基于改进Inception卷积神经网络的手写体数字识别  

HANDWRITTEN NUMERAL RECOGNITION BASED ON IMPROVED INCEPTION CONVOLUTIONAL NEURAL NETWORK

作　　者：余圣新 夏成蹊 唐泽恬 丁召 杨晨 Yu Shengxin;Xia Chengxi;Tang Zetian;Ding Zhao;Yang Chen(Guizhou Key Laboratory of Micro Nano Electronics and Software Technology,College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,Guizhou,China)

机构地区：[1]贵州大学大数据与信息工程学院贵州省微纳电子与软件技术重点实验室

出　　处：《计算机应用与软件》2019年第12期143-149,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61604046);贵州省科技计划项目(黔科合平台人才[2017]5788号);半导体功率器件可靠性教育部工程研究中心开放基金项目(20176103)

摘　　要：手写体数字识别风格变化大,而传统手写体数字识别的准确率又严重依赖于人工特征设计,一旦提取的特征不理想,识别效果就会收到非常大的影响。针对手写体识别正确率无法满足高精度的问题,设计一种高精度的手写体数字分类网络。首先使用连续非对称卷积提取图像的初步特征同时减少计算所需参数,其次使用深度可分离卷积改进Inception结构,并结合残差网络以防止梯度弥散,最后进行softmax分类。通过MNIST数据集实验,得到99.45%的识别率。为进一步提高网络识别率,在分类层使用支持向量机(SVM)代替传统卷积神经网络(CNN)的全连接层与softmax层,经交叉验证得到99.78%的识别率。结果表明,改进Inception结构能够获得更大的网络宽度,同时SVM对于CNN提取的特征的分类能力也有较好效果。The style of handwritten digital recognition changes greatly,while the accuracy of traditional handwritten digital recognition relies heavily on the design of artificial features.Once the extracted features are not ideal,the recognition effect will be greatly affected.Aiming at the problem that the accuracy of handwritten recognition cannot get the high precision,we design the classification network for handwritten numeral with high precision.Continuous asymmetric convolution was adopted to extract the initial features of images and to reduce the computational parameters.Depth separable convolution was used to improve the Inception structure,and the residual network was introduced to prevent the gradient diffusion.We applied softmax to do the classification.Experiment on MNIST data set shows that 99.45%recognition rate is obtained.To further improve the recognition rate of the network,we replace support vector machine(SVM)in the classification layer with traditional convolutional neural network(CNN)in the full connection and softmax layer.The recognition rate is 99.78%by cross-validation.The results show that the improved Inception structure can achieve a larger network width,and SVM has a better effect on the classification ability for features extracted by CNN.

关 键 词：手写体数字识别 卷积神经网络 Inception结构 深度可分离卷积 支持向量机 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]