A Comparison of HOG Feature Descriptor using SVM with Convolutional Neural Network using SVM for Image Classification
Computer vision applications are present in almost every activity in the world, the classification of images is one of them, however, it still presents some problems in the accuracy of its classifications, which can cause problems when these applications are used in systems that require a lot of certainty of classification. Many innovations, methods, and techniques have been proposed so far, they highlight with greater success the use of Convolutional Neural Networks (CNN), Support Vector Machine (SVM), selectors and feature descriptors.
In this work, we emulate an architecture that combines a convolutional neural network (CNN) with a linear SVM for image classification and other method based on histogram of orient gradient (HOG) feature descriptor using SVM, both works were compared to assess the accuracy, taking as dataset the MNIST hand-written digit dataset the Fashion-MNIST dataset.
The architecture that combines a convolutional neural network (CNN) with a linear SVM for image classification got better accuracy than the method HOG feature descriptor using SVM, with values of 99.32% and 99.10% respectively, tested with MNIST hand-written digit dataset and values of 99.29% and 87.29% respectively, tested with Fashion-MNIST dataset.