Steel Surface Defect Detection Based On Deep Learning

Steel is a basic material of many infrastructure projects and products.It’ s very important to test and control the quality of steel regularly.However,the quality of steel is seriously affected by the raw materials and processing technology.Sometimes,there are many defects on the surface of steel,such as patches,scratches,these defects could shorten the service life of steel and accelerate the damage of equipment.Therefore,it’ s necessary to research the detection technology of the defects for the quality control of steel.This paper focuses on the detection of steel surface defects.In order to obtain higher detection accuracy and efficiency,the classical Faster R-CNN is improved,it makes up for the shortcomings of traditional detections methods The maj or work of this paper is as follows:1)In order to summarize the features of defects,a lot of images in NEU-CLS are analyzed.Aiming at the phenomenon of uneven illumination in the process of image acquisition,the histogram equalization method is used to improve the quality of original image by optimizing the gray distribution;The defects in the data set are marked,and the data set is divided for training.2)In order to improve the detection accuracy of the model for small-size targets,the Faster R-CNN was improved:the backbone of the Faster R-CNN was replaced by ResNet50 from VGG16;The feature pyramid network is also used in this paper,it brings RPN two groups of anchors which are suitable for small-scale targets;The ROI Pooling is replaced by ROI Align.ROI Aligns could eliminate the errors caused by the quantization operations,It makes the detection ability of Faster R-CNN to Small-size targets stronger.3)The Resnet50 is optimized to solve the problem of poor detection accuracy of multiscale targets:The convolution kernel which size is 7 × 7 in input layer is replaced by the superposition of three small convolution kernels,so that the ResNet50 can extract more details,and its ability to express nonlinear information will be better;A convolution kernel is added to the residual block in the lower layer of resnet50 to enhance the detection accuracy for multiscale targets,and the convolution kernel is decomposed to reduce the amount of calculation caused by the change.4)The dropout is added to the full connection layer in resnet50,it could help the model to improve the generalization ability and prevent over fitting:Adam optimizer is also used to optimize the model.it makes the model more robust to parameters.At the end of this paper,the feasibility of these methods is verified by a series of comparative experiments...