SSRDet: Small Object Detection Based on Feature Pyramid Network
Lijuan Zhang, Minhui Wang, Yutong Jiang, Dongming Li, Yue Zhou
Abstract
There are still many technological challenges in recognizing small objects in complex situations, and the performance of current detection algorithms for small objects is still unsatisfactory. Most existing methods mainly use feature pyramid networks to enrich shallow features using contextual features. However, due to the inconsistency of gradients between different layers of the feature pyramid network, the shallow features cannot be fully utilized resulting in the slow improvement of small object detection accuracy. To effectively improve the small object detection algorithm, we propose a new feature pyramid network-based small object detection algorithm, SSRDet. To effectively assign positive and negative sample labels and address the issue of sample scale imbalance, we first present RFLA. Then, to overcome the gradient inconsistency between various layers and enable the full utilization of the shallow features, we extend the feature pyramid network by including a scale enhancement module (SEM) and a scale selection module (SSM). Finally, we introduced the attention module (SPAM) to filter out the background noise in the shallow feature extraction to better extract small object features. We validated our method on VisDrone2019 and AI-TOD, and our method outperformed the state-of-the-art detectors.