Research On Rapid Detection Method Of Strip Steel Surface Defects Based On Deep Feature Fusion

The surface quality problems of strip steel can seriously affect the quality of strip steel downstream products,such as automobiles,household appliances and precision electronics,causing great economic losses.Therefore,the development of online strip steel surface detecting system to quickly and accurately detect surface defects and defect types to promote process improvement and thus improve strip steel product quality has become an urgent need for steel companies to improve their core competitiveness.The excellent performance of artificial intelligence in the field of vision has driven the rapid development of online strip steel surface defect detection technology based on deep learning.However,with the continuous improvement of strip rolling technology,these methods still have problems such as data redundancy,low accuracy of defect classification and recognition algorithms,weak robustness,and complex network model.In this dissertation,to address the above problems,a strip steel surface defect detection method based on joint optimization of feature extraction and clustering,deep feature fusion,and model lightweighting is carried out with strip steel surface defect images as the research object.At the same time,an experimental platform for strip surface defect detection is built to further validate the proposed method.The main work of this dissertation includes the following four parts.(1)A joint feature extraction and clustering optimization method for the classification of strip steel surface with and without defectsTo address the problems of weak feature learning ability and poor classification accuracy when unsupervised learning is used to classify the strip steel surface with or without defects in the face of massive strip steel surface images,a joint optimization method based on feature extraction and clustering tasks is proposed to classify the strip steel surface with or without defects.The split-attention network-based image feature extraction method is constructed,combined with the Mini-Batch K-Means clustering method to obtain the pseudo-label information of the current surface image,and the optimized feedback of the joint classification pseudo-label and the back propagation of the loss of the Softmax classification model to optimize the feature extraction network performance,improve the category of the feature vector of the defective surface image and the defect-free surface image The proposed network can improve the class differentiation of feature vectors of defective and non-defective surface images,and achieve fast classification of defect-free and defective images.The experimental results show that the proposed method has a leakage rate of only about 5% for the classification of the strip surface with and without defects,which greatly reduces the amount of subsequent data processing in the detecting system and improves the detection efficiency and accuracy.2)Depth feature fusion-based strip steel surface defect recognition methodIn view of the problem of low accuracy of defect recognition due to the large scale span,inconspicuous features and diverse morphology of strip steel surface defects,a method based on depth feature fusion is proposed for the recognition of strip steel surface defects.A coder-decoder network structure based on improved split attention and feature pyramid is constructed,and a depth feature fusion module based on group normalization is designed to further extract and fuse the depth implicit features of strip steel surface defects to achieve accurate segmentation and recognition of defects.The experimental results show that the Dice of the proposed method can reach 94.24%,which provides a strong guarantee for optimizing the process regulation and thus improving the surface quality of strip steel products.(3)Lightweighting method of strip steel surface defect detection modelTo address the problems of computational time consumption,accuracy and speed of deep learning network models,an adaptive pruning method integrating ADMM and SA is proposed to perform adaptive pruning and compression of the strip steel surface defect classification and recognition model,which greatly reduces the number of parameters and computation of the model,improves the streamlining and deployability of the model,and realizes the lightweight of the detection model while ensuring the detection accuracy.Experimental results show that the number of parameters of the defect classification and identification model decreases by 45.22% and 47.23%,the computation volume decreases by 41.68% and 43.96%,and the average test time decreases by 31.02% and25.93%,providing a strong guarantee for the rapid detection of strip steel surface defects.4)Design of strip steel surface defect detection systemCombined with the actual conditions of the production site,an experimental platform for online inspection of strip steel surface defects was built to further verify the effectiveness of the proposed method.The experimental results further show that the strip steel surface defect detection method proposed in this dissertation can effectively detect the strip steel surface defect images with high accuracy,low leakage rate and false detection rate,and good robustness to the complexity and randomness of the strip steel surface defect images.