Research On Steel Plate Surface Defect Detection Based On Deep Learning

The grade and the performance of steel plate products can be affected by the quality of steel plate directly.It is very important to develop the steel plate manufacturing industry whether there are defects in the steel plate surface.Considering that there are some drawbacks in the defect detection algorithms based on traditional machine vision,including low inspection speed and poor reasoning speed.In this thesis,a new deep learning algorithm is proposed based on cascade region-based convolutional neural network(Cascade R-CNN).It has many merits such as high detection rate and high inspection speed.Thus,the proposed algorithm has important research value.The main research contents are as follows.(1)Since there are some problems which are illumination insufficient,uneven illumination and low contrast in the steel surface defect image,an image enhancement algorithm is proposed in this thesis.Firstly,the steel surface defect image is decomposed by the wavelet transform into two parts which are high and low frequency,respectively.Secondly,the Retinex algorithm based on guided filtering is used to enhance the low-frequency part.At the same time,the soft and hard threshold denoising algorithm is used to enhance the high-frequency part.Then the inverse wavelet transform is used for image reconstruction,and the reconstructed image is transformed by gamma transform.(2)To solve the problem that the data types of six types of steel plate surface defects are complex and the shapes are variable,visualization methods such as tables,histograms,and sector charts are proposed to analyze the distribution and scale characteristics of various types of defects.To select the optimal detection model,the comparative experiments are carried out in the steel plate surface defect data set.In the experiments,we mainly focus on comparing the third version of you only look once(YOLOv3)、single shot multibox detector(SSD)、faster region-based convolutional neural network(Faster R-CNN)and Cascade R-CNN.According to measurement indicators and visualization results,we can conclude that Cascade R-CNN shares better detection performance.The result indicates that Cascade R-CNN is more is practical in the defect detection on steel plate surface.(3)Aiming at the problems of low efficiency of steel plate surface defect detection and high missed detection rate and false detection rate of small target defects,an algorithm for steel plate surface defect detection based on improved Cascade R-CNN is proposed.Firstly,the data set is enhanced,and the original cascade R-CNN model is improved,Res Ne Xt-101-64 × 4D is used to enhance the feature extraction capability of the network.Then the recursive feature pyramid network(RFP)is proposed to optimize the features in the way of feedback connection,so as to better retain the details and semantic information.At the same time,a switchable hole convolution sac is proposed to replace the convolution layer of the backbone network to improve the detection performance by changing the receptive field.Finally,the soft-non maximum suppression(Soft-NMS)algorithm is introduced to reduce the influence of redundant frames on the network model.