Research On Defect Detection Method Of Small Sample Glass Based On YOLOv4

The curved glass cover is a transparent component,which is mainly used in the appearance and structure of smart terminal products.With the development of intelligent terminal products,the demand for glass cover plates has risen rapidly.However,the manufacturing process of curved cover glass is complicated,and defects such as internal bubbles,surface scratches,fingerprints,etc.are inevitably generated during the actual production and processing.In order to guarantee the quality of the product,the glass cover plate must be inspected for defects.Manual or traditional machine vision detection methods are difficult to meet the needs of current product production efficiency and quality.The use of intelligent and unmanned detection methods is a difficult problem that needs to be solved urgently in the current industry.Deep learning artificial intelligence algorithms have been extensively researched in the field of image recognition and defect detection.However,the high light transmittance and high curvature arc surface of the curved cover glass are easy to cause halo and image interference in its imaging,which greatly affects the detection of glass surface defects.In addition,the glass surface defect data obtained during the manufacturing process has problems such as small sample size,random defect characteristics,and concealment,which affect the accuracy and robustness of the deep learning algorithm model.To this end,this thesis will carry out related research on the intelligent detection of glass surface defect detection.The research contents are as follows:（1）Aiming at the problem of difficult imaging of surface defects on curved glass and large interference,by analyzing the optical imaging principle of glass and combining with the imaging characteristics of glass surface defects,the use of brightfield backlight diffuse reflection can more highlight the difference between the edge of the glass surface defect and the background.contrast.According to the above method,the glass cover plate can obtain the defect information on the flat surface and the curved surface respectively under the illumination of the annular smooth and the strip light source in two postures of laying flat and standing upright.（2）Aiming at the problem of small samples of glass surface defects,this thesis firstly enhances the glass surface defect samples offline by flipping,affine,cutting and other methods.Secondly,the model weights are initialized by means of transfer learning,and the data is enhanced online with the Mosaic data enhancement method to help the model learn more fully the feature information of glass surface defects.The detection results show that the trained YOLOv4 target detection model can effectively detect glass surface defects,and the m AP value of glass surface defect detection reaches 84.18%.（3）For small target defects,the improved k-means clustering algorithm is used to reset the model anchor value to solve the problem that the size of the original anchor frame is not suitable for small target detection of glass defects.And replace the ASPP network at the end of the feature extraction network and add the attention mechanism CBAM module to the feature fusion network,so that the detection is more focused on the center of the defect.The experimental results show that the m AP detected by the optimized YOLOv4 model is increased by 4.9%,which has better robustness and detection effect.（4）According to the application requirements of the defect detection system,a glass surface defect detection software based on the improved YOLOv4 model is designed to achieve real-time acquisition of glass image data and visualization of detection results.The detection performance of the system is evaluated and analyzed through the experimental detection results,which provides a feasible solution for the detection of glass surface defects.