【作者】 周晓宇；

【导师】 刘成明；

【作者基本信息】 郑州大学 ， 工程硕士（专业学位）， 2024， 硕士

【摘要】 瓷砖成品的质量易受技术和生产条件等因素的影响,其中表面缺陷是影响质量的关键因素。随着计算机视觉和深度学习的不断发展,基于深度学习的瓷砖瑕疵检测研究成为主流方法,但实际应用中仍面临目标瑕疵尺寸小、背景复杂和样本分布不均衡等诸多挑战,导致检测难度增加。此外,在某些工业应用场景下,边缘设备计算和存储能力有限,这对瓷砖瑕疵检测算法的参数量和计算量都提出了额外要求。因此,在工业生产中快速准确地检测瓷砖瑕疵成为了一个亟待解决的问题。针对上述问题,本文主要研究工作如下:(1)针对通用目标检测模型在瓷砖瑕疵检测任务中表现较差的问题,本文提出了一种双阶段瓷砖表面瑕疵检测算法(TDD)。首先,该算法以Cascade RCNN模型为基础架构,在主干网中加入可形变卷积(DCN),提升网络对各种尺度和形状瑕疵的特征表达能力。其次,本文还提出了一种多尺度特征融合网络(BAFPN),用作检测模型的颈部,旨在增强小目标的特征感知能力。在BAFPN中,本文采用了 BADConv卷积以缓解跨尺度融合的混叠效应,并采用BAPP模块提升网络的上下文感知能力。最后,通过级联检测头的三次目标框微调,得到更精准的检测结果。实验结果表明,TDD模型在MS COCO数据集的小目标评价指标上提高了 1.3个百分点,在天池瓷砖瑕疵数据集上mAP提高了 1.9个百分点,验证了该模型的有效性。(2)针对工业瓷砖瑕疵检测任务中对算法实时响应要求高以及边缘设备计算能力和存储资源有限的问题,本文提出了一种基于GFocal的轻量级瓷砖瑕疵检测方法。该方法以FasterNet为主干,提高了模型的粗特征提取能力并加快了检测速度。其次,在多尺度特征融合部分引入了 DFC内卷卷积,既减少了模型参数,又提高了模型的特征融合能力。然后,采用了 Quality Focal Loss和Distribution Focal Loss函数,解决了模型分类和质量预测的不一致性和不灵活性问题。最后,采用知识蒸馏的BCKD模块将骨干网替换为参数量更小的MobileNetv2。在天池瓷砖数据集上对本方法进行了实验评估,相比基准模型,改进后的模型参数量减少了 73%,推理速度达到了 26FPS,且达到了 77.9mAP的检测精度。因此,本算法在提高检测精度的同时,兼顾了参数量和推理速度。更多还原

【Abstract】 The quality of finished tiles is susceptible to factors such as technology and production conditions,among which surface defects are the key factors affecting quality.With the continuous development of computer vision and deep learning,research on tile defect detection methods based on deep learning has become a mainstream method,but practical applications still face many challenges,such as small defect object sizes,complex backgrounds,and uneven sample distributions,which leads to increased detection difficulty.In addition,in some industrial application scenarios,the edge devices have limited computational and storage capacity,which puts additional requirements computational and storage capacity of devices puts additional requirements on parameters and computation of tile defect detection algorithms.Therefore,fast and accurate detection of tile defects is an urgent problem in industrial production has become an urgent problem.To address the above problems,the main research work of this thesis is as follows:(1)In response to the problem of poor performance of general object detection models in tile defect detection tasks,this thesis proposes a two-stage tile surface defect detection algorithm(TDD).Firstly,the algorithm takes the Cascade RCNN model as the infrastructure and adds deformable convolution(DCN)to the backbone network to improve the feature expression ability of the network for defects of various scales and shapes.Secondly,this thesis proposes a multi-scale feature fusion network(BAFPN)as the neck of the detection model,aiming to enhance the feature sensing ability for small objects.In BAFPN,this paper employs BADConv convolution to mitigate the aliasing effect of cross-scale fusion and the BAPP module to enhance the contextawareness capability of the network.Finally,the more accurate detection results are obtained by three bounding box fine-tuning of the cascade detection head.The experimental results show that the TDD model improves the small object evaluation index by 1.3%on the MS COCO dataset,and the mAP improves by 1.9%on the Tianchi tile defect dataset,which verifies the effectiveness of the model.(2)In response to the problems of high requirements for real-time response of algorithms as well as limited computing power and storage resources of edge devices in industrial tile defect detection tasks,this thesis proposes a lightweight method for detecting tile defects based on GFocal.The method uses FasterNet as the backbone,which improves the coarse feature extraction capability of the model and accelerates the detection speed.Secondly,DFC involution is introduced in the multi-scale feature fusion part,which both reduces the model parameters and improves the feature fusion ability of the model.Then,the method employs Quality Focal Loss and Distribution Focal Loss functions to address the inconsistency and inflexibility of classification and quality prediction.Finally,the knowledge distillation module BCKD is used to replace the backbone with MobileNetv2 with smaller parameters.The experimental evaluation of this algorithm is carried out on the Tianchi Tile defect dataset,and compared with the baseline model,the improved model reduces the parameters by 73%,achieves an inference speed of 26 FPS,and achieves a detection accuracy of 77.9 mAP.Therefore,the present algorithm balances the parameter quantity and inference speed while improving the detection accuracy.更多还原