Automatic Explicit Content Detection in Anime: Comparison of DETR and YOLO

Explicit visual content become an inseparable issue from the internet and entertainment industry, including anime. The frequent portrayal of scantily clad female characters with exaggerated physical features raises concerns about potential societal effect, like addiction, desensitization, and overexposure to sexualized imagery. To address this issue, this study proposes an early-detection model to identify explicit visual representations and sexually suggestive elements in anime content, contributing to automated content moderation and safer media consumption. A comparative evaluation was conducted between two leading object detection architectures, DETR and YOLO, to assess their performance in detecting sexually suggestive anime content. The dataset used comprises 800 manually annotated images sourced from popular anime featuring female characters. In the DETR experiments, ResNet-50 and ResNet-101 used as a various backbone and were modified using dilated convolutions at specific stage. The best model was YOLOv9t outperformed all other models across key metrics, achieving AP@50:95 of 0.580, AP@50 of 0.967, AP@75 of 0.615, and AR of 0.905. This performance is primarily attributed to the use of the Generalized Efficient Layer Aggregation Network (GELAN) backbone, which improves feature representation, and the Programmable Gradient Information (PGI) mechanism, which enhances gradient flow and training stability. Among the DETR variants, the best-performing model was the ResNet-50 backbone with dilated convolution at stage 5, achieving an AP@50:95 of 0.479, AP@50 of 0.875, and AP@75 of 0.485. While dilated convolutions improved small object detection by expanding the receptive field, their application in early stages reduced spatial precision and hindered performance on medium and large objects.