Year: 2024
Type: dissertation
Abstract: The use of digital media files in our society, such as images and videos, is increasing as the number of users who rely on social media applications to communicate with their peers is growing. Applications that are mostly used on mobile devices that have limited access to computational resources. The increasing importance of digital media platforms leads to the need for more suitable image quality, where the field of image quality assessment plays a vital role. The use of deep learning frameworks (convolutional neural networks) to support image quality assessment tasks has shown that it can improve stateof- the-art results. The main focus of this thesis is to improve the use of the MobileNetV3 architecture, a neural network focused on resource-constrained devices, on a full-reference image quality assessment (FR-IQA) task by changing the original attention module with three different ones. To investigate this, we propose three new MobileNetV3 models by swapping the original attention module with the Convolutional Block Attention Module (CBAM), Pyramid Squeeze Attention module (PSA), and Channel Attention Module (CAM) from the DANet network and benchmark the proposed model’s performance against the original MobileNetV3 on the KADID 10K database. The experiment results showed that all three proposed models exhibited better performance on the FR-IQA task while reducing the number of parameters of the networks when compared with the original MobileNetV3. The best two variantswere the model that used the PSA and the CAM attention modules.
Author: Redinha, Duarte Guerra
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7065903960