A Wavelet-based Dual-stream Network for Underwater Image Enhancement


Ziyin Ma and Changjae Oh

School of Electronic Engineering and Computer Science, Queen Mary University of London, UK

IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2022

Abstract

We present a wavelet-based dual-stream network that addresses color cast and blur in underwater images. We handle these artifacts separately by decomposing an input image into frequency bands using discrete wavelet transform, which generates the downsampled structure image and detail images. These sub-band images are used as input to our dual-stream network that incorporates two sub-networks, multi-color space fusion network and detail enhancement network. The multi-color space network takes the decomposed structure image as input and estimates the color corrected output by employing the feature representations from diverse color spaces of the input. The detail enhancement network addresses the blurriness of the original underwater image by improving the image details from high-frequency sub-bands. We validate the proposed model on both real-world dataset and synthetic underwater datasets and show the effectiveness of the proposed method in color correction and blur removal with low computation.

Method

overview of our framework

The sub-band images with multiple frequency bands are obtained by discrete wavelet transform (DWT), which facilitates to decouple the color cast and blurry details in underwater images and separately address these artifacts with fS(·) and fD(·), respectively. fS(·) and fD(·) are constrained by the structure loss, LS, and the detail loss, LD, respectively. The output, f(I), after inverse DWT (IDWT), is constrained by the adversarial loss, Ladv, with the generative adversarial network (GAN) discriminator, C(·).

Visual comparison

We list 12 real underwater images which can be processed by three physics-based methods and four learning-based methods. 1) Choose an image to filter, and 2) click the button below to see the result.

References

Name Title Year Author Paper
ULAP A rapid scene depth estimation model based on underwater light attenuation prior for underwater image restoration 2018 W. Song et al. [Paper]
IBLA Underwater image restoration based on image blurriness and light absorption 2017 Y.T. Peng, P.C. Cosman [Paper]
UDCP Transmission estimation in underwater single images 2013 P. Drews et al. [Paper]
UWCNN Underwater scene prior inspired deep underwater image and video enhancement 2020 G. Hou et al. [Paper]
WaterNet Deep underwater image enhancement 2018 C. Li et al. [Paper]
UIE-DAL All-in-one underwater image enhancement using domain-adversarial learning 2019 P.M. Uplavikar et al. [Paper]

Sources