结合Transformer的扩散模型用于人脸美丽预测OA

Overfitting to noisy labels in the database leads to weak generalization capability and reduced prediction accuracy in facial beauty prediction tasks.To address the issue,a Transformer-integrated diffusion model for label denoising and reconstruction during training is proposed.The model learns conditional probability distributions to control the generation process through"classifier guidance,"comprising a conditional information encoder and a denoising network.First,pre-trained weights of Swin Transformer are transferred and fine-tuned to obtain preliminary predictions as output priors.Second,these priors are utilized as the mean of the endpoint in the reverse process of the diffusion model,regulating denoising transitions at each timestep.Finally,facial beauty features are extracted and fed into the diffusion model for inference to generate prediction results.Experimental validation on three facial beauty databases,SCUT-FBP5500,LSAFBD,and CelebA,demonstrates that the proposed model outperforms baseline diffusion model and existing facial beauty prediction methods.In terms of accuracy,the model achieves 76.50%,72.65%,and 81.78%on the three databases respectively,surpassing the baseline diffusion model by 0.73%,1.76%,and 1.12%,and outperforming existing facial beauty prediction methods by 1.00%,4.42%,and 0.37%.The approach effectively addresses noisy label issues,enhances prediction performance,and can be widely applied to other image classification tasks or related fields.