Manuscript received November 17, 2025; revised December 12, 2025; accepted January 22, 2025; published July 16, 2026.

Abstract—High-resolution underwater imaging with rich color details and real time performance in underwater robots is crucial in Underwater Image Enhancement (UIE). However, light absorption, deep-sea, haziness, scattering, and suspended particles degrade underwater image quality. It is leading to low contrast and color distortion, which negatively affect visual tasks. In UIE, deploying models on mobile platforms and embedded devices has challenges due to their limited computational capacity. Therefore, a lightweight and computationally efficient model is required to enhance underwater images in resource-constrained underwater robots. To overcome this issue, this paper introduces Deep2.6GNet, a lightweight U-shaped encoder-decoder architecture model. Unlike existing lightweight models, Deep2.6GNet proposes a novel integration of Depthwise Separable Convolution (DWSConv) with the simple Attention Module (SimAM). It significantly reduces Floating-point Operations Per second (FLOPs) and the number of parameters and computational demands in UIE. Moreover, experimental evaluations on the high Structural Similarity Index Measure (SSIM) and Underwater Image Quality Measure (UIQM) scores and competitive Peak-Signal-to-Noise (PSNR) values, confirmed that Deep2.6GNet offers an excellent balance between computational efficiency and enhancement quality. Deep2.6GNet outperforms other State-of-the-Art (SOTA) methods with only 2.6G FLOPs, 0.067M parameters and 0.01s for runtime on a computer with an NVIDIA GeForce MX230 GPU, Intel® Core™ i7 processor, and 24 GB of RAM. It is well-suited for underwater robots, particularly for tasks such as mapping, target detection, and navigation.

Keywords—underwater image enhancement, lightweight, computational efficiency, U-shaped encoder-decoder architecture, depthwise separable convolution, simple Attention Module (SimAM)

Cite: May Thet Tun and Tetsuya Shimamura, "Deep2.6GNet: A Lightweight 2.6 G FLOPs Architecture Model for Underwater Image Enhancement," Journal of Image and Graphics, Vol. 14, No. 4, pp. 543-550, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).