A Classical Machine Vision Pipeline for MNIST Handwritten Digit Recognition: LBP and HOG Features with From-Scratch J2 Discriminant Compression

Abstract

This paper presents the full pipeline for MNIST handwritten digit recognition using Machine Learning. The pipeline consists of three stages: (1) feature extraction using Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) descriptors; (2) feature compression using the J2 criterion (Fisher’s discriminant criterion), without using any sklearn LDA; (3) classification using Support Vector Machine (SVM) with an RBF kernel and Linear Discriminant Analysis (LDA). The MNIST dataset is used with 10,000 and 2,000 samples for training and testing, respectively, across six methods. The maximum accuracy of 95.95% was obtained with combined LBP+HOG features, J2 compression, and SVM classification, reducing the 342-dimensional features to 9 discriminant components. The results show that HOG features significantly improve digit recognition compared to LBP alone, and that the J2 criterion can be useful for maximizing class separability and drastically reducing dimensionality.