A Novel FFT_YOLOX Model for Underwater Precious Marine Product Detection

• Published in: Applied sciences 2022-07, Vol.12 (13), p.6801
• Main Authors: Wang, Peng, Yang, Zhipeng, Pang, Hongshuai, Zhang, Tao, Cai, Kewei
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Automation; Backbone; Computer applications; Datasets; deep learning; Efficiency; FFT; Fishing; Image enhancement; Marine fish; Methods; Performance evaluation; precious marine product detection; Receptive field; Robots; Underwater; Underwater robots; YOLOX
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app12136801

In recent years, the culture and fishing of precious marine product are heavily dependent on manual work, which is labor-intensive, high-cost and time-consuming. To address this issue, an underwater robot can be used to monitor the size of the marine products and fish the mature ones automatically. Automatic detection of marine products from underwater images is one of the most important steps in developing an underwater robot perceiving method. In the traditional detection model, the CNN based backbone suffers from the limited receptive field and hinders the modeling of long-range dependencies, due to the small kernel size. In this paper, a novel detection model FFT_YOLOX based on a modified YOLOX is proposed. Firstly, a unique FFT_Filter is presented, which is a computational efficient and conceptually simple architecture to capture global information of images. Then, a novel FFT_YOLOX model is introduced with fewer model parameters and FLOPs by replacing the standard 3 × 3 kernel in the original backbone of the YOLOX model with a FFT_Filter, for an underwater object detection vision task. Extensive experimental results demonstrate the effectiveness and generalization of the visual representation of our proposed FFT_YOLOX model.