Chinese mitten crab detection and gender classification method based on GMNet-YOLOv4

Abstract

The Chinese mitten crab (Eriocheir sinensis) is a unique aquaculture species in China. The accurate detection of crab targets and gender classification is crucial in guiding biomass estimation, separate breeding based on gender, and quality grading during crab breeding. Current crab gender classification methods find addressing complex backgrounds and processing images with multiple crabs challenging. Herein, we propose a lightweight crab detection and gender classification method based on the improved YOLOv4, called GMNet-YOLOv4. First, crab images with multiple backgrounds were collected to construct crab detection and gender classification datasets. Second, the lightweight GhostNet was selected as the backbone of the original YOLOv4 to extract crab features. Subsequently, the standard convolution of the neck and head network was replaced by a depthwise separable convolution in MobileNet, which further reduces the number of parameters while maintaining accuracy. Finally, the proposed method was used to detect, localize, and classify crabs using an appropriate bounding box and class, and the outputs of the model were the bounding boxes and classes (male or female). Experiments were conducted on the crab image dataset considering backgrounds, heights, and occlusion degrees. The results demonstrated a precision of 96.75%, recall of 97.07%, F1-score of 96.90%, and mean average precision (mAP) of 97.23% on the test set. Compared with the original YOLOv4, the precision of the proposed method was improved by 2.82% and the number of parameters was reduced by 82.24%. Furthermore, compared with different object detectors such as Faster R-CNN and single shot detector, the precision of the proposed method increased by 3.95% and 2.40%, the recall increased by 0.73% and 5.13%, the F1-score increased by 2.40% and 3.01%, and the mAP increased by 1.64% and 3.01%, respectively. The experimental results confirmed that the proposed method has a low memory requirement and high detection and gender classification accuracy. Additionally, it effectively detects and classifies E. sinensis based on gender.