SUR-Net: A Deep Network for Fish Detection and Segmentation With Limited Training Data

Abstract

To detect and identify fish in the wild, there are many challenges mainly due to the distinctive perception environment in the ocean. Another important issue is the lack of sufficient image data for network training. In this article, we present a technique for underwater fish detection and semantic segmentation using limited training data. Based on the U-Net architecture, various convolutional layers and residual blocks are incorporated for target detection and segmentation. A new SUR block is constructed to increase the weights of relevant features, and the residual blocks are used to learn the feature representation after concatenating the shallow and deep layers. The overall performance is further improved by mosaic data augmentation. In the experiments, the proposed technique is evaluated and compared with the convolutional neural network (CNN) models for detection and semantic segmentation. The results obtained from the open sea underwater fish datasets have demonstrated the effectiveness and feasibility of our approach with limited training data. Compared with the current neural network models, the proposed technique is able to achieve the performance of 95.04% F-1 score and 88.19% mIoU in the complex scenes. Code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Junyi8829/SURNet</uri> .