AB-YOLOv8: Attention-based Feature Extraction model for Underwater Object Detection

Abstract

Accurate and timely underwater object detection is crucial in the field of marine environmental engineering. The detection of such targets has been improved recently using techniques based on Convolutional Neural Networks (CNN). However, the processing performance of deep neural networks is typically inadequate due to their high parameter requirements. Accurate detection is difficult with current techniques when dealing with small, close-packed underwater targets. In order to overcome these problems, the proposed work combined YOLOv8 with different attention modules and proposed a novel neural network model to enhance underwater object detection capabilities. In this research, AB-YOLOv8 is proposed, which adds the attention mechanism to the original YOLOv8 design. To be more precise, the proposed work introduced four attention modules, Convolutional Block Efficient Channel Attention (ECA), Shuffle Attention (SA), Global Attention Mechanism (GAM), and Attention Module (CBAM), to create the enhanced models and train them in the aquarium dataset. Each of the attention blocks is combined with YOLOv8 to improve the performance of the entire object detection. The residual block is introduced into the CBAM to optimize the performance of the CBAM. The detailed experiments are conducted on the aquarium dataset, and various performance assessment parameters are used, like mAP, FLOPS, Params, inference time, etc. After performing the experiment, it was found that ECA gives the best result out of all attention blocks and improved mAP value by 8%, also reduced the number of parameters generated during training. To validate the work,
we also performed the experiment on the Brackish dataset, and we found that ECA outperforms other attention mechanisms with YOLOv8.