文献类型：期刊文献

中文题名：基于改进YOLOv7的网箱网衣破损识别方法

英文题名：A method for identifying the damage of cage netting based on improved YOLOv7

作者：俞国燕[1,3];苏锦萍[1,3];陈泽佳[1];陈帅兴[1];陈其菠[1];吴振陆[2]

机构：[1]广东海洋大学机械工程学院,广东湛江524088;[2]广东海洋大学数学与计算机学院,广东湛江524088;[3]广东省海洋装备及制造工程技术研究中心,广东湛江524088

年份：2023

卷号：50

期号：4

起止页码：126

中文期刊名：渔业现代化

外文期刊名：Fishery Modernization

收录：CSTPCD、、CSCD_E2023_2024、CSCD

基金：广东省区域联合基金项目(2019B1515120017);广东省海洋经济发展(海洋六大产业)专项(GDNRC[2021]42);湛江市现代海洋渔业装备重点实验室(2021A05023)。

语种：中文

中文关键词：网衣破损;精准实时识别;目标检测;视觉系统;智能网衣修补机器人

外文关键词：damaged cage netting;accurate real-time recognition;target detection;visual system;intelligent clothes repairing robot

中文摘要：网箱网衣极易破损,一旦破损未及时修补,会给养殖户造成巨大的经济损失。为实现智能化网箱网衣破损检测,本研究提出一种基于改进YOLOv7的网箱网衣破损识别方法。该方法通过在Backbone网络使用gnConv结构、Neck网络引入SimAM模块来提升模型表达能力更好聚焦网衣破损处的特征,提高模型的检测精度。Backbone网络使用深度可分离卷积,并减少激活函数和改变卷积步长,同时在Neck网络利用Bottleneck模块使用1×1卷积核的特点和使用性能更佳的Mish激活函数重构模型,以减少参数量和运算成本,实现模型检测速度的提升及尺寸的压缩。通过消融试验和对比试验结果显示,YOLOv7-C3NeHX比原YOLOv7算法的平均精度提高了3.1个百分点,精确率、召回率与F 1值分别提升了0.5、4.2与3个百分点,检测速度达到了232.56FPS,GFLOPs和模型尺寸分别占原YOLOv7的38.2%和94.3%。研究表明,改进模型能有效提高识别效率和部署的灵活性,为智能网衣修补机器人的研发提供技术支持。

外文摘要：The cage netting is prone to breakage and if not repaired in time,it can cause huge economic losses to farmers.In order to realize the intelligent damage detection of cage netting,this study proposes a damage identification method based on improved YOLOv7.We use the GNConv structure in the Backbone network and the SimAM module in the Neck network to improve the model′s expressiveness and better focus on the features at the cage netting breakage.This way the detection precision of the model is improved.Using depthwise separable convolution in Backbone networks with reduced activation functions and varying convolution step sizes.The model is also reconstructed in the Neck network using the Bottleneck module with a 1×1 convolution kernel and the Mish activation function with better performance.As a result,the number of parameters and the cost of operations are reduced,increasing in the speed of model inspection and a size reduction.The results of ablation tests and comparative tests showed that the average precision of the YOLOv7-C3NeHX algorithm is 3.1 percentage points higher than that of the original YOLOv7 algorithm,and its precision,recall and F 1 score are 0.5,4.2 and 3 percentage points higher,respectively.Detection speed up to 232.56 FPS.GFLOPs and model size account for 38.2% and 94.3% of the original YOLOv7.The improved model can effectively improve the identification efficiency and deployment flexibility,and provide technical support for the research and development of intelligent clothing repair robot.