Exploring the Impact of Different Optimi-zation Modules on ResNet-50 Performance for Low-Resolution Images

Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis

Authors

Date

2025-04-14

Department

Major/Subject

Security and Cloud Computing

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

39

Series

Abstract

The goal of this paper is to explore whether some existing computer vision optimization modules can optimize the ResNet-50 network for image classification in the presence of low-resolution image inputs. The input image size requirement for the standard ResNet-50 network is 224*224, and most of the existing computer vision optimization modules are analysed and studied in the context of standard-sized inputs, while in many real-world situations the input images are often non-standard, so it is important to explore whether they can also be applied to low-resolution images as well. In this work, We chose cifar-100 as our dataset and ResNet-50 as our baseline model, in addition, we experimented and compared the optimization of four different optimization modules on top of the baseline, and found that these four optimization modules can improve the accuracy of the final results on top of the baseline by some from a low of 1.92% to a high of 5.83%.

Description

Supervisor

Kannala, Juho

Keywords

machine learning, deep learning, image classification, residual network, convolutional neural network, channel attention

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202505203869

Collections

[dipl] Perustieteiden korkeakoulu / SCI