Semi-Supervised Semantic Image Segmentation by Deep Diffusion Models and Generative Adversarial Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pixel-level annotation for deep learning image segmentation is costly and time-consuming.
• Semi-supervised learning offers a solution to reduce annotation requirements.
• Generative Adversarial Networks (GANs) have been adapted for semi-supervised segmentation.

Purpose of the Study:

• To propose MaskGDM, a novel deep learning architecture for improved semi-supervised image segmentation.
• To integrate generative diffusion models with existing GAN architectures like EditGAN.
• To evaluate the performance of the proposed model on multi-class and binary segmentation tasks.

Main Methods:

• Developed MaskGDM, a hybrid architecture combining EditGAN principles with generative diffusion models.
• Trained and evaluated the model on multiple image segmentation datasets.
• Compared performance against established models like EditGAN and DatasetGAN.

Main Results:

• MaskGDM demonstrated improved performance in multi-class image segmentation compared to EditGAN and DatasetGAN.
• The model achieved significant improvements in binary image segmentation on the ISIC dataset.
• Quantitative results show enhanced accuracy with reduced annotation needs.

Conclusions:

• Integrating generative diffusion models enhances GAN-based semi-supervised image segmentation.
• MaskGDM offers a more efficient and accurate approach to image segmentation.
• The proposed method shows promise for various segmentation applications requiring less annotated data.