Improved Deep Residual Network And Its Application To Strip Surface Defect Identification

Cold rolled strip is an important product in the modern steel industry and has more applications in many industries such as automobile manufacturing,appliance manufacturing,container making,light industry,etc.Surface defects in strip steel during the production process are an important factor affecting product quality and can cause great economic losses to steel companies and product users.Traditional defect detection methods have the disadvantages of low accuracy,low efficiency and high cost,making it difficult to meet the needs of steel companies.In recent years,deep learning has shown great advantages in machine vision,but deep learning algorithms in online inspection of strip steel have problems such as spatial redundancy of defect image data,inconspicuous defect features,and large span of defect sizes,which cannot fully meet the needs of industrial sites.In response to the above problems,this thesis carries out the research of strip steel surface defect image recognition method based on Deep Residual Networks(Res Net)for spatial redundancy problem and multi-scale feature fusion,and the specific work is as follows:To address the problem of spatial redundancy in high-resolution strip defect images,a fast identification method of strip surface defects based on Learnable Image Resizer(LIR)and Glance and Focus Network(GFNet)is proposed.First,based on the characteristics of Dynamic Neural Networks(DNN),this dissertation proposes a GFNet-driven classification model for strip steel surface defects,which can adaptively process task-relevant regions according to different samples,significantly reduce the computational effort in the model inference phase,and facilitate the deployment of the model.Then,a joint training method of LIR and classification model GFNet is used to adjust the image size while achieving feature enhancement for the recognition model,and the enhanced image features can improve the accuracy of the classification model.Finally,Res Net-50 is used as the backbone network of the proposed method,and the effectiveness of the method is verified on the homemade dataset CR-CLS.The results show that,compared with vanilla Res Net-50,the proposed method can reduce the inference time of a single image by about 3.58 times and the model computation by about 6.11 times without sacrificing accuracy.Also,the experimental results on the publicly available dataset XLData-CLS demonstrate the effectiveness of the method.To address the problem of low defect recognition rate caused by the large span of defect size and inconspicuous defect features on the strip surface defects,a highprecision recognition model of strip surface defects based on SEPy Res Net(SENetPy Conv-Res Net)is proposed.First,a new convolution module SEPy Conv is constructed,which adopts Pyramidal Convolution(Py Conv)instead of standard convolution to obtain multi-scale fused defect features,and then SENet is used to assign weights to the obtained feature channels to further improve the feature extraction ability of convolution for strip steel defects.Then,the constructed SEPy Conv is introduced into Res Net to replace the 3×3 convolution in the residual network to construct an improved Res Net-50 named SEPy Res Net-50.Finally,the effectiveness of the proposed method is verified on the homemade dataset CR-CLS.The results show that the proposed model SEPy Res Net-50 improves the recognition accuracy by 7.66% compared with the original Res Net50 with the reduction of 0.54 G and 1.06 M in model computation and number of parameters,respectively.On another publicly available dataset X-SDD,the present method can also improve the recognition accuracy by 0.48% to 99.51%.