基于深度学习方法的传送带缺陷检测

Belt Defect Detection based on Deep Learning Approaches

摘要:

针对传送带瑕疵在图像中所占有的像素相对有限、特征相对微弱且分布不均匀的问题，设计了一个传送带缺陷检测系统，并提出了一种基于高斯混合模型(GMM, gaussian mixture model)的标签分配策略；利用特征感受野遵循高斯分布的先验信息进行高斯建模，并通过动态调整机制适应不同尺度的传送带缺陷，能够更有效地提升对微小瑕疵的捕捉能力；使用感受野距离取代交并比来衡量高斯感受野和真实标签的相似度，并通过二者之间的相似度来分配样本，从而有效提高了样本分配的准确性；使用高斯混合模型并通过期望最大化(EM, expectation-maximization)算法拟合概率分布，实现了对特征点的自适应正负样本分配，能够有效避免微小瑕疵特征微弱所导致的漏检问题；结果表明，高斯混合模型标签分配策略对传送带缺陷检测精度的提升十分明显，相对于基准网络，精度提升3.8%。

Abstract:

To address the limited pixel coverage, relatively weak features, and uneven distribution of defects in conveyor belt images, a conveyor belt defect detection system is designed, accompanied by a label assignment strategy based on Gaussian Mixture Model. Leveraging the Gaussian distribution prior information following the feature receptive fields, a Gaussian model is constructed to adaptively capture conveyor belt defects of varying scales through dynamic adjustment mechanisms, thereby enhancing the detection capability for minor defects effectively. Replacing the Intersection over Union with receptive field distance, the similarity between Gaussian receptive fields and true labels is measured, facilitating sample allocation based on their similarity, thereby improving the accuracy of sample assignment effectively. Utilizing Gaussian Mixture Model and Expectation-Maximization algorithm for probability distribution fitting, adaptive allocation of positive and negative samples for feature points is achieved, effectively mitigating the issue of missed detections caused by faint features of minor defects. Results demonstrate a significant enhancement in the accuracy of conveyor belt defect detection attributed to the Gaussian Mixture Model label assignment strategy, exhibiting a 3.8% improvement in accuracy compared to the baseline network.