An Improved Handwritten Recognition Method Using CNN-SVM Based on Successive Over-Relaxation Iteration

To achieve a more accurate and efficient handwritten recognition, in this paper we propose a model based on an improved successive over-relaxation (SOR) iteration method, building upon the original CNN-SVM hybrid model. We propose using the SOR method for SVM training and proposes the novel algorithm to select relaxation factor using CNN. Through experiments, we put the relaxation factor selected by CNN method into the SOR method to compute the support vector machine (SVM). Subsequently, the SVM under this algorithm is embedded into the CNN-SVM hybrid model. And in this experiment, we use MNIST dataset as an example to evaluate the performance of the new handwritten recognition model, the test accuracy reaches 99.21%, which is 0.14% and 0.35% higher than the original CNN-SVM and LeNet-5 respectively.