Underwater target detection algorithm based on improved YOLOv4 with SemiDSConv and FIoU loss function

• DOI: 10.3389/fmars.2023.1153416
• Publication Date: 2023-03-23T06:29:02Z
• Authors: Chengpengfei Zhang; Guoyin Zhang; Heng Li; Hui Liu; Jie Tan; Xiaojun Xue
• Keywords: Convolution (computer science)

Underwater target detection is an indispensable part of marine environmental engineering and a fast accurate method detecting underwater targets essential. Although many algorithms have achieved great accuracy in daily scenes, there are issues low-quality images due to the complex environment, which makes applying these deep learning directly process tasks difficult. In this paper, we presented algorithm for based on improved You Only Look Once (YOLO) v4 response environment. First, developed new convolution module network structure. Second, intersection over union loss was defined substitute original function. Finally, integrated some other useful strategies achieve more improvement, such as adding one prediction head detect varying sizes, integrating channel attention into network, utilizing K-means++ cluster anchor box, different activation functions. The experimental results indicate that, comparison with YOLOv4, our proposed average dataset by 10.9%, achieving 91.1%, speed 58.1 frames per second. Therefore, compared mainstream algorithms, it superior feasible applications intricate environments.