Image Classification by Denoising Diffusion Probabilistic Model for Vector on Feature Space

Abstract

The Denoising Diffusion Probabilistic Model (DDPM) has recently gained significant interest in image generation applications. However, research regarding its potential for classification problems still needs to be completed. In this study, we propose an extension of DDPM to classification tasks. Our method involves assigning a vector to each image class and predicting the class vector using a conditional DDPM with an input image as a condition. We use the nearest neighbor method for classification to predict the closest class to the vector obtained from the conditional DDPM. We conducted experiments on the CIFAR-10 dataset using randomly assigned and pretrained word vectors for class representation. The results demonstrate that random vectors lead to only a 25% accuracy rate while using word vectors achieves a much higher accuracy rate of 63%. Our findings suggest that DDPM can solve classification problems with appropriate class representation.