Feature pyramid of bi-directional stepped concatenation for small object detection

摘要

In recent years, great breakthroughs have been made in object detection. However, performance of the most algorithms declines significantly when detecting small objects in an image. Thus, multi-scale feature maps are often used to develop network variants to generate multi-scale representations. Existing feature pyramid-based methods tend to keep the number of channels consistent and fuse different scales by adding corresponding elements or channel concatenation, which is prone to lose low-level detailed feature information in feature fusion process. To solve this problem, a bi-directional stepped concatenation feature pyramid construction method based on SSD (BSCF-SSD) is proposed. The stepped concatenation strategy helps to avoid the loss of information at the current layer during the pyramid construction process, and the bi-directional tactic ensures the fusion features contain both detailed and semantic information. Furthermore, an attentional interaction module is designed to better aggregate dual-stream features to improve network performance. The proposed method improves the detection accuracy of small objects with less speed loss. Experimental results show that the method achieves 80.3% and 82.4% mAP on Pascal VOC2007 using VGG16 and Resnet50, respectively. On the special aviation object dataset UCAS-AOD, BSCF-SSD with VGG16 still achieves moderate improvement.