National Institute of Technology Rourkela

राष्ट्रीय प्रौद्योगिकी संस्थान राउरकेला

ଜାତୀୟ ପ୍ରଯୁକ୍ତି ପ୍ରତିଷ୍ଠାନ ରାଉରକେଲା

An Institute of National Importance
NIT Rourkela Inside Page Banner

Seminar Details

Seminar Title:
Development of Lightweight Deep Learning Frameworks for Robust Underwater Scene Analysis
Seminar Type:
Registration Seminar
Department:
Electronics and Communication Engineering
Speaker Name:
Arka Bairagi ( Rollno : 524ec1006)
Speaker Type:
Student
Venue:
EC Seminar Room (EC-303)
Date and Time:
30 Oct 2026 5.30 PM
Contact:
Sobhan Kanti Dhara
Abstract:

Underwater images often exhibit significant color casts, blurriness, haze, and diminished fine textures due to wavelength-dependent light attenuation and scattering. Shorter wavelengths penetrate more deeply, while longer wavelengths are absorbed more quickly, causing severe color distortions dominated by blue or green tones. These distortions affect several downstream computer vision tasks, such as object detection and scene segmentation, across various applications, including industrial inspections, intruder detection, and marine ecosystem monitoring. Initial studies in the domain primarily focused on restoration to handle image degradation for effective detection and segmentation. Recently, a plethora of work has focused on object detection and segmentation from distorted input images only. However, most of those existing methods rely on computationally intensive architectures and usually remain sensitive to complex backgrounds, color cast, varying noise levels, and illumination changes.  This necessitates the development of computationally efficient and robust techniques that can handle diverse real-world conditions for real-time applications.

The work and studies are structured into three parts: Part I: Literature Survey and Problem Statement Part II: Research Work Carried Out So Far and Part III: Future Scope. The first part reviews recent techniques for underwater scene analysis, identifies existing research gaps, and formulates the problem statements and proposed research directions. In the second part, two novel underwater image enhancement techniques are proposed: JCIC-UIE (Joint Color and Illumination Correction for Underwater Image Enhancement) and FAR-Net (Frequency-based Attention Refinement Framework for Underwater Image Enhancement). Both architectures combine an encoder-decoder with an attention mechanism to capture local and global information from degraded RGB underwater images. The performance of the proposed techniques is evaluated on benchmark datasets including LSUI, UIEB, UFO-120, and U45. The performance of the proposed techniques is compared with several state-of-the-art techniques using PSNR, SSIM, LPIPS, and UIQM. Experimental study shows that JCIC-UIE and FAR-Net outperform other state-of-the-art methods in terms of robustness and computational cost. Finally, the third part discusses future research directions and outlines a roadmap toward real-time underwater image enhancement and analysis.