Scale-Aware Regional Collective Feature Enhancement Network for Scene Object Detection

Yiyao Li, Jin Liu, Zhenyu Gao

Abstract: Deep learning-based methods are prevalent in object detection. These methods usually utilize a complex end-to-end network to locate and classify objects without overall architectures optimization, which leads to some common problems such as inaccurate multi-scale object detection, poor feature extraction robustness, and unprecise bounding box regression. To alleviate these issues, in this article, we proposed a novel Scale-Aware Multi-Branch Region Collective Feature Enhancement Network (SARCF-NET) to optimize the network from three aspects to improve the overall performance of scene object detection: detection framework, feature extraction, and bounding box regression. SARCF-NET uses multi-branch and the sale-aware training scheme to adapt multi-scale objects. Meanwhile, regional collective feature units are added to each branch to utilize both senior and junior features sufficiently. Besides, a novel Regression-Sensitive Loss function is proposed to adjust the bounding box coordinates automatically without extra computation. Extensive experiments are conducted on COCO2017 and CIFAR10, and the results demonstrate that our model is improved in accuracy while the number of parameters is significantly reduced compared with state-of-the-art models.