Research On Steel Surface Defect Detection Method Based On Deep Learning

The steel industry,as one of China’s pillar industries in several industrial disciplines,provides items that influence the quality of products in downstream industries.Due to the influence of factors such as the workshop environment,process technology,and raw materials,numerous types of flaws inevitably emerge on the surface of strip steel during the production process.These flaws have varied degrees of impact on the strip steel’s surface quality and mechanical properties.However,present steel surface defect detection is primarily accomplished by manual hand-held inspection equipment,which is more complex and inefficient,and cannot meet the requirements of real-time,precise detection,and quick feedback to the production line.Due to its adaptation to industrial situations,the deep learning-based defect detection approach provides the benefits of accurate detection and quick inference.Therefore,in order to meet the requirements of fast and accurate steel surface defect detection,this paper investigates the deep learning based steel surface defect detection method,and the main research contents are as follows:(1)The current challenges of defect detection with complex and variable flaws,minimal inter-class variance but substantial intra-class variation,are revealed by an analysis of NEU-DET,an open source dataset of steel surface defects.Rotation,mirroring,and brightness dithering are used in this paper to enhance the dataset and give data support for the following creation of defect detection models using deep learning approaches.(2)The basic model for steel surface defect detection is constructed by experimentally comparing the Efficient Det series algorithm with other mainstream target detection algorithms.In the experimental process,firstly,according to the sample point size and area ratio distribution of the steel surface defect data set,the Kmeans clustering algorithm is used to optimize the anchor frame design,secondly,the Adam optimization algorithm is selected to replace the SGD optimization algorithm for model training and tuning through comparison experiments,and finally,the experimental results show that the trained Efficient Det-D0 algorithm has higher accuracy and detection speed.(3)On the basis of the Efficient Det-D0 model based on the anchor-free idea,an Efficient Det-Free defect detection algorithm is proposed.First,a deformable convolutional block is created in this paper to reconstruct the last two layers of the feature extraction network in order to improve the model’s extraction capability for irregular defect features.Second,to address the confusing effect caused by stacking more layers of the feature fusion network and the loss of some semantic information during upsampling,we simplify the feature fusion network and embed a global context module,as well as construct an enlarged convolutional module to replace the lateral jump link.Finally,a pixel-by-pixel prediction approach is used to optimize the prediction network based on the anchor-free idea,and a GIo U loss function is used to replace the Smooth L1 loss function for the phenomena that the prediction frame does not intersect with the real frame during training,and a centrality influence factor is added to the prediction network.According to the comprehensive results,the improved algorithm is given in this research can reach 78.68 % m AP and 26.7 FPS,which is a6.45 % improvement in accuracy over the original model.