Deep Learning Techniques for Banner Image Classification

Abstract

For centuries, people use many tactics to display their disagreement or rejection of an action or behavior that is not in favor of them. Usually, they assemble in large numbers and use banners, flags, slogans, etc., to exhibit their anger and frustration. Actually, the banners & flags work as an instigator for violent behavior in a crowd, on the other side; it also helps law enforcement agencies to understand the intent of the crowd. The goal of security forces is to identify and isolate unlawful material that may instigate violent action in the crowd. Thus, identification of banners at the genesis level of the crowd is very important to control and manage the probability of violent behavior in the crowd. To support security forces in their decision-making process, here we developed a system that employs pre-trained convolutional neural network (CNN) models to automatically recognize banners in an image or video. To achieve this, the present work also releases the DIAT Banner Dataset our novel dataset, mostly based on Indian protest events that include 1017 images. Using this dataset, we train pre-trained models and illustrate its result. A novel data augmentation technique referred to as the sliding window with histogram matching is introduced for dealing with the data insufficiency problem and also demonstrates its effectiveness as it improves the accuracies of all pre-trained models by an average of 7.37%. In this study, ResNet50 exceeds other pre-trained CNN models by obtaining 99.4% accuracy, after comparing the performance of several pre-trained models.