Evaluation of Deep Learning Methods for Age-appropriate Sensitive Image Classification

Currently, communication media such as movies have a more mature grading system, but images widely disseminated on social media do not. Additionally, in the new media environment, the dissemination of sensitive images has become more covert and diversified, posing a potential threat to the mental health of adolescents. Therefore, solving the challenge of implementing age-appropriate grading for sensitive images using computer vision technology is crucial. Based on this premise, this paper first constructs an age-appropriate dataset for grading sensitive images targeted at teenagers to provide reference for image dissemination on social media platforms. Then, it evaluates the performance of eight deep learning models including ResNet101, ShuffleNet-v1, Vgg11, Vgg16, DenseNet101, MobileNet-v2, Swin Transformer, and Vision Transformer models in terms of their ability to recognize sensitive images. We finally conduct comprehensive tests on both a 2-category public dataset and a quaternary age-appropriate dataset; experimental results demonstrate that MobileNet-v2 network and Swin Transformer network exhibit superior performance on the two datasets respectively. The research findings presented in this paper serve as an important reference for managing image dissemination and classification on social media.