Single image dehazing based on vector quantization

• Main Authors: Jui-ChiangWen, 溫瑞強
• Other Authors: Shen-Chuan Tai
• Format: Others
• Language: en_US
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/69190335931091186410

碩士 === 國立成功大學 === 電腦與通信工程研究所 === 103 === The proposed method is based on McCartney’s optical haze model and uses a novel approach to estimate transmission. According to the literature, the major problem is estimating the transmission in the model-based method. This study trains plenty of haze-free and hazy images as codebooks with LBG algorithm. Then it is used to estimate transmission with matching. In order to speed up the process, the input image is down-sampled before refining with guided image filter. It not only can reduce processing time but also can preserve the quality of restored images. RGB, dark channel, and contrast values are regarded as features while training codebooks and estimating transmission. The transmission can be selected accurately because dark channel and contrast feature have complementarity. The experiment results show that the haze-free high-intensity objects can avoid over dehazing and keep the foreground of restored images more natural. The details of recovered images are also clearer.