Research On Defect Detection System Of Injection Model Based On 3D Laser Scanning

Plastic products occupy a very important position in modern industry,penetrate almost every industry field,and are also one of the commonly used materials for notebook panels.In the process of injection molding of plastic products,affected by many factors such as mold quality,raw material quality,processing technology level,etc.,there is a certain probability of defective products.To ensure the quality of the product,it is necessary to carry out surface defect detection on the notebook panel.Vision-based conventional defect detection can well detect some defects with texture outline characteristics,but sometimes there are no obvious outline characteristics due to the special texture and defects on the surface of the notebook panel,such as defects such as bulging and pitting,it is difficult to collect significant defect features using 2D imaging,and the missed detection rate is high.Therefore,this thesis designs and builds a platform to collect 3D data based on the principle of laser triangulation,and realizes automatic defect detection with the help of deep learning.The main research contents are as follows:(1)Aiming at the difficulty of collecting defect features for bulging and pitting defects on the surface of notebook panels,an RGB-D camera scheme was designed to characterize the defect features of bulging and pitting through depth images,while satisfying the collection of defect features in color space.A method of dislocation value is proposed,and a color camera is used to collect RGB images and depth images simultaneously to reduce the number of cameras on the acquisition platform.A data acquisition platform was built and calibrated combined with the 3D camera system and the motion control system.The X-axis precision of the data acquisition platform was 0.02 mm/pixel,the Z-axis precision was 0.03mm/pixel,and the acquisition speed was 20 mm/s.(2)According to the requirements of the speed of the laser centerline extraction algorithm,the extreme value method was optimized,and the template extreme value method was proposed.filter operation.At the same time,the search strategy is improved to keep the template near the laser line for searching,avoiding the global search of the laser image,and improving the speed and stability of the laser centerline extraction.On 3072*320 images,the extraction speed of a single image is stable at 8 ms.In extreme conditions,where the laser line is on the lowest side of the laser image,the extraction speed can be increased by a factor of 10.(3)The original laser image collected by the platform is processed,and the interference of other light sources other than the laser light source is filtered out by the method of channel difference,and a clear laser line image is obtained.Depth maps and RGB images are reconstructed using triangulation and misalignment.The gradient image is calculated on the depth image to make the defect features more prominent.Combine RGB images,depth images,and gradient images into a defect dataset,and use artificial synthesis and GAN adversarial network generation methods for data augmentation for subsequent deep learning model training and testing.(4)To further improve the target detection accuracy of the Mask RCNN deep learning model,the respective advantages of the convolutional neural network and the Transformer network are analyzed,and the feature extraction network model of Res-Swin-T is proposed by combining the two.Under the Mask RCNN framework,Validated using the public dataset DAGM2007.The experimental results show that adding a convolutional layer to the Transformer network can improve the local feature extraction ability of the network model.The Res-Swin-T feature extraction network is 6.5% higher than the Res Net50 with the same level of parameters,and the average accuracy m AP50 is higher than that of the Swin-Tiny network 0.8% to 99.4%.(5)Using the optimized Mask RCNN model,the collected and produced notebook panel surface defect data set is used for defect detection,and the training process is optimized by combining CIo U loss and Soft-NMS.The final detection accuracy m AP reaches 88.3%,which basically meets the accuracy of industrial inspection Require.Finally,a software detection system is designed to realize the defect detection of the notebook panel.