Violence Image Recognition Based on Improved RepVGG

The dissemination of violent images on the Internet seriously jeopardizes the physical and mental health of adolescents. Utilizing deep learning techniques enables efficient screening and analysis of vast quantities of images propagated on the Internet. In order to enhance the network's understanding of brute-force images, this paper introduces the spatial and channel attention mechanism into the original RepVGG network. To prevent neuronal death, the Gaussian error linear unit (GELU) is used instead of the original activation function. Finally, experimental results on the constructed dataset demonstrate that the method adopted in this paper achieves higher accuracy compared to traditional networks, enabling more precise identification of violent and terrorist images.