A Hybrid LeNet – LinearSVC Model in Recognizing Handwritten Digits

Abstract:
Convolutional Neural Network (CNN) and Support Vector Machine (SVM) has been the focus of research in image classifications. However, how to combine them together is rarely studied in the past. In this paper, a hybrid model of CNN and SVM was proposed. A modified LeNet that is a typical convolutional neural network with random dropout and regularizations but without the last layer is used as a feature extractor to generate feature vectors that are then fed into the backend LinearSVC for the final classifications. The entire model consists of 3 convolutional layers, 2 max pooling layers and one flatten layer. The model has been tested on the renowned MNIST dataset consisting of 70,000 images. All experiments were tested by using Tensorflow. Without data augmentations and a large epoch number to train the model, the proposed hybrid has achieved a training accuracy of 99.81% and a test accuracy of 99.31%. The experimental results proved that the proposed method is effective in the collected dataset.