DC-DETR:A Dual-Channel Cross Attention for Small Object Detection

DETR-based variant detection models have achieved remarkable results, but for small objects, there is still a problem of poor detection performance due to insufficient object feature information acquisition. Therefore, this paper proposes A Dual-Channel Cross Attention for Small Object Detection. Firstly, we introduce the cross-scale feature fusion module (CSFM) to reduce the loss of object feature information and introduce more object detail feature information by increasing the similarity of features at different scales. Secondly, we redesign the decoder layer structure and proposed a dual-channel cross attention (DCCA). A single channel is responsible for completing the classification task or the regression task. Depending on the task the channel is responsible for, the query can obtain the feature information that best satisfies the channel task, reducing the introduction of feature information irrelevant to the task. Finally, a large number of experiments are carried out on the small object dataset VisDrone 2019, and the results demonstrate the effectiveness of the proposed method, especially achieving gains of 3.6% and 3.9% in AP and APS, respectively. The entire model is referred to as DC-DETR. Compared with advanced DETR series models, DC-DETR is competitive in performance.