MNIST Dataset Heatmap Classifier

In this paper, a simple heatmap-based classifier is proposed that uses average pixel values from training data to assess image similarity to individual classes. The classifier was tested on the MNIST dataset, achieving an accuracy of 66.85% in the basic version and 81.31% in the version with region of interest (ROI) extraction and interpolation. The training process has a time complexity of O(n), and the evaluation of the classifier has a time complexity of O(1). The presented method is simple to implement and fast to train. The paper discusses potential directions for further research, including data augmentation and integration with deep learning models, such as convolutional networks, or other machine learning algorithms, e.g. k-nearest neighbors, which can significantly improve the effectiveness of the classifier.