基于特征提取和模板匹配的电网工程图纸字符识别技术  

Character recognition technology of power engineering drawings based on feature extraction and template matching

作　　者：侯凯 梅诗妍 HOU Kai;MEI Shiyan(School of Electrical Engineering and Automation,Xi′an Jiaotong University,Xi′an 710049,Shaanxi,China;Power Grid Planning Research Center,Guangdong Power Grid Co.,Ltd.,Guangzhou 510080,Guangdong,China)

机构地区：[1]西安交通大学电气工程与自动化学院,陕西西安710049 [2]广东电网有限责任公司电网规划研究中心,广东广州510080

出　　处：《沈阳工业大学学报》2025年第2期176-182,共7页Journal of Shenyang University of Technology

基　　金：广东省科技计划项目(GZHKJXM20210046)。

摘　　要：【目的】电力工程图纸在建设中是生产计划、施工及验收等环节的重要依据。然而,传统人工识别方式存在效率低、错误率高、成本高等问题,难以满足现代复杂工程项目的需求。近年来,计算机视觉技术在自动识别领域取得显著进展,但现有算法在电力工程图纸识别中仍面临识别效率低以及对倾斜和变形字符识别准确率低的难题。【方法】提出了一种基于VGG网络和Hu不变矩的电力工程图纸字符识别算法,旨在通过结合尺度自适应深度卷积特征与Hu不变矩特征,提高电力工程图纸的识别效率和准确率。利用VGG网络提取深度卷积特征,并通过自适应方式选择目标层,以实现模板与图像的尺度自适应特征提取。该方法避免了传统滑动窗技术多次提取特征的问题,仅需对每个模板和图像进行一次特征提取,大幅提升了处理效率。为解决字符倾斜和变形的难题,结合了Hu不变矩特征,利用其平移及旋转不变性作为补充特征,有效增强了对复杂字符形态识别的鲁棒性。【结果】通过对比现有算法,从识别效率和准确率两方面验证了算法的性能优势。实验结果表明,算法在识别效率和准确率上均表现出显著优势:与传统CNN字符识别算法相比,算法的执行时间约为其1/4,显著提高了处理速度;通过结合Hu不变矩特征,算法在识别倾斜和变形字符方面表现出较强的鲁棒性;采用自适应目标层选择策略后,特征提取的准确性和鲁棒性可以得到进一步提高,优于固定网络层的特征提取方式。在复杂场景下算法具有更强的适应能力,具有良好的应用前景。【结论】研究的创新之处在于:提出的尺度自适应深度卷积特征提取方法在电力工程图纸识别中可以进行单次特征提取,大幅提升识别效率;结合Hu不变矩特征的设计增强了对复杂字符形态的识别能力,特别是增强了对倾斜和变形字符的鲁棒性[Objective]Power engineering drawings are essential for production planning,construction,and check and acceptance stages in engineering projects.However,traditional manual recognition methods suffer from low efficiency,high error rates,and high costs,which makes them unsuitable for modern complex engineering projects.In recent years,significant progress has been made on computer vision technology in the field of automatic recognition.Nevertheless,existing algorithms still face challenges such as low recognition efficiency and poor accuracy in identifying slanted and deformed characters in power engineering drawings.[Methods]To address these shortcomings,a character recognition algorithm based on the visual geometry group(VGG)network and Hu invariant moment was proposed,which aimed to improve the recognition efficiency and accuracy for power engineering drawings by combining scale-adaptive deep convolutional features with Hu invariant moment features.Firstly,deep convolutional features were extracted using the VGG network,and the output layer was adaptively selected to achieve scale-adaptive feature extraction for templates and images.This approach avoided the multiple times of feature extraction in traditional sliding window techniques,only extracting features once for each template and image and thereby significantly improving processing efficiency.Secondly,to address the challenges of slanted and deformed characters,Hu invariant moment features were integrated as supplementary features.Their translation and rotation invariance were leveraged to enhance robustness against complex character shapes.[Results]The performance superiority of the proposed algorithm was validated in terms of efficiency and accuracy by comparing it with existing algorithms.The results demonstrate that the proposed algorithm offers significant advantages.Its execution time is approximately one-fourth that of the traditional convolutional neural networks(CNNs)-based character recognition algorithm,namely that the proposed algorithm has gre

关 键 词：电力工程图纸 字符识别 特征提取 模板匹配 VGG网络 深度卷积特征 归一化互相关系数 HU不变矩 

分 类 号：TM763[电气工程—电力系统及自动化]