LU2Net: A Lightweight Network for Real-time Underwater Image Enhancement

TL;DR

Underwater images suffer from absorption and scattering, hindering robotic perception. The authors present LU2Net, a lightweight U-shaped network that employs axial depthwise convolution and a channel attention module to achieve real-time underwater image enhancement with low computational costs. It delivers state-of-the-art enhancement while significantly reducing FLOPs and parameters, enabling real-time video processing on limited hardware (e.g., ~100 fps on GPU, ~12 fps on CPU). Evaluations on the LSUI dataset and real-world underwater ROV tests demonstrate higher PSNR, SSIM, and UCIQE, validating its practical utility for underwater robotics vision tasks.

Abstract

Computer vision techniques have empowered underwater robots to effectively undertake a multitude of tasks, including object tracking and path planning. However, underwater optical factors like light refraction and absorption present challenges to underwater vision, which cause degradation of underwater images. A variety of underwater image enhancement methods have been proposed to improve the effectiveness of underwater vision perception. Nevertheless, for real-time vision tasks on underwater robots, it is necessary to overcome the challenges associated with algorithmic efficiency and real-time capabilities. In this paper, we introduce Lightweight Underwater Unet (LU2Net), a novel U-shape network designed specifically for real-time enhancement of underwater images. The proposed model incorporates axial depthwise convolution and the channel attention module, enabling it to significantly reduce computational demands and model parameters, thereby improving processing speed. The extensive experiments conducted on the dataset and real-world underwater robots demonstrate the exceptional performance and speed of proposed model. It is capable of providing well-enhanced underwater images at a speed 8 times faster than the current state-of-the-art underwater image enhancement method. Moreover, LU2Net is able to handle real-time underwater video enhancement.

Figures (7)

• Figure 1: The illustration of LU2Net structure. Specially designed encoder and decoder blocks enable larger receptive fields and the adaptive adjustment of channel weights. Skip connection ensures the full utilization of multi-stage information.
• Figure 2: The comparsion of receptive field between axial depthwise convolution and traditional convolution. As the network get deeper with more convolution layers, the improvement of receptive fields in axial depthwise convolution is strengthened. Thus more details are perceived with the number of layers unchanged.
• Figure 3: The structure of axial depthwise convolution and pointwise convolution. In axial depthwise convolution, each convolution kernel operates on single input channel and the number of channel remains the same. While in pointwise convolution, each kernel combines all input channels and corresponds to an output channel.
• Figure 4: The illustration of CALayer based on channel attention module. The features of input channels are first extracted by average pooling and then adjusted by MLP-like architecture for channel weights. The adaptation of channels is carried out by element-wise multiplication of input channels and final weights.
• Figure 5: Our underwater ROV and experimental environment