Research On Image Detection And Classification Recognition Technology Of Steel Plate Surface Defects

Steel plates are one of source materials in automobile production,shipbuilding,light industry,engineering and other areas.However,during the casting procedure,because of the influence of high temperature,pressure,speed and hardware equipment,various defects on the steel surface are inevitable.These defects run through the entire production procedure,which not only affect the surface appearance,processing effect and processing cost of subsequent formed devices,but also affect their corrosion resistance,wear resistance and fatigue strength,and serious cases may lead to production accidents.Therefore,detecting and controlling the surface quality of steel plates has become the main task in steel processing enterprises,and is also an issue concerned by domestic and foreign scholars.The detection method based on machine vision has no impact on the steel plate surface and inspection equipment.Therefore,this dissertation introduces image processing and computer machine vision technology,takes the defect images as the research object,constructs the defect intelligent detection system and designs the relevant intelligent recognition algorithm.It can complete image acquisition,processing,defect recognition and detection.Specific contents include:(1)In view of the complex production field environment,it is difficult to obtain images with stable quality.An improved MSR(Multiple-Scale Retinex)steel plate defect image enhancement algorithm is proposed.On the basis of the traditional MSR algorithm,when multiple scale images are combined,the weight of each scale is automatically determined by calculating the ratio of information entropy of each image component,and the three channel images are combined to obtain the enhanced image.Compared with other common preprocessing methods,the algorithm proposed in this dissertation has higher target definition and contrast,and less background interference,which can significantly increase the image quality and improve the adaptability of the machine vision system to different lighting conditions.(2)The method of scratch defect recognition and feature parameter extraction has been studied.First,the defect edge features in an image are enhanced by Log Gabor wavelet transform and phase consistency detection.And the two-dimensional Otsu method is used for segmentation.Then the defect area can be extracted by combining morphological filtering and connected domain,and the defect location can be identified.Establishing the scratch defect model enables the calculation and preservation of geometric feature parameters,such as width,length and other parameters.For the experimental image,the maximum error is 13.62% and the minimum is 2.28%,and the average running time is0.2826 s.The maximum error and minimum error of measurement for field image detection are 15.43% and 4.15% respectively,and the average operation time is 0.3193 s.Compared with other methods,this method has smaller error and shorter running time.(3)A method of steel plate defect image classification based on combined feature selection is studied.The GLCM is used to calculate the texture characteristics,such as energy,contrast,correlation and others.Then the shape features,like area proportion,aspect ratio and roundness of the target are extracted through maximum entropy threshold segmentation.The combination features of texture and shape with strong classification ability are selected as the identification parameters.The FOA is used to optimize the parameters of PNN network structure.Compared with other methods,the FOA-PNN algorithm based on texture and shape combination features has an accuracy of 94.5% and a test time of 23.7ms.What’s more,the open dataset NEU-CLS is subjected to the proposed algorithm,yielding an accuracy of 93.8% and a test time of 25.3ms.And the recognition accuracy of the test field image is 91.7%,which generally meets the detection requirements.(4)A lightweight steel plate surface defect detection method is proposed.Firstly,the steel plate defect image collected in this dissertation are expanded to increase samples.Then the proposed defect region extraction of steel plate by using image segmentation and pixel difference is used to get the target bounding box,which insteads of manually drawing box to assist the annotation of defect datasets.The average intersection ratio IOU and detection speed of the proposed method are 0.87 and 457 ms,respectively.Finally,a lightweight target detection model is designed.The YOLOV5 model is adopted as the basic network frame,the Mobilenet V2 is substituted for its backbone part and CBAM is added to optimize features.NEU-DET and enhanced labeled images in this dissertation are used to construct the defect dataset to test.The model in this dissertation has a test accuracy of 0.924 and a reasoning speed of 29.8ms.Although the accuracy of the improved method is lower than that of YOLOV5,the amount of parameters,calculation and reasoning time are far less than that,and the comprehensive performance is improved.Compared with SSD and Faster R-CNN models,the performance is better than them,and it also meets the real-time requirements of industrial detection with high accuracy.Finally,based on the above research methods,these intelligent algorithms are embedded in the expert software of the background monitoring center for debugging and verification.Users can remotely control the front-end camera through the PC terminal,and view the field video and save test data in real time,so as to conduct quantitative and qualitative analysis and evaluation on the quality of the entire steel plate.It provides theoretical basis and research ideas for intelligent detection of steel plate surface defects.