Maritime Small Object Detection in UAV Images based on Improved YOLOv5s

As countries around the world emphasize and develop marine resources, the frequency of maritime accidents is increasing year by year. However, most of the current object detection algorithms are unable to accomplish both high detection accuracy and smaller parameter sizes in Unmanned Aerial Vehicle (UAV) images for small objects. To improve the accuracy of small object detection, the backbone network incorporating the attention mechanism is proposed and the original spatial pyramid pooling (SPP) structure is improved to enhance the ability of the model to extract object features. The single detection head in the YOLOv5 model is replaced by a decoupled detection head to improve the accuracy and robustness of the detection. In addition, to reduce the parameters of the model, the neck network of the original model is also lightened and improved in this paper. Based on the above improvements, the SCD-YOLO algorithm is proposed in this paper, which is specialized for small object detection algorithm in maritime environments in UAV images. Finally, the improved algorithm is thoroughly analyzed and evaluated in this paper on self-made marine datasets. The ablation experiments and comparative experiments show that the detection performance of the SCD-YOLO algorithm is significantly improved, and the value of mAP is increased from 87.2% to 91.0%.