基于改进Faster R-CNN绪下茧形态识别与计数方法的研究  

Research on the morphological recognition and count of undertone cocoons based on the improved Faster R-CNN

作　　者：杨青青 邵铁锋 孙卫红 梁曼 YANG Qingqing;SHAO Tiefeng;SUN Weihong;LIANG Man(College of Mechanical and Electrical Engineering,China Jiliang University,Hangzhou 310018,China)

机构地区：[1]中国计量大学机电工程学院,浙江杭州310018

出　　处：《中国计量大学学报》2023年第2期224-230,240,共8页Journal of China University of Metrology

基　　金：浙江省基础公益研究计划项目(No.LGG20E050014)。

摘　　要：目的:为解决解舒试验过程中绪下茧人工识别与计数劳动强度大等问题,提出一种基于改进Faster R-CNN绪下茧形态识别与计数方法。方法:首先,根据解舒试验过程中绪下茧呈现的不同形态,将绪下茧分为新茧、中茧、薄茧3种。采集绪下茧图像,构建数据集,使用ResNet50残差网络作为Faster R-CNN的特征提取网络,提取3种绪下茧形态图像特征;其次,调整区域建议网络(RPN)中锚点(Anchor)的比例,使检测结果中的目标矩形框更加精确;再次,将SENet注意力模块加入到特征提取网络中;最后,在形态识别的基础上,统计绪下茧数量。结果:改进算法训练的模型对绪下茧的3种形态的平均准确率达到了86.37%,召回率达到了90.3%。检测的平均速度0.17 s/幅。结论:该算法满足绪下茧形态识别与计数的要求。Aims:In order to solve the problem of labor intensity,this paper proposed a method of morphological identification and counting of undertone cocoons based on the improved Faster R-CNN.Methods:Firstly,according to the different forms of the cocoons in the relieving test process,the cocoons were divided into three forms:the initial,the middle and the end.The undertone cocoon images were collect to construt the data set.The ResNet50 residual network was used for the feature extraction network of Faster R-CNN to extract the features of three undertone cocoon morphological images.Secondly,the proportion of Anchor points in RPN was adjusted to make the target rectangle box in the detection result more accurate.Then,the SENet attention module was added to the feature layer.Finally,on the basis of morphological recognition,the number of undertone cocoons was counted.Results:The experimental results showed that the average accuracy of the model trained by the improved algorithm reached 86.37%;and the recall rate reached 90.3%.The average speed of detection was 0.17 s econd per image.Conclusions:The algorithm satisfies the requirement of recognition and the counting of the undertone cocoon shape.

关 键 词：解舒试验 Faster R-CNN算法 SENet注意力模块 绪下茧识别与计数 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]