基于生成对抗网络的手写数字重叠图像分离与识别  

Separation and recognition of overlapping handwritten digit images based on generative adversarial networks

作　　者：韦家成 董然 蔡成涛 林小竹[1] 宋慧佳 王翔宇 WEI Jiacheng;DONG Ran;CAI Chengtao;LIN Xiaozhu;SONG Huijia;WANG Xiangyu(School of Information Engineering,Beijing Institute of Petrochemical Technology,Beijing 102617,China;College of Intelligent Systems Science and Technology,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]北京石油化工学院信息工程学院,北京102617 [2]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2024年第11期2226-2234,共9页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(52171332);北京市教委科技一般项目(KM202210017007);国家级大学生创新创业训练计划(2021J00193).

摘　　要：为解决手写数字重叠给识别带来的障碍,探索识别重叠手写数字的高效方法,本文提出一种采用生成对抗网络分离重叠手写数字的方法,将重叠手写数字分离成2个单独的数字后再进行识别。分别使用卷积层和反卷积层构建判别器和生成器,充分提取数字特征,减少模型参数量;融入自注意力机制,快速有效提取数字显著区域特征;对生成器和判别器进行谱归一,约束网络梯度;改进模型损失函数,提高生成器分离数字的质量。在通过MNIST数据集构造的数据上进行试验,结果表明:本文提出的方法对重叠手写数字的识别准确率达95.91%;峰值信噪比和结构相似性指数分别为22.11和0.8961,相比CapsNet网络模型有了显著提升。To tackle the challenges posed by overlapping handwritten digits in recognition,this paper proposes a high-efficiency method for recognizing such digits using a generative adversarial network-based approach.This method separates overlapping handwritten digits into two distinct digits before recognition.The discriminator and generator are constructed using convolutional and deconvolutional layers,respectively,to effectively extract digit features while minimizing the number of model parameters.Additionally,a self-attention mechanism is incorporated to quickly and effectively extract features from significant regions of the digits.Spectral normalization was applied to the generator and discriminator to constrain the network gradient.Additionally,the model loss function was optimized to enhance the quality of the generated separated digits.Experiments conducted on data constructed from the MNIST dataset show that the proposed method achieved a recognition accuracy of 95.91%for overlapping handwritten digits.Furthermore,the peak signal-to-noise ratio and structural similarity index reached 22.11 and 0.8961,respectively,representing a significant improvement compared with the CapsNet network model.

关 键 词：生成对抗网络 重叠手写数字分离 字符分割 字符识别 重叠目标识别 自注意力机制 深度学习 神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]