基于改进YOLOv4的烟条拉线头缺陷检测

针对烟条透明塑料外包装上拉线头的缺陷检测中传统图像处理的误报率高的问题,提出了一种基于YOLO深度学习算法和算法相结合并具有自学习优化功能的改进方案。即先使用传统Hough变换圆检测,将检测出缺陷的图像二次经过YOLO算法。为能将YOLO算法达到高性能、高精度的效果,本文对YOLO的网络结构进行改造,提出了AAS-YOLO(adaptive anchor size with YOLOv4),使其具备兼容动态尺度锚定边框的功能,可以实现将传统算法的部分计算结果作为自学习参数贡献给AAS-YOLO算法；并通过去除贡献低的BN通道,精简了网络结构,减少冗余计算。通过实验数据来看,改进后的AAS-YOLO算法提高了定位精度和检测速度；改进后的方案降低了拉线头缺陷检测的误报率。

A Defect Detection of Tobacco Rod Pull Ears Based on Improved YOLOv4

Aiming at the high false alarm rate of traditional image processing in the defect detection of the pull thread on the transparent plastic outer packaging of cigarette rods, an improved scheme based on the combination of the YOLO deep learning algorithm and the algorithm and with the self-learning optimization function is proposed. That is, the traditional Hough transform circle detection is used first, and the detected defect image is passed through the YOLO algorithm for the second time. In order to achieve high-performance and high-precision effects of the YOLO algorithm, this article reforms the network structure of YOLO and proposes AAS-YOLO (adaptive anchor size with YOLOv4), which is compatible with the dynamic scale anchor frame function, which can be realized Part of the calculation results of the traditional algorithm are used as self-learning parameters to contribute to the AAS-YOLO algorithm; and by removing the BN channel with low contribution, the network structure is simplified and redundant calculations are reduced. According to the experimental data, the improved AAS-YOLO algorithm improves the positioning accuracy and detection speed; the improved scheme reduces the false alarm rate of the defect detection of the cable head.