【作者】 张振；

【导师】 彭晨；

【作者基本信息】 上海大学 ， 控制科学与工程， 2023， 硕士

【摘要】 根据《“十四五”智能制造发展规划》的目标,轮胎制造产业需要从传统的粗放式生产向自动化、信息化、数字化的智能生产转变。在这个转变过程中,基于轮胎X光图像的质检流程自动化成为整个产线智能升级改造最关键的难题。然而,受制于有限的装备技术储备水平,现阶段国内厂商仍采用人工方式进行缺陷检测。为了解决人工检测所面临的成本高和效率低等问题,本文结合计算机视觉领域中的先进的目标检测技术,针对轮胎X光图像进行了缺陷检测算法研究,为轮胎产线智能升级改造提供助力。本文的主要工作内容如下:(1)针对轮胎人工质检漏检和误检率高等问题,本文提出了一种CNSYOLOv5s缺陷检测算法。首先,采用优化后的CNB(Convolution with a Larger Receptive Field and Narrower Bottleneck)模块替代主干网络中的C3模块,增强主干网络的特征提取能力,减少细节信息丢失;然后引入三维、无参的Sim AM注意力机制模块,提高特征融合网络对缺陷特征信息的关注度,改善小目标缺陷的检测能力;预测层采用CIo U＿Loss损失函数,加速网络模型的收敛。与改进前算法相比,提出的改进算法平均准确率指标(m AP)从79.5%上升到82.9%,平均回收率指标(AR)从72.5%上升到85.4%。进一步的消融实验验证了改进的算法既能够提取更多缺陷特征信息,也能够更多关注有用特征信息,实现缺陷的有效检测。(2)针对轮胎缺陷数据集小及数据不平衡等问题,本文提出了一种基于Retinex-CGAN的二维类内Mixup数据增强算法。考虑到轮胎X光图像整体低灰度值会导致整体图像模糊,前景帘线和背景的低对比度会导致缺陷特征不明显等实际问题,首先采用基于Retinex的Auto＿MSRCR算法对X光图像进行去模糊和对比度拉伸,在此基础上训练无监督的Cycle GAN网络以生成更多的保留缺陷特征的缺陷图像,最后将生成的缺陷图像与原始缺陷图像进行二维类内Mixup融合,增强缺陷特征的同时,实现了扩充缺陷样本集的目的。与数据增强前对比,本文提出的检测算法的m AP和AR指标分别提高了1.0%和2.1%,实验验证了该增强算法的有效性。(3)针对已有单通道模型鲁棒性差的问题,搭建了多通道融合的缺陷检测模型。考虑到实际生产中轮胎类型的多样性,单通道算法往往受到X光机成图质量低等因素的不利影响,考虑到算法的健壮性,构建多通道融合的缺陷检测模型。分别对图像进行不同的预处理,输入到多个单通道模型中分别进行训练。预处理操作包括直方图均值化、伽马变换、线性拉伸等。使用非极大值抑制算法处理多个通道模型的预测输出,将融合后的输出作为最终检测结果。实验显示提出的多通道模型的健壮性更高,m AP指标提升了0.3%,AR指标提升了0.5%。基于上述理论与方法,研发了轮胎缺陷智能检测软件系统。该软件系统采用服务端、业务端、数据段三层MVC设计架构,实现了对轮胎运行状态实时检测及可视化界面展示、缺陷实时报警、离线模型更新,并通过数据库保存缺陷信息记录,便于企业管理人员查询统计。更多还原

【Abstract】 According to the goals of the "14th Five-Year Plan" intelligent manufacturing development plan,the tire manufacturing industry needs to transform from traditional extensive production to automated,informationized and digital intelligent production.In this transformation process,the automation of the quality inspection process based on tire X-ray images has become the most critical problem for the intelligent upgrade of the entire production line.However,subject to the limited level of equipment technology reserves,domestic manufacturers still use manual methods for defect detection at this stage.In order to solve the problems of high cost and low efficiency faced by manual inspection,this paper combines advanced object detection technology in the field of computer vision to study the defect detection algorithm for tire X-ray images,so as to provide assistance for the intelligent upgrading of tire production lines.The main work of this article is as follows:(1)Aiming at the problems of tire manual quality inspection leakage and high false detection rate,this paper proposes a CNS-YOLOv5 s defect detection algorithm.Firstly,the optimized CNB(Convolution with a Larger Receptive Field and Narrower Bottleneck)module is used to replace the C3 module in the backbone network to enhance the feature extraction capability of the backbone network and reduce the loss of detailed information.Then,a three-dimensional and parameter-free Sim AM attention mechanism module is introduced to improve the attention of the feature fusion network to defect feature information and improve the detection ability of small target defects.The prediction layer uses the CIo U＿Loss loss function to accelerate the convergence of the network model.Compared with the pre-improvement algorithm,the average accuracy index(m AP)of the proposed improved algorithm increased from 79.5%to 82.9%,and the average recovery index(AR)increased from 72.5% to 85.4%.Further ablation experiments verify that the improved algorithm can not only extract more defect feature information,but also pay more attention to useful feature information to achieve effective defect detection.(2)Aiming at the problems of small tire defect data set and data imbalance,this paper proposes a two-dimensional intra-class Mixup data augmentation algorithm based on Retinex-CGAN.Considering that the overall low gray value of the tire X-ray image will lead to the overall image blurring,and the low contrast of the foreground cord and background will lead to the lack of obvious defect features,firstly,the Auto＿MSRCR algorithm based on Retinex is used to deblur and contrast stretch the Xray image,on this basis,an unsupervised Cycle GAN network is trained to generate more defect images that retain defect features,and finally the generated defect image is fused with the original defect image by two-dimensional intra-class mixup.While enhancing the defect characteristics,the purpose of expanding the defect sample set is realized.Compared with the data enhancement,the m AP and AR indexes of the proposed detection algorithm are increased by 1.0% and 2.1%,respectively,and the effectiveness of the enhancement algorithm is verified by experiments.(3)Aiming at the problem of poor robustness of existing single-channel models,a multi-channel fusion defect detection model is constructed.Considering the diversity of tire types in actual production,the single-channel algorithm is often adversely affected by factors such as low X-ray machine image quality,and considering the robustness of the algorithm,a multi-channel fusion defect detection model is constructed.Different preprocessing of images is performed and input into multiple single-channel models for training.Preprocessing operations include histogram meaning,gamma transform,linear stretching,and more.The prediction output of multiple channel models is processed using the non-maximum suppression algorithm,and the fused output is used as the final detection result.Experiments show that the robustness of the proposed multi-channel model is higher,with the m AP index increased by 0.3% and the AR index increased by 0.5%.Based on the above theories and methods,the intelligent detection software system of tire defects of Shandong Zaozhuang Ba Yi rubber tire factory was developed.The software system adopts the three-layer MVC design architecture of server,business and data segment,which realizes real-time detection and visual interface display of tire running status,real-time alarm of defects,offline model update,and saves defect information records through the database,which is convenient for enterprise managers to query statistics.更多还原