作者：

郭培林，陈金水，卢建刚（浙江大学，浙江 杭州 310058）

戴柏炯（杭州朝阳橡胶有限公司，浙江 杭州 310018）

史敦禹（中策橡胶（建德）有限公司，浙江 建德 311607）

孙洪林（杭州中策清泉实业有限公司，浙江 杭州 314100）

摘要：轮胎是我国国民经济的重要支柱，利用X光机对轮胎进行质量检测在整个轮胎生产过程中是极其重要的一道工序。目前国内工厂普遍采用肉眼观察轮胎X光图像进行识别，存在效率低下、人工成本高等一系列问题，因此采用计算机视觉技术进行自动识别是今后的发展方向。本文将目标检测算法Faster R-CNN应用于轮胎质检，并加以改进：（1）在模型中融合FPN（Feature Pyramid Network，特征金字塔网络），用以解决轮胎瑕疵尺度变化大的问题；（2）在算法中融合背景特征信息，对候选框进行重排名，增加检测模型最终的检测精度。通过对某轮胎厂提供的轮胎X光图像进行瑕疵检测对比表明，这些改进措施提高了检测模型的mAP（mean Average Precision）指标，具有良好的应用前景。

关键词：轮胎X光图像；瑕疵检测；深度学习；背景特征

Abstract: The tire industry is an important part of China's nationale conomy. The use of X-ray machines for tie quality inspection is an extremely important process in the tire production process.At present, domestic factories generally use the naked eye to observe tire X-ray images for recognition. There are a series of problems such as low efficiency and high labor cost. Therefore,automatic identification by computer vision technology is the future development direction. In this paper, the Faster R-CNN is applied to tire quality inspection, and the following improvements are made: Firstly, FPN is integrated in the model to solve the problem of large changes in tire defect scale; secondly, the background feature information is integrated in the algorithm, the RoIs is re-ranked to increase the final detection accuracy of the detection model. The comparison of defect detection on the tire X-ray image shows that these improvement measures improve the mAP value of the detection model and have a good application prospect.

Key words: Tire X-ray image; Defect detection; Deep learning; Background feature

在线预览：轮胎X光图像瑕疵检测Faster R-CNN算法改进研究

摘自《自动化博览》2020年8月刊