Research On Steel Surface Defect Detection Based On Convolutional Neural Network

In steel industry,steel surface defect detection is the key to ensure product quality.However,there are some problems in the actual defect detection,such as the imbalanced number of various types of defects and difficulties in locating and classifying defects.Therefore,in order to solve the above problems,obtain more accurate information of the type and location of steel surface defects,and improve the accuracy of steel surface defects detection,this paper studies the detection of steel surface defects based on convolutional neural network technology,and the main work is as follows:(1)For the difficult problem of defect location and classification in steel surface defect detection,a steel surface defect detection model GAUDeeplab based on improved Deeplabv3+ is proposed to obtain the type and location information of steel surface defects through the image semantic segmentation model.Among them,on the basis of the encoder-decoder structure of the Deeplabv3+ model,the residual network Res Net50 is used as the backbone network,while the attention mechanism module is introduced to improve the decoder structure,so that the high-level semantic information can effectively play a guiding role on the shallow detail information,more accurately select the detail information and recover the defect type information without losing the detail information in the process.The experimental results on the image data of steel surface defects provided by Severstal Steel Company show that the GAUDeeplab model can more accurately segment defects and process the details of defect edges more smoothly in the process of detecting various defect information.Compared with the original Deeplabv3+ model,the Io U is increased by 2.78%,and the Dice coefficient is increased by 0.66%.(2)In view of the problem that steel surface defect image samples are few and the number of various defects is unbalanced,data enhancement method is adopted to increase training samples,and Dice coefficient is adopted to reduce the negative impact of the imbalance in the number of samples for each type of defect.The classification of steel surface defects is regarded as a binary classification problem,and only the defective and non-defective classes are considered,and a combined loss function is proposed.The loss function consists of the Dice loss function with different weights and the binary cross entropy loss function,which can effectively improve the training effect during the model training process.(3)A multi-scale feature fusion module MFF is designed to capture fine-grained multi-scale feature information for the problem of easy loss of fine defect information in the process of steel surface defect detection.The module is added to the encoder of the GAUDeeplab model to propose a steel surface defect detection model based on multi-scale feature fusion.The experimental results on Severstal steel surface defect image data show that the multi-scale feature fusion module is effective in reducing the loss of fine defect information while improving the detection rate of steel surface defects.Compared with the GAUDeeplab model,the Dice coefficient of M-GAUDeeplab is increased by 0.42%,and the Io U is increased by 0.33%.