Research On Steel Defect Detection Based On Self-Supervised Contrastive Representation Learning

Steel is an important industrial raw material that plays a vital role in a range of industries,including industrial manufacturing and construction,so steel production capacity is vital to the economic development of countries around the world.It is important to note that steel surface defect detection technology has become an important factor in the steel industry’s drive to improve productivity and move toward automation,and it plays an essential role in the quality control of steel production.However,the automated detection of steel surface defects is still a challenging task so far.Steel surface defect detection refers to the detection of scratches,cracks,foreign bodies,corrosion,holes,etc.on the steel surface.At present,most companies are still using manual inspection methods for steel surface defects product detection,this completely human-dependent detection method is inefficient.With the development of artificial intelligence technology,many deep learning-based defect detection methods have been proposed and applied to various defect detection fields,but most of such methods are based on supervised learning that rely on a large amount of labeled data.Steel surface defects have single color,small defect targets and many types of defects appear simultaneously,which makes it difficult for non-professionals to distinguish and summarize various types of defect samples,making it extremely expensive to obtain labeled data in this field.The lack of labeled data will greatly affect the effectiveness of steel defect detection based on deep learning methods.In this paper,we construct a novel self-supervised steel surface defect detection model by learning better embedding feature representation of the defect on large amounts of unlabeled steel surface defect data,which can achieve excellent results in downstream steel surface defect detection tasks.The main research of this paper is as follows:(1)The image embedding strategy commonly used in contrastive learning methods destroys the spatial structure of images and is not suitable for pre-training for steel surface defect detection tasks.Therefore,we retain convolutional feature maps to mine robust local and spatial features of steel surface defect data,which can enhance the representation of the upstream model and make it suitable for transfer to the steel surface defect detection task.Since the data change from vectors to feature maps in the subsequent mapping structure,we design to take advantage of the convolutional structure for mapping instead of the original MLP head.(2)In order to eliminate the negative impact of random enhancement of contrastive learning,which produces noise on the multi-target coexistence dataset,this paper introduces the discrete Earth Mover’s Distance(EMD)metric to measure the contrastive matching similarity of targets.Meanwhile,the semantic correspondence between different targets is established to mitigate the influence of noise during the training,and these strategies can help overcome the problem that the contrastive learning method using random enhancement is not suitable for steel surface defect datasets.(3)Extensive experiments prove that the self-supervised steel surface defect detection framework(SCRL-EMD)proposed in this paper can achieve superior results as compared to state-of-the-art approaches on two public downstream defect datasets,NEU-DET and GC10-DET.Compared to the baseline model,it achieves 4.1% and 6.8%m AP improvement on the two datasets,respectively.Besides,this paper also validates that our model can achieve higher accuracy improvement on smaller downstream datasets,which reveals the greater potential of our approach in utilizing more readily available unlabeled data.