Handwritten digit recognition based on SVM
Wendi Tang, Fei Li · 2024
Handwritten digit recognition (HDR) is one of the most classic application scenarios in machine learning and is often employed to validate algorithms and techniques. The task poses multiple challenges due to variations in writing styles and angles. Traditional feature extraction methods are intricate, and machine learning algorithms have difficulty handling complex nonlinear relationships and rely on features. Deep learning algorithms require a considerable amount of data and computing resources, and there exist problems such as gradient disappearance, explosion, and poor interpretability. In this paper, an algorithm model consisting of an encoder, a decoder, and an image classifier is proposed. Features are extracted through convolutional neural networks, image representation is optimized, and multi-classification support vector machines are utilized for classification. In light of the lack of challenge in MNIST datasets, variational autoencoders are utilized to generate numerically extended datasets in diverse styles. Experiments demonstrate that the proposed method outperforms the classical CNN algorithm by nearly 10% on EMNIST datasets, and the accuracy of the generated datasets is 99.4%. At the same time, the influence of different kernel functions on the performance of the model is also explored.