高新技术产业的不断创新推动了半导体器件朝着多种类和微型化方向发展，厂家通过在半导体产品表面刻印标识字符来检测产品的生产质量，以便更好的管理产品，而传统的字符识别主要依靠人工完成，不仅识别的效率低，而且人工成本高，因此基于机器视觉的字符自动化识别应运而生。为了满足自动化生产条件，本文以车间生产为背景，以芯片表面的刻印字符为对象，研究基于机器视觉的半导体表面字符质量检测系统，该系统能够减少人工参与，提升系统的识别精度和效率，对实现半导体产品字符的自动识别具有重要实用价值。

按照生产需求和待检测芯片确定系统研究内容，主要包括机械平台设计搭建、系统软件设计和图像处理三部分。通过机械设计方案确定硬件设备的选型，并完成半导体表面字符质量检测系统机械结构的搭建。基于Microsoft Visual C++ 6.0编译环境实现系统控制软件，实现对系统的实时控制。系统的重点在于对采集图像的有效处理。在图像处理方面，首先对图像进行平滑，通过分析字符表面噪声特点分别利用三种平滑算法对图像进行处理，选取最优平滑算法。其次在字符分割算法研究方面，首先对基于边缘检测和基于投影的分割算法进行研究分析，通过对比芯片表面字符特征，提出改进的投影分割算法，实验结果表明，该算法能够快速准确地实现字符的准确分割。最后在字符识别算法研究方面，针对传统模板匹配算法进行研究，并基于该算法的不足提出基于两字符距离的匹配算法，取得了较好的识别效果；研究了基础的卷积神经网络算法模型，针对实际问题对网络进行改进，实现目标字符的识别。同时应用残差网络的迁移学习实现字符分类。

在搭建平台上以30*30芯片为标准进行测试，分别应用模板匹配和神经网络两种不同算法对待测字符进行目标识别，识别的准确率分别是87.9%和94.6%，实验结果表明，相较于模板匹配，神经网络算法具有较好的应用效果，满足实际生产要求。

The continuous innovation in the high-tech industry has promoted the development of semiconductor devices in a variety of and miniaturized directions. Manufacturers mark the quality of products by marking logo characters on the surface of semiconductor products to better manage the products, while traditional character recognition mainly relies on Manually completed, not only the recognition efficiency is low, but also the labor cost is high, so the automatic recognition of characters based on machine vision came into being. In order to meet the requirements of automated production, this paper takes workshop production as the background and takes the printed characters on the surface of the chip as the object to study the semiconductor surface character quality detection system based on machine vision. This system can reduce manual participation and improve the recognition accuracy and efficiency of the system. Realizing the automatic recognition of characters of semiconductor products has important practical value.

The system research content is determined according to the production needs and the chip to be tested, which mainly includes three parts: mechanical platform design and construction, system software design and image processing. Determine the type of hardware equipment through the mechanical design plan, and complete the construction of the mechanical structure of the semiconductor surface character quality detection system. The system control software is realized based on the Microsoft Visual C++ 6.0 compilation environment, and the real-time control of the system is realized. The point of the system lies in the effective processing of the collected images. In terms of image processing, the image is first smoothed, and the three smoothing algorithms are used to process the image by analyzing the characteristics of the character surface noise, and the optimal smoothing algorithm is selected. Secondly, in terms of character segmentation algorithm research, firstly, the edge detection-based and projection-based segmentation algorithms are researched and analyzed. By comparing the character characteristics of the chip surface, an improved projection segmentation algorithm is proposed. The experimental results show that the algorithm can quickly and accurately realize the character Accurate segmentation. Finally, in the research of character recognition algorithms, the traditional template matching algorithm is studied, and based on the shortcomings of the algorithm, a matching algorithm based on the distance between two characters is proposed to achieve a better recognition effect; the basic convolutional neural network algorithm model is studied. To improve the network according to practical problems, to achieve the recognition of target characters. At the same time, the use of residual network migration learning to achieve the classification of characters.

The 30*30 chip was used as the standard for testing on the construction platform, and two different algorithms, template matching and neural network, were used to perform target recognition on the characters to be tested. The recognition accuracy rates were 87.9% and 94.6%, respectively. The experimental results show that the comparison For template matching, the neural network algorithm has a good application effect and meets the actual production requirements.