基于DeepLab V3+估测小龙虾虾头虾钳占比及分级

Grading crayfish by estimating the proportion of crayfish head and pincers based on DeepLab V3+

目的:完善小龙虾分级工作。方法:搭建小龙虾图像拍摄平台,获取小龙虾原始图像,创建分割虾头、虾钳和虾尾3个部位的语义分割数据集。分析小龙虾虾头、虾钳、虾尾3个部位实际质量与数据集中对应部位像素大小之间的相关性,总结得到根据整虾中虾头虾钳占比进行分级的小龙虾分级新标准。使用数据集训练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,语义分割评价指标为平均交并比(MIoU)、平均像素准确率(MPA)和像素准确率(PA)。结果:小龙虾语义分割测试集的MIoU为94.35%,MPA为96.56%,PA为99.44%,测试集小龙虾分级准确率为85.56%。结论: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。

Objective:To achieve reasonable and effective grading of live crayfish,and improve the work of grading crayfish.Methods:The construction of crayfish image shooting platform,to obtain the original image of crayfish,and the semantic segmentation dataset which segmented the three parts of the crayfish head,crayfish pincers,and crayfish tail was created.The correlation between the actual weight of three parts and the corresponding pixel size in the dataset was analyzed,and a new grading standard for crayfish which was according to the proportion of head and pincers in the whole crayfish was summarized.The DeepLab V3+neural network was trained using the crayfish semantic segmentation dataset,and the test set was used to test the semantic segmentation effect of the model and the accuracy of crayfish grading.Semantic segmentation evaluation criteria were mean intersection over union(MIoU),mean pixel accuracy(MPA)and pixel accuracy(PA).Results:The MIoU of the crayfish semantic segmentation test set was 94.35%,the MPA was 96.56%,and the PA was 99.44%.The accuracy of crayfish grading in the test set was 85.56%.Conclusion:The DeepLab V3+model can accurately segment crayfish images and estimate the proportion of crayfish head and pincers,and the model can complete the crayfish grading task.