CATEGORIZING SHOCKING VISUAL CONTENT ON SOCIAL MEDIA USING A DEEP LEARNING APPROACH

Abstract

The prevalence of shocking visual content has significantly increased due to the extensive utilization of social media platforms. Deep learning methods have emerged as highly effective tools for automatically categorizing such visual contents. The primary objective of this study was to develop a deep learning-based framework specifically focused on classifying shocking visual content extracted from social media platforms. To accomplish this, it was crucial to establish a comprehensive understanding of the defining characteristics of shocking visual content, encompassing various manifestations of shocking content across diverse platforms and contexts. The categorizing of shocking visual content aligns well with deep learning techniques. A model was constructed to categorize shocking visual content obtained from widely used social media platforms such as Facebook, Instagram, and Twitter. To create an appropriate dataset, visual content were collected from public profiles, posts, and user submissions likely to contain shocking content. Essential preprocessing steps were implemented to ensure thorough cleaning and proper preparation of the dataset for training. Next, ShockNet, a convolutional neural network that can distinguish between shocking and non-shocking visual content, is designed. The final step involves testing and training the designed model using the original datasets,augmented datasets, resized datasets as well as both the augmented and resized datasets. Additionally, the designed model is compared to several pre-trained convolutional neural network models, including VGG16, DenseNet121,InceptionV3 with attention, InceptionV3, ResNet50, and ResNet50 with attention. The dataset contains 15266 shocking and non-shocking visual content. Using multiple scenarios the hyperparameters were selected promising results using the 80:20. Through multiple experiments, ShockNet model demonstrated promising results in terms of categorizing metrics report among all models, using both the augmented and resized dataset. The ShockNet model achieved a training accuracy of 99.62%, a testing accuracy of 99.9% and Categorizing report accuracy of 97%. This study developed ShockNet, a deep learning framework for classifying shocking visual content on social media, achieving high accuracy and contributing to improved content moderation.