基于深度学习的高压水鱼片切段装置控制系统设计  

Design of system with high-pressure water jet based on deep learning

作　　者：肖哲非 马田田 张军文[1,2,3] 李彦龙 沈建 XIAO Zhefei;MA Tiantian;ZHANG Junwen;LI Yanlong;SHEN Jian(Fishery Machinery and Instrument Research Institute,Chinese Academy of Fishery Sciences,Shanghai 200092,China;Key Laboratory of Ocean Fishing Vessel and Equipment,Ministry of Agriculture and Rural Affairs,Shanghai 200092,China;National R&D Branch Center for Aquatic Product Processing Equipment,Shanghai 200092,China;Dalian Polytechnic University,Collaborative Innovation Center of Seafood Deep Processing,Dalian 116034,Liaoning,China)

机构地区：[1]中国水产科学研究院渔业机械仪器研究所,上海200092 [2]农业农村部远洋渔船与装备重点实验室,上海200092 [3]国家水产品加工装备研发分中心,上海200092 [4]大连工业大学海洋食品精深加工关键技术省部共建协同创新中心,辽宁大连116034

出　　处：《渔业现代化》2023年第5期79-89,共11页Fishery Modernization

基　　金：中国水产科学研究院基本科研业务费资助(2020TD70)。

摘　　要：为了实现鱼片的自动化、节能化切段操作,设计了一种基于深度学习的识别鱼片轮廓并规划切割路径的控制系统。使用深度学习模型对采集到的鱼片图像进行实例分割,根据识别结果规划切割路径,并由固定有高压水喷嘴的切割执行机构完成鱼片切段作业。目标检测试验中,在Mask R-CNN的基础上改进算法,加入Mosaic数据增强方法,主干网络使用ResNetXt50并加入注意力机制SKNet。改进后的模型检测精度达到98%,平均检测时间为0.058 s,相对于以ResNet50为主干网络的原模型,在检测时间仅增加0.001 s的基础上将检测精度提高了9.9%。鱼片切段试验中,平均切割合格率达到90%。研究表明,该系统可准确快速识别鱼片轮廓,高效完成鱼片切割作业,为未来向着智能化方向发展的水产品加工行业提供了新思路。The automatic fish slicing equipment can improve production efficiency and save labor costs,and the use of high-pressure water for fish processing is an environmentally friendly and sustainable production method.Using high-pressure water cutting does not alter the physical and chemical properties of the object,meeting the requirements of food safety.Most of the existing fish fillet cutting equipment have a fixed cutting distance and cannot intelligently cut according to the size of the fish fillet.In order to study the automation and energy-saving cutting device for fish slices,a control system based on deep learning was designed to identify the contour of fish slices and plan the cutting path,and herring fish slices were used as cutting samples for testing.First,the deep learning model was used to perform instance segmentation on the collected fish slice images.Then,based on the output from deep learning model and actual requirements,the cutting path was planned.Finally,the cutting path was input to the motion controller,and the cutting execution mechanism with high-pressure water nozzle was used to complete the fish slicing operation.In the object detection experiment,the model was improved based on Mask R-CNN,incorporating Mosaic data augmentation method and using ResNetXt50 as the backbone with the attention mechanism SKNet.The improved model achieved a detection accuracy of 98%and an average detection time of 0.058 s.Compared to the original model with ResNet50 as the backbone,the detection time increased by only 0.001 s,while the detection accuracy improved by 9.9%.In the fish slicing experiment,the average cutting pass rate was 90%.The control system proposed in this article has good flexibility and scalability,and the method of modifying the planned path based on actual needs can be applied to meet various cutting requirements in production.In the next step of research,we plan to incorporate 3D scanning equipment and combine 3D point cloud information to meet more cutting requirements such as equal weight

关 键 词：鱼片切断装置 控制系统 深度学习 Mask R-CNN 

分 类 号：TP271[自动化与计算机技术—检测技术与自动化装置]