Research On Metal Sheet Surface Defect Detection System Based On Deep Learning

As one of the common forms of metal materials,metal sheet is widely used in machinery,automobiles,construction and other fields.At present,the demand and quality requirements of various industries are also increasing.Some uncontrollable defects will occur on the surface of the metal sheet during processing.The presence of surface defects not only affects the appearance of the finished product but also has a certain impact on quality.In order to ensure the quality of the metal sheet that flows into the market and also to improve the competitiveness of the enterprise,the detection of the surface quality of the metal sheet has become a necessary link.Although there are many different types of surface defect detection methods and methods,the combination of the actual situation and the combination of the multiple types and uncertain distribution of surface defects of metal plates still mainly rely on manual visual inspection to detect surface defects of metal plates.Due to the low efficiency and poor accuracy of manual visual inspection,it is not enough to meet the inspection needs of enterprises.Therefore,it is of great significance to research and design a set of technical solutions that can accurately detect and identify surface defects of different types of metal plates.Through the related knowledge research of traditional visual inspection,image processing and classic model of deep learning,image classification,etc.,a defect detection system with higher detection accuracy,faster detection speed and lower cost is designed.The main research contents include:Design of hardware module of defect detection system.Introduce and introduce the overall work flow of the inspection system,and design and select the hardware parts such as camera,lens,and light source of the inspection system according to theactual inspection requirements.After completing the selection of hardware design,build and debug the experimental platform.Design of image preprocessing module of defect detection system.This module first performs a certain preprocessing operation on the image through the image preprocessing algorithm.Subsequently,the pre-processed image is tested for suspected defects.If no suspected defects are detected,the plate is considered to be a normal detection plate.If a suspected defect is detected,the suspected defect area is extracted and the extraction result is transmitted to the improved Inception-The V3 model performs accurate detection and identification.Design of accurate detection and identification module for suspected defects.After analyzing and comparing the defect detection method based on R-CNN + SVM,the defect detection method based on GAN + LBP and the detection method based on the Inception series model,the Inception-V3 model in the Inception series model was used to achieve accurate detection of surface defects of metal plates And identification,in order to meet the actual detection needs,the Inception-V3 model is improved accordingly.Production of data sets.Constructed a data set containing one type of normal metal plates and five types of defective metal plates.To solve the problem of insufficient data samples,the original data was expanded by random cropping,translation,and flipping.The expanded data set has a total of 5200 pictures.There are2,000 plates,and there are 3200 metal plates with five types of defects: bottom leakage,dirty spots,scratches,orange peel and bruises.According to the ratio of about 8.5: 1,600 pictures were selected to form a test set to test the model performance.The experimental results show that the overall detection accuracy based on the improved Inception-V3 model reaches 91.15%,which is significantly better than the other two detection methods.Development and design of detection system software.Develop the testing software supporting the testing system and use the sheet metal image to verify the software testing function and actual working conditions.In the actual detection system,the overall detection accuracy has been improved to 92.42% because the image has undergone image preprocessing before inputting the deep learning mode detection type.For the detection time,the detection time of the system for a single non-defective metal sheet is within 1.5s,and the detection of defective metal is within 3.2s to meet the actual use requirements.