摘要： 针对牡蛎形状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 ,本研究构建了一种基于改进YOLOv10 的轻量级牡蛎形状识别模型 ,以提高牡蛎形状识别的自动化和准确性 。首先 ,将 YOLOv10n 网络结构的骨干网络替换为PP-LCNet 结构;其次 ,在骨干网络中的PSA 部分添加聚焦线性注意力模块;然后 ,颈部网络的上采样模块替换为DySample 动态上采样算子 ;最后 ,将原激活函数替换为 ARelu 激活函数 。结果显示 :改进后的模型浮点计算数、参数量及大小相比原YOLOv10n 基线模型分别降低 20. 7%、22. 2%和 22. 4% ,同时精确度相较于原模型提高了 3. 8个百分点 ,达到 94. 4%。本研究为牡蛎形状识别提供了一种有效的解决方案 ,也为开发牡蛎形状的自动识别和分类系统提供了技术支持。

关键词: 深度学习, 牡蛎, 形状识别, 改进 YOLOv10, 轻量级

Abstract: To overcome the inefficiency, high expense, and subjective inconsistency of manual grading, we introduce a lightweight oyster shape recognition approach built upon an enhanced YOLOv10 framework. . First,the backbone of YOLOv10n is replaced with PP -LCNet; second,a Focused Linear Attention module is inserted into the PSA blocks of the backbone; third,the upsampling operator in the neck is substituted with the DySample dynamic upsampler; ultimately, the original activation is superseded by the AReLU function. Compared with the original YOLOv10n,the enhanced model trims floating- point operations by 20. 7% ,parameters by 22. 2% ,and overall size by 22. 4% ,while raising precision to 94. 4%—a 3. 8-point improvement. The proposed approach not only provides an effective solution for oyster shape recognition but also offers technical support for developing automatic identification and classification systems,thereby advancing the intelligent development of oyster sorting technology.

Key words: deep learning, oyster, shape recognition, improved YOLOv10, light weight