EFP-YOLO: A quantitative detection algorithm for marine benthic organisms

Abstract

Underwater object detection technologies are widely applied in the field of marine scientific research. In real underwater environments, a large number of dense, blurred, and small-sized objects appear. It is difficult for current general-purpose object detectors to detect these samples. To solve these problems, this paper proposes a quantitative detection algorithm EFP-YOLO (Enhance the Ability of Feature Extraction and Perception Ability of YOLOX) for marine benthic organisms that improved on YOLOX. Firstly, IGACSP (Cross Stage Partial Module with Interactive Global Attention) is embedded in the backbone to enhance the ability for extracting features of small-sized and blurred objects, which fuses local and global information in a parallel interactive way. To improve the capacity for detecting multiscale objects, EPPM (Efficient Pyramid Pooling Module) fuses feature maps from different scales that provide sufficient contextual information. Secondly, FSM (Feature Selection Module) and FAM (Feature Alignment Module) are introduced into the feature fusion part for spatial feature extraction and feature alignment, respectively. The modules provide accurate boundary localization information for the detection of dense objects. Finally, ATFhead (Asymmetric Task Focused head) is proposed to increase the scale perception, spatial perception, and task perception of the detection head. It improves the classification and localization abilities of the model, thus achieving accurate counting of marine benthic organisms. The experimental results show that compared with YOLOX-S, mAP0.5:0.95 is improved from 59.7% to 64.6% on the DUO dataset. EFP-YOLO demonstrates superior performance in the quantitative detection of marine benthos compared to other object detection algorithms. This can effectively advance marine scientific research and promote biological conservation efforts. To show the generality of EFP-YOLO, we use a generic object detection dataset for experiments. You can access the code for EFP-YOLO publicly on https://github.com/llllllvv/EFP-YOLO.