基于多头注意力机制的非规则形状鱼片部位分割方法研究

Research on a Part Segmentation Method for Irregularly Shaped Fish Fillets Based on a Multi-Head Attention Mechanism

摘要: 针对非规则形状鱼片不同部位边界特征不明显、自动分割精度不足的问题，以线激光扫描获得的200个三文鱼片点云为研究对象，根据鱼体形态特征将其划分为上鱼背、鱼腹、鱼腩、后段和尾部五个区域，并建立相应的部位分割数据集。在PointNet++第一层集合抽象模块中引入多头注意力机制，强化局部点云特征及不同部位之间的特征联系，构建Multi-PointNet++分割模型。以PointNet++和PointCNN为对比模型开展试验，结果表明，Multi-PointNet++的平均交并比和总体精度分别达到0.822和0.911，较PointNet++分别提高0.042和0.024，较PointCNN分别提高0.029和0.019，其中边界较复杂的后段区域交并比由0.588提高至0.681。该方法能够减少鱼片部位交界区域的误分类，为高价值鱼类的精细化分割提供参考。

Abstract: To address the difficulties in automatically and accurately separating irregularly shaped fish fillets with indistinct boundaries between different parts, a fish fillet part segmentation method based on a multi-head attention-enhanced PointNet++ is proposed. Salmon fillets were scanned using a line laser to obtain surface point clouds. According to their morphological characteristics, each fillet was divided into five regions: upper dorsal, abdominal, belly, posterior, and tail regions. A dataset containing 200 annotated fish fillet point-cloud samples was established. A multi-head attention mechanism was incorporated into the first set abstraction layer of PointNet++ to improve the extraction of local features and correlations among different parts. The Multi-PointNet++ segmentation model was then constructed. Experimental results showed that the proposed model achieved a mean intersection over union (mIoU) of 0.822 and an overall accuracy (OA) of 0.911, representing improvements of 0.042 and 0.024 over the original PointNet++, and improvements of 0.029 and 0.019 over PointCNN, respectively; the IoU of the posterior region, where the boundary was more complex, increased from 0.588 to 0.681. The proposed method effectively reduces misclassification near the boundaries of adjacent parts and provides a technical reference for the fine-grained segmentation of high-value fish.