基于改进Faster-RCNN的烟标缺陷检测  

Cigarette Case Defect Detection Based on Improved Faster-RCNN

作　　者：李殷昊 李莹[1] 王凯彬 何自芬[1] LI Yin-hao;LI Ying;WANG Kai-bin;HE Zi-fen(Faculty Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学机电工程学院,云南昆明650550

出　　处：《软件导刊》2023年第3期179-183,共5页Software Guide

基　　金：国家自然科学基金项目(61761024)。

摘　　要：针对品检机上的烟标缺陷检测算法进行改进,根据烟标缺陷检测固有特性,融合目标检测算法,提高烟标缺陷检测精度,实现传统烟标缺陷检测算法不能实现的实时缺陷分类。用改进后的Faster-RCNN代替传统的模版匹配,对CCD线阵相机采集到的4K烟标图像进行烟标缺陷检测,通过加入特征图金字塔网络FPN,烟标缺陷检测准确率上升1.4%。在FPN中使用内卷卷积,对比不同大小内卷卷积核的烟标缺陷检测精度发现,使用5×5的内卷卷积核的烟标缺陷检测精度提升最大,达1%,且使用内卷卷积模型的单张烟标检测时间未增加。最后通过迁移学习使烟标缺陷检测精度进一步提升,达97.9%。结果表明,所提算法能够实现烟标缺陷检测定位和分类,且具有较高准确率。The cigarette case defect detection algorithm on the quality inspection machine is improved,and the objection detection algorithm is integrated according to the inherent characteristics of the cigarette case defect detection to improve the accuracy of the smoke case defect detection.The improved Faster-RCNN is used instead of the traditional template matching,and the 4K cigarette case image collected by the CCD line scan camera is used for cigarette case defect detection.By adding the FPN layer,the accuracy of cigarette case defect detection increased by 1.4%.By using involution in the FPN,by comparing the detection accuracy of cigarette case defects of different sizes of involution kernels,it is found that the detection accuracy of cigarette case defects using 5×5 involutional kernels is greatly improved,increasing by 1%,and the detection time of a single cigarette case using the model of internal convolution did not increase.Finally,the accuracy of cigarette case defect detection is further improved through transfer learning,reaching 97.9%.The results show that the proposed algorithm can realize the positioning and classification of cigarette case defect detection,and has a high accuracy rate.

关 键 词：图像处理 CCD线阵相机 烟标缺陷检测 内卷卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]