基于改进VGG-16和朴素贝叶斯的手写数字识别
2020-06-19王梅李东旭
王梅 李东旭



摘 要: 为了解决手写数字识别困难和准确率问题,提出基于改进VGG?16和朴素贝叶斯的手写数字识别,主要通过归一化和双线性插值对图像进行预处理,然后通过改进的VGG?16网络框架对图像进行特征提取和特征融合,通过LDA方法进行数据降维,最后通过朴素贝叶斯分类器进行分类。在MNIST数据集中进行实验,获得了99.36%的准确率。实验结果验证了卷积神经网络与朴素贝叶斯结合后可以有效地提高识别准确率。
关键词: 手写数字识别; VGG?16网络; 朴素贝叶斯分类器; 图像预处理; 特征提取; 数据降维
中图分类号: TN919?34; TP391.41 文献标识码: A 文章编号: 1004?373X(2020)12?0176?06
Abstract: The handwritten digit recognition based on naive Bayes and improved VGG?16 is proposed to improve the difficult recognition and low accuracy in handwritten digital recognition. The images preprocessing are performed by means of the normalization and bilinear interpolation, and then the improved VGG?16 network framework is used to extract and fuse the features of the preprocessed images; the data dimension reduction is carried out by means of the LDA method, and then the data classification is conducted by means of the naive Bayesian classifier. The experiments were conducted in MNIST, yielding an accuracy of 99.36%. The experimental results prove that the convolution neural network combined with naive Bayes can effectively improve the recognition accuracy.
Keywords: Handwritten numeral recognition; VGG?16 network; naive Bayes classifier; images preprocessing; feature extraction; data dimension reduction
0 引 言
数字识别一直是计算机视觉与深度学习领域研究的热点问题。在财务处理、金融管理、车牌识别、邮政编码识别和无人驾驶等领域应用广泛。然而对于手写数字,人工的手写数字笔体不一,导致图像本身的复杂性和不确定性,使识别更加困难。例如,日常生活中邮政编码识别时,微小的错误可能导致巨大损失。
传统的方法都存在特征提取不充分导致准确率不高的问题,如模板匹配法、统计决策法、句法结构法、模糊判别法、逻辑推理法和神经网络法等[1?5]。因此,解决手写数字图像识别的关键在于特征提取,其直接影响最后分类器的识别准确率。部分学者为了对特征进行高效的提取来提高准确率,提出的文献[6]基于旋转投影统计特征的手写数字识别方法通过抽取数字的轮廓和骨架来提取几何特征,文献[7]中基于贝叶斯网络的手写数字识别的研究与应用以及文献[8]中基于主成分分析法及贝叶斯分类器的手写数字识别,虽然在原有的基础上提高了准确率,但是未考虑特征是否提取充分,准确率依旧不足的问题。……
