An Adaptive Enhanced Residual Attention Method for Small Object Detection

Introduction: In the field of object detection, small object detection has been a challenging problem. Existing CenterNet mainly focuses on deep features while ignoring shallow features, and there is also strong similarity object interference, which leads to insufficient detection ability for small objects. To solve these problems, this paper proposes a small object detection method AR-CenterNet based on the Adaptive Enhanced Context Model and Residual Attention Mechanism. Methods: Firstly, to enhance the feature representation capability, an Adaptive Enhanced Context Model (AEC) is designed, which balances the contextual information of shallow features at different scales. context information of different scales of shallow features and fuses them with deeper features by different scales of convolutional expansion. In addition, to reduce the influence of strong interference objects, the RAM (Residual Attention Mechanism) module is proposed, which reduces the interference of surrounding features by introducing the residual attention mechanism, recognizes the channel and spatial attributes of the small object features by using the coordinate attention mechanism, and preserves the original feature information through the jump connection. Results: Experimental results show that AR-CenterNet achieves excellent small object detection performance on PASCAL VOC, RSOD, and KITTI datasets. Conclusion: The method has important application value in the field of intelligent transportation.