Event Title

UR-191 - Automated Image Colorization Through EfficientNet

Streaming Media

Document Type

Event

Start Date

28-4-2022 5:00 PM

Description

Automatic Image Colorization is the procedure of transforming a gray-scale image into a colored image without any human intervention. This field is highly researched and strongly applicable to the real world due to: historic importance, data generation/augmentation, and human satisfaction. The main objective of this research is to develop an artificial intelligence feature extraction method that implements color into a gray-scale image. To solve this problem, I relied on transfer learning through the EfficientNet model. The problem was treated in multiple parts, those being: the processing of images into features, feature extraction using the model, and then colorization via the luminosity channel. My model outperformed the base model marginally but saved vastly on the time constraint even within my limited resources. As stated in the baseline paper, the implementation of EfficientNet generates colored images without fail. However, a more complex model, as well as an even more complex cost function, is required to truly evaluate automated image colorization.

This document is currently not available here.

Share

COinS
 
Apr 28th, 5:00 PM

UR-191 - Automated Image Colorization Through EfficientNet

Automatic Image Colorization is the procedure of transforming a gray-scale image into a colored image without any human intervention. This field is highly researched and strongly applicable to the real world due to: historic importance, data generation/augmentation, and human satisfaction. The main objective of this research is to develop an artificial intelligence feature extraction method that implements color into a gray-scale image. To solve this problem, I relied on transfer learning through the EfficientNet model. The problem was treated in multiple parts, those being: the processing of images into features, feature extraction using the model, and then colorization via the luminosity channel. My model outperformed the base model marginally but saved vastly on the time constraint even within my limited resources. As stated in the baseline paper, the implementation of EfficientNet generates colored images without fail. However, a more complex model, as well as an even more complex cost function, is required to truly evaluate automated image colorization.