毕业生知识库

精密曲面光学元件是诸多先进科学装置中的关键零部件之一，其表面质量是影响光学系统整体性能的重要因素，对元件表面进行检测意义重大。传统的人工目视方法无法实现缺陷的精准定量检测，而曲面光学元件非平整的表面特性对自动光学检测设备提出了更高的要求。本文针对曲面光学元件检测装置成像质量要求高、轨迹规划精度要求高、缺陷检测算法性能要求高的技术难题开展研究，主要研究内容和贡献如下：

1）搭建了一套可以对300mm以下大口径曲面光学元件进行2微米（大倍数镜头）至10微米（小倍数镜头）高分辨率缺陷检测的工业视觉装备。具体地，为了使相机可以对复杂曲面进行高精度成像，本文设计了一套由三个平移轴以及两个旋转轴构成的精密五轴运动平台，在实现运动机构多自由度运动灵活性的同时保证运动精度。同时，本文还根据高分辨率可变倍工业相机在不同倍率下的成像特点，设计了基于暗场照明原理的多角度环形照明系统以及基于明场照明的同轴光照明系统。设备可以在15分钟内完成200mm口径曲面光学元件表面的高精度成像与缺陷定量检测任务，检测可重复性达到了95%。

2）提出了一种基于五轴运动机构的曲面光学元件表面成像路径自适应生成方法。因为高分辨率成像系统景深较小，光学元件表面成像位置精度要求很高。本文保持显微相机清晰成像时工作距和视野大小固定，通过几何关系推理得到了一种曲面光学元件表面成像视点生成方法，视点位置计算过程简洁高效。同时，本文还基于五轴机构耦合性低的特点，设计了曲面元件表面成像时五轴机构的运动规划方法。检测装置针对曲面光学元件表面的全口径成像方法摆脱了传统的相机建模、运动学关系构建等复杂过程，简化了路径生成与运动规划方法。同时，本文还设计了一种曲面光学元件快速对心方法，提高了检测流程效率。

3）针对精密光学元件表面缺陷尺寸小、与背景对比度低，在图像中较为微弱的特点，本文设计了一种基于逆注意力机制的图像分割算法提高检测设备对微弱缺陷的检测能力。网络通过特征提取与融合模块以及特征增强模块增加特征融合密度，提高对底层特征的利用率，实现小尺寸缺陷边界的精确判别。同时，网络还加入了逆注意力引导模块，通过逆注意力机制引导网络提升对于与背景对比度较弱的微弱缺陷的识别能力。实验表明本文提出的方法与现有经典分割方法相比能够在参数量较小、推理速度较高的情况下更好地检测精密光学元件表面的微弱缺陷，并具有更高的鲁棒性。

4）设计了一种基于缺陷合成与残差图引导的光学元件图像无监督异常检测方法。针对光学元件表面图像缺陷数据收集困难，同时在大口径曲面光学元件定制化小批量的生产模式下产线快速迁移的需求，本文针对光学元件表面缺陷设计了一种基于随机线段生成的微弱缺陷合成方法，通过缺陷图像训练重构网络，以此提高网络对于微弱缺陷的感知和重构能力。本文还将重构前后图像的残差图输入判别网络，并设计了残差权重引导的损失函数，提升网络对于缺陷的识别能力。实验表明本文设计的方法在光学元件表面异常检测任务中有着良好的效果，性能相比于其他方法有大幅提升。

Precision spherical optical elements are one of the key components in many advanced scientific devices, and their surface quality is an important factor affecting the overall performance of optical systems, it is significant to detect the surface of elements. Traditional artificial visual methods cannot achieve accurate and quantitative detection of defects, while the non-flat surface of spherical elements pose high requirements for the automatic optical inspection equipment, which is one of the urgent issues to be solved in the industry. In this thesis, research is conducted on the technical challenges of curved optical component detection devices, such as requirements on high imaging quality, high trajectory planning accuracy, and high performance for defect detection algorithms. The main contents and contributions are as follows:

1) A set of industrial vision system capable of high-resolution defect detection from 2 micrometers (large magnification lens) to 10 micrometers (small magnification lens) for large aperture spherical optical elements below 300mm has been built. In order to image the complex surfaces with high accuracy, this thesis designs a precision five-axis motion platform composed of three translational axes and two rotational axes, which ensures the motion accuracy while achieving the flexibility of the motion mechanism with multi-degree of freedom. At the same time, according to the imaging characteristics of high-resolution variable magnification industrial cameras at different magnification, this thesis also designed a multi-angle annular lighting system based on the principle of dark field illumination and a coaxial light lighting system based on bright field illumination. The system can complete high-precision imaging and the quantitative detection tasks for defects for spherical optical elements with a diameter of 200mm within 15 minutes, with a detection repeatability of 95%.

2) An adaptive generation method for surface imaging paths of spherical optical elements based on a five-axis motion mechanism is proposed. Due to the shallow depth of field of the high-resolution imaging system, the accuracy of imaging positions on the surface of optical elements is highly required. Based on the fixed field of view of the microscopic imaging camera during clear imaging, this thesis proposes a method for generating viewpoints for surface imaging of spherical elements through geometric relationship reasoning. The process of viewpoint position calculation is simple and efficient. At the same time, based on the low coupling characteristics of the five-axis mechanism, this thesis also designs a motion strategy for the five-axis mechanism during surface imaging process. The detection device eliminates the complex process of traditional camera modeling and kinematic relationship construction for full aperture imaging of spherical elements, simplifying the path generation and motion planning method. This thesis also designs a fast aligning method for curved optical elements, which improves the efficiency of the detection process.

3) Aiming at the characteristics of the weak surface defects on precision optical elements such as small size and low contrast with the background, this thesis designs an image segmentation algorithm based on reverse attention mechanism to improve the detection ability for weak defects. The network uses feature extraction and fusion modules and feature enhancement modules to increase the density of feature fusion, improve the utilization of features from lower layers, which achieves accurate identification of small defect boundaries. A reverse attention guidance module has also been adopted in the network, which guides the network through the reverse attention mechanism to improve its recognition ability for weak defects that have weak contrast with the background. Experiments show that the method proposed in this thesis can detect weak defects on the surface of precision optical elements better with less parameters and higher inference speed, and has greater robustness compared to existing classical segmentation methods.

4) An unsupervised anomaly detection method for optical element images based on defect synthesis and residual image guidance is designed. Aiming at the difficulty of collecting data on surface defects of optical elements, and the need for rapid migration of production lines in the production mode of customized and small batches of large-aperture spherical surface optical components, this thesis designs a weak defect synthesis method based on random line generation for optical element surface defects. The reconstructive network can be trained through defect images to improve the network's ability to perceive and reconstruct weak defects. In this thesis, residual images of pre and post reconstruction images are input to the discriminative network, and a residual weight guided loss function is designed to improve the network's ability to identify defects. Experiments show that the method designed in this thesis has a good performance in optical element surface anomaly detection tasks, and its performance is significantly improved compared to other methods.