This is the course project for both CSDS 600 - Deep Generative Models and CSDS 600 - Machine Learning and Causal Inference at Case Western Reserve University.
Team member consist of me, Ruilin Jin, and Tuo Liang.
Here's the abstract for our final project, more detailed final report can be found in my homepage.
The field of image inpainting has witnessed substantial growth, fueled by the need for advanced techniques in digital forensics, image restoration, and object removal. Traditional methods, however, often fall short in maintaining semantic coherence. Our research introduces a approach to image inpainting that integrates causal reasoning within a novel generative process, leveraging the strengths of Variational Autoencoders (VAEs) and Structural Causal Models (SCMs). We propose a model that adapts the Causal Layer of CausalVAE, enhanced by the structural elements of NVAE, to address three primary challenges: creating expressive neural networks, scaling up training for larger image sizes and groups, and maintaining training stability. This model focuses on the coherent reconstruction of missing or corrupted image regions through an understanding of causal relationships among features. Our approach not only enhances the semantic coherence and realism of inpainted images but also fosters interpretability in the latent space, paving the way for more reliable and comprehensible image restoration processes.