Surface Defect Detection of Mobile Phone Covers Based on Improved BiSeNet V2
Bo Liu,Tingting Wang,Jie Liu
DOI: https://doi.org/10.3788/aos240659
2024-01-01
Acta Optica Sinica
Abstract:Objective Mobile phone glass covers are crucial components of display screens, influencing optical display performance and user experience. Mass production necessitates rigorous defect detection due to inevitable defects like scratches and dirt. Current methods like manual inspection and traditional machine vision fall short of meeting high- volume, high accuracy demands. Addressing these challenges, we propose an improved BiSeNet V2- based semantic segmentation method for accurate and efficient surface defect detection of mobile phone covers with diverse defect types, complex shapes, and challenging background distinctions. Methods We introduce an improved BiSeNet V2 semantic segmentation method tailored for detecting defects on mobile phone covers. Initially, we enhance defect features using a fusion weighted image difference method to aid network model extraction, addressing unclear detect imaging and low background discrimination issues. Building upon BiSeNet V2, enhancements include a detailed branch network using grouped dilated convolution and residual structures to enhance multi- scale defect feature extraction and preserve shallow feature details. A channel attention mechanism adaptively calibrates feature channel importance, bolstering defect recognition. A multi- scale feature fusion method in the decoding network restores lost local detail, enhancing defect segmentation accuracy and precision. Results and Discussions In response to various surface defects in mobile phone cover plates, such as multiple types of defects, small affected areas, a wide range of size variations, unclear defect features, and challenges in distinguishing them from the background. We adopt a lightweight semantic segmentation network based on BiSeNet V2 for defect detection. The network model is enhanced and optimized from four key aspects: improving defect feature representation, incorporating group dilation convolution, integrating attention mechanisms, and fusing multi- scale features. These enhancements effectively boost the accuracy of segmenting small and slender defect targets, enabling the classification and detection of multiple defect types. The refined model strikes a good balance between detection accuracy and detection speed. Eight samples of defect images from mobile phone cover plates are selected for detailed analysis and comparison (Fig. 14). Results demonstrate that the improved network model excels in extracting defect features and recovering shallow details, achieving notably comprehensive segmentation results for slender scratch defects, surface dust, and foreign objects. Significant improvements are observed in detecting small point- shaped targets and in distinguishing similar defects, enabling overall detection efficacy across different defect types on mobile phone cover surfaces. In practical production processes, defect detection on mobile phone cover plates prioritizes metrics such as defect sample detection, missed detections, and false detections of normal samples. We quantitatively analyze the defect detection performance on mobile phone cover plates. Utilizing the improved BiSeNet V2 network model, 354 defect samples and 100 normal samples are examined, and the detection outcomes for various defect types in image samples are statistically analyzed. Experimental results (Table 5) indicate that the proposed method achieves a defect detection accuracy of 91.27 degrees o, a misclassification rate of 5.02 degrees o, a defect detection rate of 96.29 degrees o, a leakage detection rate is 3. 71 degrees o, and a normal sample misdetection rate of 1.00 degrees o. Thus, the improved network model significantly enhances defect detection capabilities across diverse surface defects of mobile phone covers. Conclusions We propose an improved BiSeNet V2 semantic segmentation network for detecting surface defects on mobile phone cover plates. Initially, defect image features are enhanced through weighted image difference processing, enhancing contrast between defects and background to facilitate comprehensive defect feature extraction. Subsequently, employing group dilation convolution and squeeze- and- excitation (SE) attention mechanism enhances bilateral feature extraction, accommodating defect features at diverse scales while minimizing loss of feature detail and channel- wise feature map calibration for adaptive response. Lastly, a multi- scale feature fusion method in the upsampling decoding network restores lost detail information, enhancing segmentation accuracy across various defect types and reducing missed detection rates. Experimental results demonstrate superior detection performance of the improved BiSeNet V2 network model compared to alternative semantic segmentation networks for diverse defect types on mobile phone cover plates.