Belt Defect Detection Based on Deep Learning Methods
To address limited pixels,relatively weak features,and uneven distribution of defects in conveyor belt images,a con-veyor belt defect detection system is designed,and a label assignment strategy based on Gaussian mixture model is propopsed.The Gaussian distribution prior information of the feature receptive fields is applied to build the Gaussian model,which can adapt to defects in different belts through dynamic adjustment mechanisms,thereby effectively improving the detection capability for minor defects.The intersection over union with receptive field distance is replaced to measure the similarity between Gaussian receptive fields and true labels,the sample based on their similarity is allocated to effectively improve the accuracy of sample assignment.The Gaussian mixture model and expectation maximization algorithm are used to implement the probability distribution fitting,achieve the adaptive allocation of positive and negative samples for feature points,and effectively avoid the missed detections caused by minor defects.Ex-perimental results show that the Gaussian mixture model label assignment strategy increases a significant accuracy of conveyor belt de-fect detection,improving the accuracy by 3.8％compared to the baseline network.
NETL
NSTL
万方数据
