Scene Classification with Simple Machine Learning and Convolutional Neural Network

Abstract

Scene classification has drawn a great deal of attention among scientists in the past decade. It focuses on developing a method that can be used to classify scene images into pre-defined categorizes. Unlike human, identifying discriminative features are still the main burden to accomplish this goal. As such, a question arises to determine the appropriate level of feature descriptors, i.e., fine-grained texture/shape or recognizable objects/components. In this paper, we like to compare if a simple machine learning method can compete with the deep learning algorithm. Since the former requires a feature vector of each training instance, we have come up with a method to extract object-based features from images. YOLOv3 that is a pre-trained model for object detection is exploited to produce the object-based features. Following that, four machine learning methods are evaluated using this training set, while the deep learning model is built using original images. The experiment shows simple machine learning models with object-level features can match the performance of deep learning, which is more demanding in terms of computational resources and time.