Handwritten Digit Recognition Based on SVM

本文主要研究基于支持向量机(SVM)的手写数字识别方法。通过对SVM算法的原理进行深入分析，结合手写数字图像的特征提取方法，构建了一个高效的手写数字识别系统。初始时，固定惩罚参数C (选用0.1)和γ值(选用scale)，采用RBF核和linear核函数进行训练，最后对RBF核函数和linear核函数的各项性能进行比较，如召回率、准确率等，选取针对手写数字识别这一场景下最优的核函数。最后实验结果表明，RBF核函数的各项性能指标如准确率、召回率、F1函数等值在这种场景下要优于linear核函数。最后选取RBF核函数，再进一步调整惩罚参数C和γ值以训练出最优的模型。This paper mainly studies the handwritten numeral recognition method based on Support Vector Machine (SVM). Through the in-depth analysis of the principle of SVM algorithm, combined with the feature extraction method of handwritten digital images, an efficient handwritten digital recognition system is constructed. At the beginning, the penalty parameters C (0.1) and γ (scale) are fixed, and RBF kernel and linear kernel function are used for training. Finally, the performances of RBF kernel function and linear kernel function, such as recall rate and accuracy rate, are compared, and the optimal kernel function for handwritten numeral recognition is selected. Finally, the experimental results show that the performance indexes of RBF kernel function, such as accuracy, recall and F1 function equivalence, are better than linear kernel function in this scenario. Finally, RBF kernel function is selected, and then the penalty parameters C and γ are further adjusted to train the optimal model.