Automated Detection of Offensive Images and Sarcastic Memes in Social Media Through NLP
T Purnima, Koteswararao Ch · International Journal of Advanced Computer Science and Applications · 2024
In this digital era, social media is one of the key platforms for collecting customer feedback and reflecting their views on various aspects, including products, services, brands, events, and other topics of interest. However, there is a rise of sarcastic memes on social media, which often convey contrary meaning to the implied sentiments and challenge traditional machine learning identification techniques. The memes, blending text and visuals on social media, are difficult to discern solely from the captions or images, as their humor often relies on subtle contextual cues requiring a nuanced understanding for accurate interpretation. Our study introduces Offensive Images and Sarcastic Memes Detection to address this problem. Our model employs various techniques to identify sarcastic memes and offensive images. The model uses Optical Character Recognition (OCR) and bidirectional long-short term memory (Bi-LSTM) for sarcastic meme detection. For offensive image detection, the model employs Autoencoder LSTM, deep learning models such as Densenet and mobilenet, and computer vision techniques like Feature Fusion Process (FFP) based on Transfer Learning (TL) with Image Augmentation. The study showcases the effectiveness of the proposed methods in achieving high accuracy in detecting offensive content across different modalities, such as text, memes, and images. Based on tests conducted on real-world datasets, our model has demonstrated an accuracy rate of 92% on the Hateful Memes Challenge dataset. The proposed methodology has also achieved a Testing Accuracy (TA) of 95.7% for Densenet with transfer learning on the NPDI dataset and 95.12% on the Pornography dataset. Moreover, implementing Transfer Learning with a Feature Fusion Process (FFP) has resulted in a TA of 99.45% for the NPDI dataset and 98.5% for the Pornography dataset.