基于改进YOLOv5s和DeepLabV3+的摄像头模组瑕疵检测

Defect Detection of Camera Modules Based on Improved YOLOv5s and DeepLabV3+

摘要:

为了克服摄像头模组现有检测方法的局限性，提出了一种基于改进YOLOv5s和DeepLabV3+的摄像头模组瑕疵检测方法，以满足摄像头模组工业生产过程中外观与功能检测的需求；针对摄像头模组表面瑕疵检测中存在难度大、目标小、种类多的问题，采用了基于改进YOLOv5s的外观定性检测方法；该方法引入多维协作注意力机制，并结合基于NWD的损失函数优化，有效提高了模型对小目标的检测能力；测试结果表明，改进后YOLOv5s的平均精度mAP达95.7%，相比原始模型提升了8.5%，同时，每秒帧率FPS为38.5，基本满足工业实时检测的要求；此外，针对需要进行定量检测的组件（如脖子胶区域），进一步研究了一种基于DeepLabV3+语义分割的脖子胶定量分析方法；通过提取区域边界，并计算其面积与长宽比特征，有效评估模组的组装质量并识别潜在功能瑕疵；相比传统方法，该方法能够同时实现摄像头模组的外观与功能检测，同时保障检测的精度与速度，并为其他工业领域的质量控制和瑕疵检测提供了有益借鉴与参考，具有较高的应用价值。

Abstract:

To overcome the limitations of existing inspection methods of camera modules, proposed a camera modules defect detection method based on improved YOLOv5s and DeepLabV3+, designed to meet the requirements for appearance and functionality inspection in industrial production. To tackle the challenges of surface defect detection in camera modules, including high complexity, small target sizes, and diverse defect types, an improved YOLOv5s-based qualitative appearance inspection method is introduced. The method incorporates a Multimodal Co-Attention (MCA) mechanism and an optimized loss function based on Normalized Wasserstein Distance (NWD), significantly enhancing the model's ability to detect small targets. Experimental results show that the improved YOLOv5s achieves a mean Average Precision (mAP) of 95.7%, an 8.5% improvement over the original model, with a frames per Second (FPS) of 38.5, meeting the requirements for real-time industrial inspection. Additionally, for components requiring quantitative analysis, such as the neck glue area, a DeepLabV3+based quantitative analysis method is proposed. By extracting region boundaries and calculating area and aspect ratio features, this method effectively evaluates assembly quality and identifies potential functional defects. Compared to traditional methods, the proposed approach achieves simultaneous inspection of appearance and functionality while ensuring both accuracy and speed. It provides valuable insights for quality control and defect detection in other industrial fields, demonstrating significant application potential.