Transformers-Based Few-Shot Learning for Scene Classification in Child Sexual Abuse Imagery

Sexual abuse affects many children globally, with over 36 million reports in the past year. The vast amount of multimedia content exceeds law enforcement’s analysis capacity, necessitating reliable automated classification tools. While effective, deep learning methods require extensive data and costly annotations that are restricted to law enforcement. This Master’s thesis addresses these challenges using Transformer-based models for classifying indoor scenes, where such content is often found. Utilizing few-shot learning, the study reduces the need for extensive annotations, comparing classic few-shot models with Transformer-based models and exploring different methods for feature vector aggregation. The findings show that aggregating vectors using the mean is most effective, achieving 73.50 ± 0.09% accuracy with just five annotated samples per class. Evaluated with the Brazilian Federal Police, the model achieved 63.38 ± 0.09% balanced accuracy on annotated child sexual abuse indoor scenes, indicating the technique’s potential to aid preliminary screening efforts.