Software Department Discusses PhD Dissertation on Image Fusion Using Attention-Based Deep Learning

Duhaa Fadill Abbas
The Software Department at the College of Information Technology is set to discuss a PhD dissertation titled "Image Fusion Using Attention-Based Deep Learning Techniques" by PhD candidate Rusel Mohammed Neama , under the supervision of Dr. Tawfiq Abdulkhaleq Abbas. The discussion is scheduled for Thursday, March 6, 2025, at 9:00 AM in the Conference Hall of the College of Information Technology.

The dissertation highlights Image fusion is a technique used to combine information from multiple images into a single, more useful image. It addresses several challenges, including the limited information in individual images, as they often capture only a portion of the available scene details. Additionally, it reduces noise, as sensor-captured images can be affected by noise, obscuring important details. It also improves spatial and spectral resolution by leveraging the strengths of different imaging sensors and enhances object detection and recognition, making image fusion technology widely applicable in fields ranging from medical diagnosis to military surveillance.
The proposed system introduces an approach for merging visual and thermal images using deep learning networks, employing an encoder-decoder network to extract features and reconstruct the image while incorporating multi-headed attention and spatial attention to enhance and merge the features. The approach consists of four phases: The first phase, pre-processing, includes resizing, gray-level conversion, and normalization, resulting in normalized grayscale images.
• The second phase, feature extraction, utilizes visual and thermal images as input and consists of four blocks, each extracting features at different significance levels.
• The third phase, feature fusion, also consists of four blocks, each containing two branches where features are dynamically adjusted during training and merged using concatenation to preserve image details and prevent information loss, resulting in integrated characteristics.
• The fourth phase, image reconstruction, employs seven blocks to reconstruct the fused image from the features extracted in the previous phase. Experimental results demonstrate.
that the proposed approach outperforms many modern methods, as it has been implemented and evaluated using four types of datasets: visual and thermal images, medical images, and aerial images. The model was trained using a dataset different from the one used in testing, specifically the LLVIP dataset for visible and thermal images.