Automated Semantic Labelling of Images Generated with Deep Diffusion Probabilistic Models

Images labeled with semantic information are of paramount importance for training and evaluating supervised deep learning models. However, manual labeling is a costly process in terms of time. On one hand, the recent DatasetGAN model tries to reduce this gap by generating images along with pixel-wise semantic information through a multi-layer perceptron. On the other hand, diffusion models allow for synthesizing high-quality images improving Generative Adversarial Networks (GANs). This work presents a new supervised model to generate labeled datasets based on the combination of both methodologies. A diffusion model is used as an image generator, and a multi-layer perceptron uses its internal state to predict the segmentation masks. Experiments on the ImageNet \(512\times 512\) dataset showed a significant performance in creating the output labels, and applying the labels in semantically meaningful ways to different ImageNet classes.