混合样本训练方式的手写数字识别  

Handwritten digital recognition of mixed sample training

作　　者：汪雅琴 夏春蕾[1] 戴曙光[1] Wang Yaqin;Xia Chunlei;Dai Shuguang(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院

出　　处：《电子测量技术》2018年第18期52-56,共5页Electronic Measurement Technology

摘　　要：近年来,图像识别技术被应用在通信、军事、公安侦测、生物医学等众多领域,而LeNet-5模型作为卷积神经网络模型的典型代表受到广泛青睐。在MNIST字符库上运用该模型,通过优化卷积层的样本训练方式,即将原来以每批固定输入样本数量、固定迭代次数的训练方式,优化为以每批不同输入样本数量、不同迭代次数的混合训练样本方式。优化后的训练方式能够减少预处理工作量,加快识别速度。通过实验可知,在相同的训练时间内,分别采用传统样本训练方式和混合样本训练方式,识别率可以提高0.15%左右;改变混合样本的组合方式,识别率也会不断改变,甚至相差0.18%。实验结果表明在保证样本训练时间相等的前提下,优化后的混合样本输入方式可以得到更低的测试错误率,即更高的识别率。In recent years, image recognition technology has been applied in many fields such as communications, military, public security detection, and biomedicine, and the LeNet-5 model has been widely favored as a typical representative of the convolutional neural network model. Use this model on the MNIST character library and optimize the sample training method of convolution layer. That is to say, the training method that uses the number of fixed input samples per batch and the number of fixed iterations is optimized to be a mixed training sample mode with different numbers of input samples per batch and different iterations. The optimized training method can reduce preprocessing workload and speed up recognition. Through experiments, we can see that in the same training time, using the traditional sample training method and the mixed sample training method respectively, the recognition rate can be increased by about 0.15%; changing the combination of mixed samples, the recognition rate will continue to change, even a difference of 0.18%.The experimental results show that the optimized mixed sample input method can get lower test error rate, ie higher recognition rate, under the premise of equal sample training time.

关 键 词：图像识别技术 LeNet-5模型 卷积神经网络 MNIST字符库 

分 类 号：TN37[电子电信—物理电子学]