DCA-YOLO: Non-Prominent Feature Object Detection Using the Dynamic Convolution Attention YOLO Model

Camouflaged object detection is one of the most challenging problems in computer vision. Object detection depends on feature extraction and processing, so detecting targets with less distinct features is difficult. To enrich the experiments regarding the performance of such objects on YOLO-related models, this paper conducts comparative experiments on a non-prominent feature objects dataset and explores the impact of dynamic convolution and attention mechanisms on the detection capability of YOLO models. This paper proposed a YOLO-based model called DCA-YOLO, which combines dynamic convolution and attention mechanisms. Dynamic convolution provides flexible and powerful feature extraction capabilities, while attentional scale sequence fusion further improves detection performance through effective feature fusion and weight assignment. Our YOLO based model with the proposed components, performs well in diverse scenarios. The proposed model and other 4 YOLO models were tested on certain real-world dataset about targets disguised in the background. Results showed that our proposed model can effectively detect non-distinctive targets.