毕业生知识库

在精密电子制造领域中，极其微弱的缺陷也会对产品的功能带来损害，需要及时识别并剔除。基于 AI 算法的视觉缺陷检测方法以其非接触、高精度、高效稳定的优点，正在工业质检环节中得到广泛应用。然而，如何在流水线上快速精准地实现被测产品在被检区域的高精度成像，以及如何解决复杂纹理背景对微弱缺陷检测造成的干扰等问题，成为实现微弱缺陷检测亟待研究解决的系列问题。 因此，本文针对智能手机摄像头规则纹理背景下的微弱缺陷检测任务，搭建了一套缺陷检测系统，围绕高精度快速视觉对位方法、工件表面被检测区域定位方法和规则纹理背景下微弱缺陷检测方法三项关键技术展开研究。本文主要的工作和贡献有：

1. 提出了一种基于双目视觉的高精度快速直接对位方法，该方法仅需一 次对位操作，即可将手机屏幕精确地调整至目标位姿，而无需伺服控制。首先，设计了一种基于改进非线性阻尼最小二乘法的远心视觉系统标定方法，建立了双目视觉系统中图像坐标系与局部世界坐标系之间的关系。其次，为了将坐标从局部世界坐标系转换为以运动平台旋转中心为原点的统一世界坐标系，设计了一种基于角度约束的旋转中心标定方法。第三，构建了一种基于形状匹配的两阶段特征点检测方法来检测手机屏幕的特征点。在此基础上，获得了手机屏幕的位姿信息。然后根据工件的当前和目标位置和姿态计算对位指令，在一次对位操作后完成工件精准对齐。最后，针对手机屏幕的对位任务进行了一系列的标定和对位实验。实验表明，对位误差＜0.020mm，对位时间＜20ms，表明了该方法的有效性。
2. 设计了一种基于自监督图像配准的工件表面被检测区域定位方法。针对手机屏幕摄像头规则纹理背景下被检测区域难以有效定位的问题，本文设计了 一种基于自监督范式的图像配准模型，通过直接估计待测图像与基准图像之间的仿射变换参数，将待测图像配准至基准图像位姿，实现了待测图像中不规则被检测区域的有效定位。该方法通过人工生成待测图像与基准图像之间的仿射变换矩阵，通过自监督的方式缓解了图像配准任务中真实仿射变换参数难以获取的问题。实验结果表明，该方法可以有效引导模型学习图像配准特征，实现精确的规则纹理背景图像配准，完成了工件表面被检测区域定位任务。
3. 给出了一种规则纹理背景下微弱缺陷分割方法。针对具有规则纹理背景的手机屏幕摄像头，本文实现了一种基于多尺度有偏融合的缺陷分割模型，通过执行更加偏向图像低层特征的特征融合策略，有效分割出摄像头的微裂纹和彩虹纹两种微弱缺陷。此外，为了可以获取更加紧凑的图像块来提升算法效率，本文实现了一种实用高效的基于质心引导的图像块切分策略。最后，本文通过知识蒸馏策略进行了模型压缩工作，获得轻量化缺陷分割模型，以适应工业现场对视觉检测模型的轻量化需求。实验结果表明了该方法可以精确实时检测出手机屏幕摄像头的微裂纹和彩虹纹两种微弱缺陷。

In the field of precision electronics manufacturing, even extremely weak defects can damage the function of products and therefore need to be detected in a timely manner. The AI algorithm-based visual defect detection method has been widely used in industrial quality inspection with the advantage of non-contact, high accuracy, efficiency and stability. However, achieving high precision imaging of the inspected area imaging of the product on the assembly line quickly and accurately, as well as mitigating the interference problem caused by complex texture backgrounds for faint defect detection are the series of problems that need to be solved urgently to realize faint defect detection. Therefore, in this paper, a mobile phone screen camera defect detection system is established for the task of detecting faint defects with regular texture background, focusing on three key technologies: a high-precision fast alignment method based on binocular vision, a method for extracting the inspected area on the workpiece surface and a method for detecting faint defects with regular texture background. The main work and contributions of this paper are as follows.

(1) A high-precision fast alignment method based on binocular vision is proposed, which can accurately adjust the mobile phone screen to the target position and pose after one-shot alignment operation without servo control. Firstly, a calibration method of the telecentric vision system based on an improved nonlinear damped least-squares method is proposed to establish the relationship between the image coordinate system and the local world coordinate system in the binocular vision system. Secondly, in order to transform the coordinates from the local world coordinate system to a unified coordinate system with the platform’s rotation center as the origin, an angle constraint-based rotation center calibration method is proposed. Thirdly, a two-stage feature point detection method based on shape matching is proposed to detect the feature points of the workpiece. Based on these, the position and pose of the workpiece are obtained. Then the alignment commands are calculated based on the current and the target position and pose of the workpiece, enabling the accurate alignment to be accomplished in one operation. Finally, a series of calibration and alignment experiments were conducted. the experiments and results show that the alignment errors are within ±0.020mm and the time taken to calculate alignment commands is less than 20ms, which demonstrates the effectiveness of the method.

(2) To solve the problem that the detected area is difficult to be extracted effectively under the regular texture background of mobile phone screen camera, a self-supervised image registration method is proposed to align the image to be detected to the reference image position and pose by directly estimating their affine transformation parameters. The method simulates the corresponding affine transformation of the object in the field of view by randomly generating the affine transformation, which alleviates the difficult problem of obtaining real data labels in the image alignment field. Experimental results show that the method performs high accuracy, fast inference speed on the image alignment task with regular texture backgrounds, and good adaptability to different data sets.

(3) Aiming at the difficulty of weakness of defects in regular texture background of mobile phone screen camera with regular texture background, a defect segmentation model is implemented based on multi-scale biased feature fusion, which effectively segments weak defects in images by implementing a feature fusion strategy that is more biased to visual low-level features containing precise spatial location information of defects. In addition, in order to speed up defect detection in industrial sites, a practical and efficient center-of-mass-guided image patch slicing strategy is proposed in this paper for obtaining more compact image blocks. Finally, model compression experiments are carried out through a knowledge distillation strategy to obtain a lightweight defect segmentation model to adapt to the requirements of industrial sites. The experimental results demonstrate that the method can accurately detect two kinds of weak defects in mobile phone screen camera in real time, which are microcracks and rainbow lines.