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:


작성자 관리자 날짜 2021-06-08 22:43:49 조회수 3


1.     Introduction

2.     Main Algorithm and Principle

3.     Demo

1.     Introduction

-       Images of outdoor scenes are often degraded by turbid medium, such as dirt particles and water droplets, in the atmosphere. Light is scattered or attenuated as it travels through these particles, resulting less image radiance reaching the imaging sensor. Scene information is further corrupted by so called the “air-light” originated from ambient light reaching the sensor by the scattering medium. It can be said that the image degradation due to air-light and the attenuated light from the scene is directly linked to the distance from the scene to the imaging sensor. Image degradation extends to scene color content since light attenuation through a turbid medium is not spectrally uniform.

-       There have been a number of techniques developed for haze removal to enhance scene visibility and restore color content for outdoor applications such as intelligent surveillance system, remote sensing and object recognition since the algorithms for feature detection, filtering and photometric analysis depends heavily on scene radiance. Due to their effectiveness, some of these methods are becoming an integral part in consumer and computational photography and computer vision.

2.     Main Algorithm and Principle

-       Hazy Imaging Model

n  The dehazed image J can be restored from the hazy image by estimating the atmospheric light A and the transmission t from the hazy imaging model.


-       Atmospheric light estimation

n  A is estimated by quad-tree subdivision using a transformed image. A gray-scaled input image is transformed by minimum filtering in a non-overlapping fashion to eliminate the adverse effects due to bright values of a local object. Then, we search the most haze-opaque region by quad-tree subdivision and estimate the atmospheric light as the color vector which minimizes the Euclidean norm, ║ (rpgpbp) – (1, 1, 1) ║


-       Transmission estimation

n  The transmission estimation is optimized by maximizing objective function which is comprised of two functions (entropy and information fidelity).

n  Image entropy:  

n  Information fidelit:y 

n  Objective function: 

n  Local transmission optimization: 



3.     Demo

댓글 (0)

등록된 댓글이 없습니다.
작성 권한이 없습니다.