Publication Date
Spring 2026
Degree Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Engineering
Advisor
Magdalini Eirinacki,; Bertin Cordova-Diba; Jahan Ghofraniha,
Abstract
A critical process in digital image generation is the construction of images through layers. Professionals may not always have access to these layers and are often limited to the final static image, which makes editing and refinement difficult and cumbersome. To address this gap, research has explored image layer decomposition, the process of breaking an image down into intermediary layers for subsequent editing. However, a limitation of these applications is that they are rigid in the types of layers that are generated. This work introduces a novel artificial intelligence framework: Layer-RL. Layer-RL is grounded in reinforcement learning and focuses on advancing the capabilities of image layer decomposition to be more dynamic. Layer-RL sequentially generates layers, with an agent adding one at a time until the target image for decomposition is reconstructed. Unlike prior approaches, which often restrict decomposition to a fixed number or type of layers, this method enables a variable number, allowing images to be broken into their most fundamental components. This flexibility enhances both interpretability and usability. The resulting outputs are semantically meaningful and directly editable, offering a powerful tool for professionals to use. By resolving the challenges associated with achieving both adaptability and succinctness in layer definitions, Layer-RL constitutes a significant advancement toward meeting the growing demand for efficient and high-fidelity editing of generated image content.
Recommended Citation
Raven, Lindsey A., "Image Layer Decomposition Using Deep Reinforcement Learning" (2026). Master's Theses. 5776.
DOI: https://doi.org/10.31979/etd.my53-t3b9
https://scholarworks.sjsu.edu/etd_theses/5776
