Abstract: In order to improve the accuracy of marine ship recognition in multiple targets and foggy environments, a marine ship recognition model SE-NMS-YOLOv5 is proposed based on improved YOLOv5 deep learning. The model is combined with Dark channel defogging algorithm, SE(squeeze-and-congestion) attention mechanism module and improved non-maximum suppression model for training and testing of ship data sets. The results showed that in the ship recognition task, there was the accuracy of 90.6%, recall rate of 89.9% and SE-NMS-YOLOv5 F1 value of 90.5%, and compared with YOLOv5 model, the detection effect is improved by 6.3%, 4.8% and 5.8%. Compared with YOLOv4, the model improved 19.1%, 19.0% and 19.3%. In foggy conditions, the accuracy, recall rate and F1 value of SE-NMS-YOLOv5-Dark channel model were 88.1%, 87.2% and 87.6%, compared with SE-NMS-YOLOv5 model, the detection results are improved by 13.8%, 13.3% and 13.5%, respectively. The findings indicate that the marine ship recognition method based on SE-NMS-YOLOv5 effectively solves the problem of low accuracy of marine ship detection on multiple targets and foggy conditions, and improve the overall effect of ship detection and recognition.
2022, Vol. 37 
