Target Detection of Forward-Looking Sonar Image Based on Improved YOLOv5

Forward-looking sonar is a commonly used underwater detection device at present, but the detection accuracy is poor due to the complex underwater environment, small target highlight area and fuzzy feature details. Therefore, this paper proposes a forward sonar image target detection model based on You Only Look Once Version 5 (YOLOv5) network using transfer learning method. First, the YOLOv5 network is pretrained with COCO data set. Then the pre-training model is fine-tuned according to the training set of forward-looking sonar images. Before fine-tuning, the traditional k-means clustering is improved. The intersection over union (
$IoU$
) value is used as the distance function to cluster the labeling information of the training set of the forward-looking sonar image. The results of clustering serve as the initial anchor frame of the training network. This operation greatly improves the detection speed. Second, due to the characteristics of weak echo intensity and small target area of forward-looking sonar image, an improved feature extraction method of CoordConv was proposed to give corresponding coordinate information to high-level features which improves the accuracy of network detection regression. Finally, the fine-tuned network is used to detect the target in the forward-looking sonar image. The experimental results show that the improved model based on YOLOv5 network is superior to the original YOLOv5 network and other popular deep neural networks for target detection in the forward-looking sonar image, which has a reference significance for underwater target detection. The CoordConv-YOLOv5 network based on transfer learning proposed in this paper shows the best performance in both detection accuracy and detection speed. Detection accuracy mAP@0.5:0.95 can reach 56.95%, and detection speed can reach 9ms.