Removing Backscatter to Enhance the Visibility of Underwater Object


Supervisor: Lap Pui Chau

Underwater vision enhancement via backscatter removing is widely used in ocean engineering. However, due to the existence of dust-like particles and light attenuation, underwater images and videos always suffer from the problems of low contrast and color distortion.

In this project, we investigated the underwater light propagation process from a physical standpoint, studied the state-of-the-art methods of sloving the problem and finally proposed a novel and effective method based on underwater optical model and fusion technique to overcome the backscatter problem.

In general, state-of-the-art strategies to solve this issue can be categorized into two parts, one is physical-based model, which studys its physical process and then reverse derivation, while another is more focus on image processing techniques. Experiment shows that carefully designed physical-based models often achieve excellent in specific situations, but are not able to generate great results for varieties of underwater environments all the time. By contrast, image processing methods are more flexiable and perform good under various environments, although they cannot always guarantee to give good contrast and texture information.

In order to solve the underwater imaging issue well and also take the universality of the model into consideration, we investigated to incorporate the merits of both physical-based model and image processing model, while suppress their drawbacks at the same time. Hence, we came up with a new technique which utilize a novel fusion strategy to fuse the restored results from physical mdoel and image processing model. Our proposed method is mainly consist of three steps:

  1. We decomposed the input image into two components, reflectance channel and illuminance channel;
  2. We utilized color correction technology and dehazing technology to handle these two components separately;
  3. Finally, in order to rebuild result well, we applied the Gaussian and Laplacian pyramids based multi-scale fusion to reconstruct the target image while exposedness, saliency maps are utilized as weights to assist the fusion task.

The experimental results show that the proposed method is able to greatly improve the quality of distorted underwater images. By introducing the underwater image quality metric measurements, we also analyze the intrinsic information and objective feature indexes of restored images via different methods. In general, our method outperforms state-of-the-arts among sets of test images captured in different water environments and is demonstrated to be well-performed and effective. sample

Research Engineer, Institute of High Performance Computing (IHPC)

My research interests include vision grounding, natural language processing, computational linguistics and machine learning.