摘要： 中国鱿鱼加工产业规模庞大，鱿鱼足作为加工过程中的主要副产品之一，其切片作业当前仍以人工操作为主。受到鱿鱼足形状不规则的客观限制，人工切割普遍存在切割精度欠佳以及产品成品率不足等问题。针对该现状，提出一种面向鱿鱼足的高精度定量切割方法，首先，搭建线激光扫描平台，采集鱿鱼足的点云数据，并通过点云匹配与三维重建技术合成完整的鱿鱼足点云模型；其次，构建一种改进的生成对抗网络模型，生成高质量的三维完整模型；最后，将定量切割问题建模为多目标优化问题，采用改进模拟退火算法在点云模型上开展切割路径的仿真试验。试验结果显示，改进后的生成对抗网络模型在交并比上比基准模型提升6.2%，在交叉熵上降低了42.4%。在切割模拟试验中，限定单片质量误差≤8%，鱿鱼足的平均成品率可达87.3%。本方法显著提升了鱿鱼足的切割精度与物料利用率，为水产品高值化加工提供了新的技术思路。

关键词: 鱿鱼足加工, 定量切割, 生成对抗性网络, 模拟退火算法, 多目标优化 

Abstract: Squid is a marine mollusk, stands out as one of the most commercially valuable seafood resources globally, with profound economic significance for China’s marine fisheries sector. As a key species in China’s aquatic product supply chain, squid contributes significantly to the national marine catch volume: data from 2024 shows that China’s total squid catch reached 317,325 tons, accounting for 32.97% of the country’s overall marine catch. This substantial output underscores the urgent need for efficient and precise processing technologies to maximize the economic value of squid products, particularly in the segment of squid tentacle slicing, an essential step in producing value-added products such as frozen squid slices, canned squid, and ready-to-eat seafood snacks. Against this backdrop, this project proposes a novel quantitative cutting solution for squid tentacles, integrating line laser scanning, 3D point cloud reconstruction, improved deep learning, and optimized algorithmic decision-making. The implementation process consists of three core stages: First, a line laser scanning platform was constructed to capture the 3D morphological information of squid tentacles. Given that a single-angle laser scan can only obtain partial point cloud data, the platform performs multiple laser scans of the same squid tentacle from different angles. The acquired multi-angle incomplete point cloud datasets are then processed through point cloud matching and surface reconstruction. This process ultimately synthesizes a complete, high-resolution 3D point cloud model that accurately represents the entire morphological structure of the squid tentacle, including details such as suction cup distribution and local diameter variations. Second, an improved Generative Adversarial Network deep learning model was established to address the potential inefficiency of multi-angle scanning in industrial scenarios. The key improvement lies in integrating an attention mechanism into both the autoencoder and decoder modules of the original GAN architecture. This attention mechanism enables the model to dynamically weight and emphasize valuable feature information during the learning process, while downplaying irrelevant or noisy data. The trained model can efficiently reconstruct the complete 3D structure of a squid tentacle from a single incomplete point cloud, significantly reducing scanning time while maintaining morphological accuracy. Comparative experiments show that the improved GAN model outperforms the baseline GAN model by 6.2% in the Intersection over Union (IoU) index and by 42.4% in the Cross-Entropy (CE) index, demonstrating its superior performance in 3D structure reconstruction. Finally, the quantitative cutting of squid tentacles was transformed into a multi-objective optimization problem, with the core objectives being: minimizing the weight error of each sliced piece and maximizing the overall utilization rate of the squid tentacle. To solve this problem, the Simulated Annealing  algorithm was improved by incorporating domain-specific constraints from squid processing. Cutting tests were then conducted on the 3D point cloud models of squid tentacles using the improved SA algorithm. Experimental results confirm that under the constraint of a single-slice weight error≤8%, the average utilization rate of squid tentacles reaches 87.3%, a significant improvement of 27.3–37.3 percentage points compared to the 50–60% utilization rate of manual slicing. In summary, this project develops a comprehensive quantitative cutting technology for squid tentacles that integrates 3D sensing, intelligent reconstruction, and optimized decision-making. It effectively addresses the inefficiencies, low precision, and high waste of traditional manual and mechanical cutting methods, providing a feasible technical solution for the industrial upgrading of the squid processing industry and laying a foundation for the intelligent transformation of aquatic product processing.

Key words: squid tentacles processing, quantitative cutting, generating adversarial networks, simulated annealing algorithm, multi-objective optimization