Research Of Improved U-Net Based Detection Method For Camera Module Assembly Adhesive Defects

Automatic dispensing machines have a high efficiency and yield rate when dispensing camera modules,but they are not completely immune to glue defects,which can lead to poor adhesion on the work surface and seriously affect the stability of the product.In recent years,machine vision and deep learning have made extremely rapid research progress in the field of non-destructive testing.Applying machine vision and deep learning to the detection of defects in the camera module assembly adhesive is of great significance to improve the quality of camera module products.For the problem that it is difficult to obtain a large number of glue samples for labeling in industrial production scenarios,this paper uses data augmentation methods,such as scale and color space transformations,to expand the data and at the same time improves the random erasure algorithm.In order to further enhance the supervision information of the samples,the mask is integrated into the samples as the generation condition,and Res-UNet is used as the generation model to build a Region-Maskgenerative adversarial network.The network model can expand the data of colloidal defects and glue,and significantly enhance the diversity of colloidal defects and glue samples.Aiming at the problem that the U-Net partition network cannot meet the real-time requirements in industrial applications.In this paper,a lightweight U-Net containing12 convolutional blocks is designed.At the same time,the convolutional attention module and context-guided module have made lightweight improvements,and the improved modules are embedded in the lightweight U-Net.Ablation experiments demonstrate the effectiveness of module improvement and module embedding.The model achieved 81.21% MIo U and 94.96% MPA for the four categories of background,glue,scratches and bubbles.In the segmentation task of the colloid defect and glue,compared with U-Net models of different backbones,the model in this paper has a significant speed advantage with almost the same guaranteed accuracy.Compared to the lightweight real-time segmentation networks ENet and CGNet,the model in this paper has advantages in terms of speed and accuracy.Finally,the leakage and false detection rates of industrial inspection metrics are introduced,and use a variety of methods to detect dispensing defects and colloidal defects.The deployment of the network model based on the existing hardware system.The experimental results show that the segmentation and classification network model designed in this dissertation can meet the real-time and accuracy requirements in industrial inspection.