Creating Image Dataset With Semantic Diffusion Model
Name
Triin Schaffrik
Abstract
Semantic diffusion model is a neural network that allows image generation based on segmentation. The ability to generate images instead of manually creating them would save time in creating segmented image datasets. The goal of this bachelor’s thesis is to generate images using the semantic diffusion model and evaluate whether they improve the results of the segmentation model. 1640 images are generated based on the ADE20K training dataset. The generated images are added to the same training set used to train the DeepLabv3+ segmentation model. The results of the DeepLabv3+ models are evaluated based on accuracy, recall, and F1-score and compared with each other. The outcome includes the images created by the semantic diffusion model and a comparison of the model trained with these images with the model trained purely on the ADE20K dataset.
Graduation Thesis language
Estonian
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Joosep Kivastik
Defence year
2023