Improving the Performance of Intelligent Photo Compliance Detection Method Based on Diffusion Model for Telecom User Management

As the user information management requirements in the telecommunications industry increase year by year, operators need to invest a large amount of manpower in auditing communication network registration photos. We employ the Stable Diffusion Model to generate telecommunications user registration photos, effectively avoiding detection issues such as class imbalance and privacy breaches. The diffusion model creates fake registration photos using prompt contexts adapted from typical photo non-compliance cases. Face detection and image classification models are utilized for the task of photo compliance detection. Experiments demonstrate that incorporating the images generated by the Diffusion Model enhances the classification capabilities for non-compliant photos, with peak performance achieved at a merge ratio of 10:1. Additionally, the Face+ models incorporating facial information outperform the base CNN and ViT models. This work provides a valuable reference for improving intelligent audit technologies used by telecommunications operators.