摘要： 梭子蟹（Portunus trituberculatus）具有较高的营养和经济价值，其外观的完整性直接影响其价格。针对梭子蟹分选过程中存在的效率低、成本高和受主观因素影响较大的问题，该研究提出了一种基于改进ConvNext模型的梭子蟹缺陷智能识别方法。首先利用重参数化重聚焦卷积替换ConvNext模型Block模块中的深度可分离卷积；其次在ConvNext模型的Block模块前以及下采样模块后加入坐标注意力机制。改进后的模型在复杂情况下的识别准确率最高，为98.90%，比ConvNext模型高出3.32%。同时，精确率、召回率和F1分数等各项指标均优于其他模型，展现出良好的泛化性和鲁棒性。结果表明，改进后的ConvNext模型能够有效解决梭子蟹分拣过程中存在的问题，为开发梭子蟹缺陷的自动识别和分类系统提供了技术支持，推动梭子蟹分拣技术的智能化发展。

关键词: 深度学习, 三疣梭子蟹, 识别缺陷, 改进ConvNext, 引入CA注意力

Abstract: Portunus trituberculatus has high nutritional and economic value, and its appearance integrity directly affects its market value. At present, the sorting of Portunus trituberculatus faces challenges due to low levels of automation in traditional methods. Traditional sorting of Portunus trituberculatus is generally done manually, which has the problems of low efficiency and high labor costs, and human observation is easily affected by subjective factors. To solve the sorting problem of Portunus trituberculatus, this paper proposes a defect recognition model for crabs based on an improved ConvNext model , which can accurately identify the feet of Portunus trituberculatus and carry out intelligent recognition and grading. By accurately identifying the feet of Portunus trituberculatus, intelligent recognition and grading are carried out. Using re-parameterized refocusing convolution to replace depthwise separable convolution in the ConvNext model block, the network is able to capture richer and more detailed features. Coordinate attention mechanisms are added before the Block module and after the Downsample module of the ConvNext model to enhance the model's attention to key features and discard irrelevant features. The results show that the improved model has the highest accuracy on the validation set, at 98.90%, which is 3.32% higher than the original ConvNext baseline model network. The proposed algorithm ensures effectiveness. This study contributes to achieving the goal of intelligent and precise grading in the Portunus trituberculatus industry, and is anticipated to replace traditional methods to improve the sorting efficiency of Portunus trituberculatus. This can provide technical support for the further development of an automatic recognition and classification system for defects in Portunus trituberculatus.

Key words: deep learning, Portunus trituberculatus, defect identification, improved ConvNext, introducing CA attention