Image & Vision

Investigates Image data fusion techniques that combine image and track data from multiple sensors to achieve improved accuracies and more specific inferences than could be achieved by the use of a single sensor alone. Our aim is to explore the state-of-the-art image processing algorithms for achieving effective data fusion as in:


low visibility and biased color in underwater images

작성자 관리자 날짜 2021-06-08 23:09:46 조회수 592



1.     Introduction

2.     Main Algorithm and Principle

3.     Demo

1.     Introduction

-       Images of underwater scenes are degraded by turbid medium, such as water particles and plankton, in lakes, oceans, and rivers. Image degradation extends to scene color content since light attenuation through a turbid medium is not spectrally uniform. This results in low visibility and biased color in underwater images.

-       In underwater environments, it is important to acquire the clear image since the most computer vision algorithm, from low-level image analysis to high-level object recognition, usually assume that the input image is the scene radiance.

-       Underwater images captured from optical sensor mounted in UUV (Unmanned Underwater Vehicle) are used in various fields such as inspection of plants, seabed exploration, searching for wrecks and the exploration of natural resources and it have an advantages in the visual impression, the immense information content in texture and object reflectance as well as the possibility of an intuitive visual inspection compared to underwater imaging based on sonar sensors.

2.     Main Algorithm and Principle

-       Color Correction

n  Underwater images have biased color due to the multi-scattering effects and different absorption of wavelengths of light. Most underwater images show green or blue since the green and blue light travel the longest distance through water for its shortest wavelength.  This condition may result color distortion in a restored scene radiance. Therefore, it is customary that underwater image is processed by color correction before we restore the scene radiance. We adjust the color shifting using biasness image and average of gray-scaled image



-       Restoring Scene Radiance

n  After color correction, we estimate the homogeneous background light and transmission to restore the scene radiance. The candidate region for estimating the homogeneous background light is selected by quad-tree sub-division using image performed by min operation in local patch. Then, the homogeneous background light is selected as the k-th color vector which minimizes the Euclidean norm, ║ (rkgkbk) – (1, 1, 1) ║ among pixels in the finally selected candidate region. Then, transmission is optimized with conditional entropy criteria.



3.     Demo

댓글 (0)

등록된 댓글이 없습니다.
작성 권한이 없습니다.