Research On Steel Defect Detection Algorithm Based On Deep Learning

As the basic raw material of modern industry,steel is widely used in national defense,petroleum,construction,equipment,and other fields.In the process of producing steel products,due to factors such as raw material impurities,excessive cooling,manufacturing technology,etc.,there will be cracks,patches,scratches,inclusions,or other kinds of defects on the produced steel.Therefore,defect detection is an important part to ensure the quality of steel products.In order to introduce intelligent and automated technology into steel defect detection,improve the efficiency of steel defect detection,and reduce labor costs,an in-depth study of the detection algorithm of steel defects is carried out in this paper based on the computer vision method.Firstly,aiming at the detection difficulties of steel defect data,combined with the research findings in the field of general target detection,an improved method in a single domain is proposed.In addition,the training of deep learning models relies on a large amount of labeled data,which is very time-consuming and labor-intensive to collect and label,but due to the problems of too many small targets and defects don't have obvious boundaries,it is particularly difficult to obtain lots of target domain labeled data,which is not conducive to the large-scale deployment of intelligent product lines.This paper makes an in-depth study on this pain point in industrial production.Through the domain adaptive detection model proposed in this paper,other domain data are used to improve the effect on the target domain,so as to reduce the need for labeling data on the new production line.To sum up,the research contents of this paper are as follows:1.For the specific steel defect detection scenario,based on the general target detection framework Faster R-CNN model,targeted improvement methods are proposed from the data level and model level to alleviate the difficult problems such as complex background,prominent small target problems,and blurred boundary in steel defect detection,so that the model achieves good results in the task of steel defect detection.2.In order to alleviate the difficulty of obtaining a large amount of labeled data in steel defect detection,based on the transfer learning theory and DA Faster R-CNN model,a domain adaptive detection model based on pseudo tags and feature weighted alignment by category is proposed in this paper.The model combines the idea of domain adaptation at the instance level and feature level at the same time.In the case of insufficient target domain data,by taking other production line data as auxiliary training data,and overcoming the problem of inconsistent instance label space in different domains,the detection effect in the target domain is effectively improved,thereby reducing the dependence on the target domain data.3.For the algorithm improvement in the single domain scenario and cross-domain scenario proposed in this paper,sufficient comparative experiments and ablation experiments are done to verify the effectiveness of the improved method.