Technical gazette, Vol. 32 No. 3, 2025.
Original scientific paper
https://doi.org/10.17559/TV-20240508001530
Advancements in Efficient Underwater Image Restoration Using ETransMapNet for Enhanced Dehazing
C. P. Indumathi
; Department of Computer Science and Engineering, University College of Engineering, BIT Campus, Anna University Tiruchirapalli-24, Tamilnadu, India
Haya Mesfer Alshahrani
; Department of Information Systems, College of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia
N. A. Natraj
; Symbiosis Institute of Digital and Telecom Management, Symbiosis International (Deemed University), Pune, Maharashtra, India
*
C. H. Sarada Devi
; Department of CSE Meenakshi College of Engineering, India
* Corresponding author.
Abstract
Underwater (UW) information is essential for advancing human exploration and utilization of the underwater world, including fields such as UW Paleology, UW Target Detection, UW Object Tracking, UW Surveillance, and related activities. Visual media like movies and images enhance our natural understanding of underwater objectives. In the past decade, underwater photo restoration and enhancement have gained increasing attention. This study proposes a novel approach employing the recently developed Convolutional Neural Network (CNN) for dehazing, named ETransMapNet (Efficient Transmission Map Network). ETransMapNet is designed with convolution layers and nonlinear activations to execute four sequential processes: nonlinear regression, local maxima detection, multi-scale decomposition, and convolutional feature extraction. Unlike traditional CNNs, ETransMapNet replaces the initial layer's Rectified Linear Unit (ReLU) activation with a convolution layer utilizing a Maxout activation function. ETransMapNet extracts features using three convolution kernels of different sizes (3 × 3, 5 × 5, and 7 × 7). The method suppresses noise in the estimated transmittance map, while local extremum values maintain local consistency within the transmittance map. This study adopts Bilateral ReLU (BReLU) for normalizing network outputs within a 0 to 1 range. Additionally, Adaptive Bilateral Filtering (ABF) is applied to remove redundant artifacts from the predicted transmission map. White balancing addresses color divergence, and Laplacian pyramid fusion combines the color-corrected and dehazed images. In the final stage, the resultant image is transformed into the Wavelet and Directional Filter Banks (WDFB) domain for denoising and edge enhancement. Performance metrics reveal that the proposed ETransMapNet approach improves performance by 38% - 50% compared to previous methods.
Keywords
Adaptive Bilateral Filtering (ABF); Convolutional Neural Network (CNN); ETransMapNet; transmission map denoising; underwater image restoration
Hrčak ID:
330579
URI
Publication date:
1.5.2025.
Visits: 473 *