to the influence of many factors, such as the deformation and occlusion of the object in the actual scene. The more interference factors, the more complicated the semantic information, so we need a deeper network to extract deep information. However, deep neural networks often suffer from network degradation. To prevent the occurrence of degradation on deep neural networks, we put forth a new model using a newly-designed Pre-ReLU, which inserts a ReLU layer before the convolution layer for the sake of preventing network degradation and ensuring the performance of deep networks. This structure can transfer the semantic information more smoothly from the shallow to the deep layer. However, the deep networks will encounter not only degradation, but also a decline in efficiency. Therefore, to speed up the two-stage detector, we divide the feature map into many groups so as to diminish the number of parameters. Correspondingly, calculation speed has been enhanced, achieving a balance between speed and accuracy. Through mathematical demonstration, a Balanced Loss (BL) is proposed by a balance factor to decrease the weight of the negative sample during the training phase to balance the positives and negatives. Finally, our detector demonstrates rosy results in a range of experiments and gains an mAP of 73.38 on PASCAL VOC2007, which approaches the requirement of many security systems.