Research On Defect Images Classification Of Steel Surface Based On Deep Active Learning

Steel surface defects will seriously affect product quality.Traditional manual detection and machine vision methods based on image processing are inefficient,while traditional deep learning methods based on convolutional neural networks rely on a large number of labeled samples,and defect labeling in the industrial field requires strong professional knowledge,and the cost of sample labeling is high.In order to reduce the cost of labeling,improve the utilization of labels and the efficiency of model classification,and use fewer label samples to achieve high-precision identification of image defects in a short period of time,this paper takes the surface defects of steel plates as a breakthrough,and combines the active learning method with deep learning.The following researches are carried out on the efficient and rapid identification method of the surface defect pictures of the steel plate:1.The open source hot-rolled steel sheet surface defects and self-built milled steel sheet surface defects are used as experimental data;the operating environment for the defect image classification and recognition experiment is built,and the performance evaluation standards of each defect classification model in this paper are introduced.2.Aiming at the over-fitting problem in the classification of steel plate surface defects based on deep learning,a defect image classifier fused with structural regularization terms is proposed as an improved regularization method,which can enhance the regularization ability of the model and reduce training.time and computational cost.Using the NEU(Northeastern University)hot-rolled steel plate defect data set,the deep learning model required for the experiment was designed,and the regularization ability of the proposed improved method was verified by experiments.3.It is proposed to introduce the active learning method using uncertainty measurement into the deep learning training process.Through the analysis of the training process,an improved defect classification model is proposed by combining two types of uncertainties: model’s uncertainty on labeled defect sample set and on unlabeled defect sample set.This method can further reduce the cost of manual labeling and improve the utilization rate of images with labels.The effectiveness of the proposed improved method is verified by experiments.4.The proposed improved regularization method and improved defect image sampling method are experimentally verified by using the self-built milling steel plate surface defect data set.The original collected data set is amplified by geometric methods such as scaling and rotation,and the amplified defect data is divided into training set and validation set according to the ratio of 80% and 20%,and the original collected defect data is used as the test set.Through experiments to analysis the classification effect of improved regularization method in deep active learning and the difference between improved defect sampling method and traditional methods.Theoretical analysis and relevant tests results demonstrate that: the surface defect classifier based on the structure regularization term proposed in this paper can improve the training speed while ensuring the classification accuracy,and effectively alleviate the overfitting problem of the deep learning model in the classification of steel surface defects.At the same time,the proposed improved defect sampling method is also better than the traditional uncertainty sampling method,and can achieve better classification results under the condition of fewer labels.The deep active learning algorithm proposed in this paper,which combines the improved regularization method and the improved defect sampling method,can achieve efficient classification of steel plate surface defects while ensuring accuracy,and can effectively reduce the cost of labeling.