Automatic Shadow Detection from a Single Image

• Main Authors: Shun-Wen Cheng, 鄭舜文
• Other Authors: 李明穗
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/72901432280737229354

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 101 === Shadow detection is one of major topics in computer vision. Shadow usually causes interference when preforming image segmentation, tracking, face detection, and object recognition. Thus, detecting shadow region is an important step before shadow removal process. It is a challenging task that most of method need users to provide strokes to indicate location of shadow and only few of methods can automatically detect shadow region. This thesis proposed a method that can detect shadow automatically. The pairwise classification is performed to classify the relationship between all the pair of patches in the input image, and Gaussian Mixture Model (GMM) is applied to obtain main color statistic of the shadow patches and enhance accuracy of the detection, then the input image is upsampled to find the small area of shadow. Finally, the detected shadow region is refined by mean-shift segmentation. The experimental result shows the proposed method can generate accurate shadow map without any user assistance.