雾环境下的船舶目标检测研究OA北大核心CSTPCD

To effectively avoid ship collisions and overcome difficulty in ship identification, and low detection accuracy in foggy environments, this paper first builds a dataset for ship detection in foggy environments. Then, improvements are made on the YOLOv5 model. Specifically, the GSConv module is employed to replace the CBS module in the Head section to make the depth separable convolution closer to the separable convolution, improving model accuracy. The Slim-Neck paradigm is introduced to further boost the model's average accuracy and reduce computational complexity. Additionally, the binary cross-entropy loss function is replaced with a polynomial loss function to enhance the model's accuracy. The SIoU Loss is introduced to address the deficiency in direction between the real box and predicted box, thereby improving training speed and inference accuracy. Our experimental results show the model reaches 95. 7％ in mAP0. 5 score, 0. 9％ higher than that of the baseline YOLOv5 model while the FLOPs is down by 2. 1G. Our study demonstrates the fog-based ship detection model achievesa better accuracy and has a lighter model structure and thus it has great potentials for application in improving the accuracy and efficiency of ship detection in foggy environments.