中国制造2025背景下电子工业技术快速发展，其产品自动化生产和封装技术都相对比较成熟。但由于生产过程中的诸多因素，导致产品外观时常出现划痕、斑痕、压痕等外观缺陷，严重影响产品质量。而目前外观缺陷检测仍以人工检测为主，其检测效率低，准确度不稳定，且人力成本高，这些都严重影响产品产量。基于计算机视觉和人工智能技术的外观缺陷检测技术也逐渐发展起来。但传统图像处理方法对图像特征的表征相对比较低级，而且容易受到光照、噪声等干扰，导致缺陷检测效果和性能仍不理想。为此，受到深度学习技术启发，本文利用图像处理、神经网络语义分割等方法对电子元器件中的薄膜电容外观缺陷检测展开研究。
为了避免其他目标对检测结果的干扰，同时提升后续处理速度，研究了薄膜电容外观快速分割方法。在研究了基于阈值、区域和图论的分割算法的特点基础上，提出了基于HSV空间OTSU算法的快速分割算法。进而对分割结果进行形态学处理，去除毛刺和孔洞，得到薄膜电容外观的完整区域图。
由于传统图像特征提取方法对内容的表征比较低级，为此，提出了融合Gabor卷积网络进行外观缺陷预处理，以增强特征。进而，介绍了经典的深度语义分割网络FCN，UNet和SegNet，在考虑像素级分割效果和算法效率下，采用了基于DeepLabV3+的外观缺陷快速检测模型。
但由于光照干扰，薄膜电容器件边界上往往出现部分黑色阴影或白色条纹，极易被误检为缺陷，导致良品率低。为此，将器件表面轮廓图作为引导信息，以抑制边界附近光照的干扰。对比了Canny，HED和RCF边缘轮廓提取方法，结果表明RCF方法提取的边缘轮廓图对器件外观图像噪声抑制效果更好，更加适合用于引导图。
基于上述改进，提出了融合Gabor卷积网络和轮廓信息引导DeeplabV3+的薄膜电容外观缺陷检测方法，并在实测数据集上验证了提出方法的检测结果视觉效果、客观评价指标mIoU和分类准确度。最后设计实现了基于PyQt和Pytorch的算法部署和可视化界面，并对检测效果和性能进行展示。本文在薄膜电容外观缺陷检测方向的研究成果具有较好应用价值。
Under the context of "Made in China 2025", the technology of electronic industry is developing rapidly, and its product automatic production and packaging technology are comparatively mature. However, owing to many factors in the production process, the product appearance often has surface defects such as scratches, spots and indentation, which seriously affects the product quality. At present, the surface defect detection is still dominated by artificial detection, which has low detection efficiency, unstable accuracy and high labor cost, which seriously affect the product output. Surface defect detection technology based on computer vision and artificial intelligence technology is also gradually progressed. However, the image feature representation by traditional image processing methods is relatively low, and it is easy to be disturbed by illumination and noise, resulting in the unsatisfactory effect and performance of defect detection. Therefore, inspired by deep learning technology, this topic uses image processing, neural network semantic segmentation and other methods to study the surface defect detection of capacitor devices in electronic components.
For fear of avoiding the disturbance of background object on the detection results and improve the subsequent processing speed, a fast surface segmentation method of capacitor devices is studied. Based on the study of the characteristics of segmentation algorithms based on threshold, region and graph theory, a fast segmentation algorithm based on Otsu algorithm in HSV space is proposed. Then the segmentation results are morphologically processed to remove burrs and holes, and the complete area map of the capacitor surface is obtained.
Because the traditional image feature extraction method has low representation of content, a surface defect preprocessing method based on three-channel Gabor convolution network is proposed to enhance the feature. Then, the classical deep semantic segmentation networks FCN, UNet and SegNet are introduced. Considering the pixel-level segmentation and efficiency, a fast surface defect detection model based on deeplabv3+ is adopted.
However, due to light interference, some black shadows or white stripes often appear on the boundary of capacitor devices, which is very easy to be misdetected as defects, resulting in low yield. Therefore, the device surface profile is used as guidance information to control the disturbance of light near the boundary. The edge contour extraction methods of Canny, HED and RCF are compared. The results show that the edge contour extracted by RCF method has better noise suppression on the device surface image, and is more suitable for guidance map.
Based on the above improvements, a surface defect detection method of film capacitance integrating Gabor convolution network and contour information as guidance information is proposed, and the visual effect, MIoU and classification accuracy of the proposed technique are confirmed on the dataset. Finally, the algorithm deployment and visualization interface based on PyQt and Pytoch are designed and implemented, and the detection effect and performance are displayed. It has good application value in the research of surface defect detection of capacitor devices.
