Soft-shell crab detection model based on YOLOF

• Published in: Aquaculture international Vol. 32; no. 4; pp. 5269 - 5298
• Main Authors: Zhang, Zhen, Liu, Feifei, He, Xinfeng, Wu, Xinyu, Xu, Meijuan, Feng, Shuai
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.08.2024; Springer Nature B.V
• Subjects: Adaptability; Agricultural practices; Algorithms; Biomedical and Life Sciences; Breeding methods; Convolution; Crustaceans; Decapoda; Farming; Freshwater & Marine Ecology; Inspection; Life Sciences; Marine crustaceans; Moulting; Zoology; Deep learning; YOLOv5; Soft-shell crab; Mixed Convolution; Crab Apartment
• ISSN: 0967-6120; 1573-143X
• DOI: 10.1007/s10499-024-01426-2

The detection technology of soft-shell crab breaks the traditional soft-shell crab breeding method. Conventional farming practices rely heavily on human labor for the identification of soft-shell crab birth, and manual inspections are prone to issues such as delayed and inaccurate results. In this paper, we propose an improved high-precision detection algorithm called YOLOF based on YOLOv5. This algorithm achieves precise detection of soft-shell crab. To address the absence of soft-shell crab detection datasets on the internet, this paper establishes the Crab Molting Detection Dataset (CMP) for the purpose of training neural networks. The detection algorithm has undergone several improvements in the following aspects: the introduction of the FassterNetT1 feature extraction network to replace the original CSPDarknet53 backbone network significantly reduces model complexity and computational overhead; The incorporation of an additional auxiliary prediction feature layer of size 5 × 5 enhances the model’s adaptability for detecting various molting states of crabs; the activation function GeLU was introduced, to enable the model to extract features more smoothly and reduce information loss during feature extraction; additionally, the definition of Mixed Convolution (MConv) was proposed to replace Partial Convolution (PConv) in the FasterNet Block. MConv exhibits comparable speed to PConv while possessing high-performance convolution capabilities, allowing the backbone network to extract more abundant semantic information; The introduction of the Content-Aware ReAssembly of FEatures (CARAFE) operator, a universal upsampling algorithm, enhances the extraction of crab shell contour information, thereby improving the model’s accuracy in detecting soft-shell crab; addition of an Explicit Visual Center (EVC) in the network neck to strengthen the model’s grasp of long-range dependencies and critical information in the image, enhancing the model’s robustness. The experimental results demonstrate that YOLOF achieves a 5.4% improvement in mean average precision compared to the original YOLOv5s algorithm. The average precision for the classes “shedding_crab” and “end_unshell” increased by 1.7% and 16.3% respectively. The model running speed is 28FPS, indicating that the proposed YOLOF algorithm can accurately and efficiently identify the molting states of crabs. This study provides the theoretical foundation and technical support for the realization of an automated system for soft-shell crab cultivation.