Authors:

Aryan Shahu

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

The purpose of this study is to find better ways to test and improve image classification accuracy with different image styles (ink wash, oil paint, water paint). Our work aims to create a tool that is an evaluation framework for image generation models. This tool is meant to analyze various image generation models and benchmark them based on various image accuracy criteria. One of these criteria is the image style, which have been broken up into three specific styles: ink wash, oil paint, and water paint. The purpose of this sub-team is to increase the accuracy at which the image classification model categorizes generated images. We are working on testing generative models (text to image) and evaluating its performance through quantitative metrics. Current generative models still fail to respond to complicated prompts and could deliver inaccurate results. Our goal is to reach a better accuracy in predicting the performance of these models by applying multiple methods such as hyper parameter optimization. Also, we are planning on using other versions of the Vision Transformer. One version we already applied is Compact Convolutional Transformer and we are aiming to apply more models. Another idea to improve the accuracy is to use more training data. After applying the Compact Convolutional Transformer, we were able to get 93% total accuracy. We aim for a better accuracy after applying the other methods discussed in this research. Keywords: Vision Transformer; Style Classification; Compact Convolutional Transformer; Hyper Parameter Optimization; Model Accuracy

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Aryan Shahu