基于Yolo算法的塑料检测定位系统  

Plastic Detection and Positioning System Based on Yolo Algorithm

作　　者：沈文杰[1] SHEN Wen-jie(Fujian Vocational College of Agriculture,Fuzhou 350119,China)

机构地区：[1]福建农业职业技术学院,福建福州350119

出　　处：《塑料科技》2020年第11期115-118,共4页Plastics Science and Technology

摘　　要：为提高塑料检测定位任务的准确率,基于Yolo算法提出一个塑料检测定位模型。针对模型的基础网络层,设计3个由小尺寸卷积层组成的残差模块,以期捕获多样化、抽象塑料特征,以避免梯度消失现象的出现。由仿真分析结果可知:应用数据增强技术可拼接4个图像为1个图像样本,模型的检测定位准确率为0.9497;应用Concat方式拼接特征图后,模型的检测定位准确率提升约1%;应用Hinge损失函数优化模型训练过程,检测定位准确率提升约2%;应用SPP捕获21个多层次的塑料特征向量,促使模型特征表达,最终测试定位准确率为0.9825。In order to improve the accuracy of the plastic detection and positioning task,a plastic detection and positioning model was proposed based on the Yolo algorithm.Aiming at the basic network layer of the model,three residual modules composed of small-size convolutional layers were designed to capture diversified and abstract plastic features to avoid the appearance of gradient disappearance.According to the simulation analysis results,it can be seen that the application of data enhancement technology can stitch four images into one image sample,and the detection and positioning accuracy of the model is 0.9497.After applying the Concat method to splice the feature maps,the detection and positioning accuracy of the model is increased by about 1%.The Hinge loss function is used to optimize the model training process,and the detection and positioning accuracy is increased by about 2%.The SPP is used to capture 21 multi-level plastic feature vectors to promote the expression of model features.The final test positioning accuracy rate is 0.9825.

关 键 词：塑料 Yolo 残差模块 

分 类 号：TQ320.6[化学工程—合成树脂塑料工业]