Deep residual convolutional neural network based on hybrid attention mechanism for ecological monitoring of marine fishery

Abstract

Understanding the ecological environment, population abundance, and growth status of marine organisms in the marine fishery is important to promote its sustainability. However, existing manual detection methods can cause some damage to marine ecology and are difficult to meet the demand for fast and accurate detection. In addition, light, shadows, and disturbances in the marine ecosystem can affect the effectiveness of intelligent detection methods. To address these problems, a deep residual convolutional neural network (DRCNN) based on hybrid attention mechanism (HAM) is proposed to detect marine organisms. The hybrid attention mechanism obtains key information from both channel and space dimensions of the input image. And the residual module is added to the deep convolutional neural network to iteratively extract image features while avoiding error accumulation. It is demonstrated experimentally that the HAM-DRCNN model achieves 93.17% image localization accuracy with a processing speed of 20.95 frames/s. Compared with the YOLOv5 and Faster R-CNN, the mean average precision of species classification is 91.36%, which is an improvement of 0.93% and 9.39%, respectively. In addition, excellent results are achieved on two other benchmark datasets. The method can accurately locate and complete the classification of marine organisms in underwater images and has practical application value.