【作者】 王子豪；

【导师】 胡志刚；

【作者基本信息】 武汉轻工大学 ， 机械工程， 2024， 硕士

【摘要】 提升小龙虾分级工作的准确率和效率,有助于后续小龙虾活体售卖、初级加工和精深加工等工作。通过引入机器视觉技术,创建小龙虾图像数据集,训练深度学习神经网络,能够实现对活体小龙虾的无损、快速和精确分级,具体内容如下:(1)搭建小龙虾图像采集系统,详细说明小龙虾拍摄规范与环境要求。获取小龙虾原始图像后,创建分割虾头、虾钳和虾尾三个部位的小龙虾语义分割数据集,创建包含青虾和红虾的小龙虾分类数据集。(2)结合小龙虾图像语义分割,提出一种新的小龙虾分级标准,即依据虾头和虾钳在整虾中的比例对小龙虾进行分级。通过构建小龙虾整虾、虾头、虾钳和虾尾的实际重量与对应图像区域像素大小之间的相关性模型,验证新分级标准的可行性。(3)训练DeepLab V3+神经网络,对小龙虾图像进行语义分割并对小龙虾进行分级。小龙虾语义分割数据集图像共290张,其中训练集图像200张,测试集图像90张。使用训练数据集图像训练DeepLab V3+神经网络,并用测试集检验模型语义分割效果以及小龙虾分级的准确率。语义分割主要评价指标为平均交并比(MIo U)、平均像素准确率(MPA)和像素准确率(PA)。小龙虾图像语义分割测试集的MIo U为94.35%,MPA为96.56%,PA为99.44%,同时测试集小龙虾分级准确率为85.56%。试验结果表明,DeepLab V3+模型可以准确分割小龙虾图像并估测虾头虾钳占比,模型能够完成小龙虾分级任务。(4)通过添加Ni N模块以及微调模型结构的方式优化原始ResNet-18神经网络,实现对小龙虾青虾和红虾的分类。小龙虾分类数据集共1000张图像,将其分为训练集(800张)和测试集(200张)。模型训练时使用图像增广技术,对小龙虾数据集进行扩增。使用小批量数据进行预训练,确定模型较优超参数设置。使用训练数据集训练原始ResNet-18、优化ResNet-18和Goog LeNet模型,三种模型测试集的分类准确率分别为89%、94%和78%,同时优化模型训练时间与原始模型相比减少59.26%。试验结果表明,ResNet-18模型优化效果显著,优化ResNet-18模型能够完成小龙虾分类任务。(5)搭建小龙虾分级实物装置,该装置主要由视觉采集系统、小龙虾输送系统和气动执行系统三部分组成。同时,设计配套的小龙虾分级和分类检测软件界面。更多还原

【Abstract】 Improving the accuracy and efficiency of crayfish grading work contributes to subsequent tasks such as live crayfish sales,primary processing,and advanced processing.By integrating machine vision technology and creating a dataset of crayfish images,training deep learning neural networks enables non-destructive,rapid,and precise grading of live crayfish.The specific details are as follows:(1)Establishing a crayfish image acquisition system involves detailed specifications for photographing crayfish and environmental requirements.After obtaining the source images of crayfish,a semantic segmentation dataset is created for the segmentation of three parts:the head,claws,and tail of the crayfish.Additionally,a crayfish classification dataset is created,including both green and red crayfish.(2)Integrating crayfish image semantic segmentation,a novel grading criterion is proposed based on the ratio of the head and claws to the whole crayfish.By establishing a correlation model between the actual weight of the whole crayfish,head,claws,and tail,and the corresponding pixel sizes of the image regions,the feasibility of the new grading criterion is validated.(3)Training the DeepLab V3+ neural network for semantic segmentation and grading of crayfish images.The crayfish semantic segmentation dataset comprises 290 images,with200 images allocated for training and 90 for testing.The DeepLab V3+ neural network is trained using the training dataset,and the model’s semantic segmentation performance and accuracy in crayfish grading are evaluated using the test dataset.The primary evaluation metrics for semantic segmentation are the Mean Intersection over Union(MIo U),Mean Pixel Accuracy(MPA),and Pixel Accuracy(PA).For the crayfish semantic segmentation test set,MIo U is 94.35%,MPA is 96.56%,and PA is 99.44%.Simultaneously,the accuracy of crayfish grading on the test set is 85.56%.The experimental results demonstrate that the DeepLab V3+ model accurately segments crayfish images and estimates the ratio of the head and claws,enabling successful crayfish grading tasks.(4)Optimizing the ResNet-18 neural network by incorporating Ni N modules and finetuning the model structure to achieve classification of green and red crayfish.The crayfish classification dataset comprises 1000 images,divided into a training set(800 images)and a test set(200 images).Image augmentation techniques are employed during model training to augment the crayfish dataset.Pre-training with mini-batch data is conducted to determine optimal hyperparameters for the model.Training is performed on the original ResNet-18,optimized ResNet-18,and Goog LeNet models using the training dataset.The classification accuracies on the test set for the three models are 89%,94%,and 78%,respectively.Furthermore,the optimized model reduces training time by 59.26% compared to the original model.The experimental results demonstrate significant improvement in the optimization of the ResNet-18 model,indicating that the optimized ResNet-18 model effectively accomplishes the crayfish classification task.(5)Building a physical device for crayfish grading,the device mainly consists of a visual acquisition system,crayfish conveyance system and pneumatic actuation system.Additionally,designing a complementary software interface for crayfish grading and classification detection.更多还原