基于轻量级网络的PCB芯片文字识别  

PCB Chip Text Recognition Based on Lightweight Network

作　　者：蒋子敏 刘宁钟[1] 沈家全 JIANG Zi-min;LIU Ning-zhong;SHEN Jia-quan(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106

出　　处：《计算机技术与发展》2021年第12期55-60,共6页Computer Technology and Development

基　　金：国家自然科学基金(61375021)。

摘　　要：随着人工智能的兴起,深度学习技术在多个领域有了广泛的应用与发展,将深度学习技术与PCB芯片文字识别相结合,实现具体的场景应用,具有重要意义。基于深度学习进行文字识别,完成图像中文字信息的自动获取,进一步提高了准确率,极大地节约了人工及时间成本。但传统深度学习模型巨大的参数量以及内存消耗等限制了其在小型设备如移动终端上的应用与发展,难以满足人们日益增长的需求。基于此,该文提出了一种基于轻量级网络的文字识别算法LWTR。该算法框架主要包括卷积神经网络(进行特征提取)、循环神经网络(进行标签预测)以及CTC(实现转录),最终得到预测序列。为减小模型参数量,进行通道数的统一并采用多路小卷积及堆叠Dense Layer充分提取特征。同时,为加速网络的收敛,提高模型的泛化能力,引入BN归一化。结果表明,该算法在PCB芯片数据集中文字识别准确率达到了89.58%,与现有文字识别算法相比,在准确率几乎没有下降的情况下具有更小的模型与更快的速度。With the rise of artificial intelligence,deep learning technology has been widely used and developed in many fields.It is of great significance to combine deep learning technology with PCB chip text recognition and realize specific domain applications.Finishing text recognition based on deep learning to realize automatic acquisition of text information in images has further improved recognition accuracy and greatly saved labor and time costs.However,the huge parameter amount and memory consumption of the traditional deep learning models limit its application and development on small devices such as mobile terminals,and it is difficult to meet the increasing needs of people nowadays.Therefore,a text recognition algorithm LWTR is proposed based on lightweight network.The algorithm framework mainly includes convolutional neural network for feature extraction,recurrent neural network for label prediction and CTC to achieve transcription,and finally gets the predicted sequence.In order to reduce the amount of model parameters,unified number of channels,multi-path small convolution and stacked Dense Layer are used to fully extract features.At the same time,in order to accelerate the convergence of the network and improve the generalization ability of the model,Batch Normalization(BN)is introduced.The results show that the proposed algorithm has achieved 89.58%of the text recognition accuracy in our PCB chip dataset.And compared with the existing text recognition algorithms,it has a smaller model and faster speed with almost no decrease in accuracy.

关 键 词：PCB芯片 文字识别 轻量级神经网络 卷积神经网络 循环神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]