Research On Steel Defect Object Detection Algorithm Based On Object Shape Diversity And Data Imbalance

Steel defect object detection can assist the factory to complete the quality control inspection in the production process.In image-based defect detection,the defect can be regarded as a kind of object contained in the image.Defect detection refers to locating the position of the defect from the image and classifying the type of defect.Defect objects with morphological diversity causes that the size and shape of the ground-truth boxes of the defect object are quite different.It is difficult for anchor-based algorithm to detect defect objects with large shape changes by simply setting the anchor box size parameter.And there are some detection boxes with high classification score but low location quality that do not fit well with the shape of the object.Besides,defect tasks usually have data imbalance problems,and the proportion of data in some categories is quite different.The majority instance category has training tendency,resulting in low detection accuracy of the minority instance category.According to the shape diversity of defect objects and data imbalance,the main work of this paper is as follows:For the problem that the defect objects have shape and scale diversity in the defect detection task,this paper designs a label assignment module by utilizing the object shape and scale.Specifically,when selecting the candidate positives,the proposed module considers shape difference between ground-truth and anchors in addition to the center distance,and assign different number of candidate positive samples for different size of objects to focus more on the characteristics learning of large-sized objects in the training process.In order to reduce the number of detection boxes with high classification scores but low localization quality,this paper designs a classification score correction loss function fused with localization information,so that positive samples with high localization quality can obtain high classification scores,and the detection boxes with higher localization quality are easier to match with the ground-truth boxes.For the data imbalance of different categories,this paper adopts two model ensemble methods for balanced training:Method 1 trains a singleclass weak detector of the majority instance category,and integrates with a weak detector of the minority instance category and other categories into a strong detector;Method 2 trains a single-class weak detector of the majority instance category,a single-class weak detector of the minority instance category,and a multi-class weak detector for other categories,then integrate them into a strong detector.By training the majority instance category and the minority instance category separately,the training tendency of the model to the majority instance category is alleviated when the two are jointly trained.In addition,for the small amount of defect data,a hybrid data augmentation method is designed for the defect dataset,to increase the data diversity of a single image.And this paper uses the semi-supervised pre-training model method to perform category-free training on the defect dataset,and then uses the model transfer method to fine-tune the parameters to improve the detection accuracy of the minority instance category.The experimental results show that the algorithms in this paper can increase the successful matching probability of the detection boxes with a more similar shape to the ground-truth boxes,and thus improve the detection accuracy.The balanced training strategy for the defect dataset can effectively improve the accuracy of the minority instance category,then improve the average detection accuracy of the model.