CBS-YOLO for ship target detection in complex marine environments

In intricate marine surroundings, problems such as target blockage and diminutive object dimensions frequently give rise to poor accuracy and low efficiency in conventional ship detection approaches. To tackle these hurdles, a ship target detection algorithm founded on CBS-YOLO and customized for such complex scenarios is proposed. Firstly, a small object detection layer is embedded into YOLOv5, which markedly enhances the model’s accuracy when it comes to detecting minute objects within images. Secondly, the CBAM attention mechanism is incorporated into both the Backbone and Neck segments, with the intention of strengthening the model’s capability to portray image features. Thirdly, the BiFPN is employed in the neck network, enabling the comprehensive utilization of feature information at diverse levels and consequently augmenting object detection performance. Through a succession of comparative experiments, it has been corroborated that the optimized CBS-YOLO algorithm surpasses the YOLOv5s detection algorithm by conspicuous margins. Precisely, it attains a 3.1% increment in accuracy, a 3.7% growth in mAP_50:95, and a 2.7% enhancement in recall. These outcomes evidently manifest that CBS-YOLO demonstrates outstanding performance in maritime ship target detection.