Adaptive Segmentation Techniques for Object Region Extraction from Images with Complex Lighting Environment

Automated image segmentation of objects in complex lighting environments is a difficult task since it is not possible to pre-define a threshold to distinguish the object from the background. It is ascertained that an effective threshold ctite1ion for such objects depends on the negative binomial distribution of the object region with respect to the background lighting. Based on this concept, a novel adaptive threshold selection method for automatic segmentation of the lumen region in endoscopic images is presented in this thesis. More precise lumen region and boundary are obtained by an integrated neighborhood search based region growing procedure whose speed is significantly enhanced by a quad-tree structure decomposition method. A boundary thinning and connecting algorithm, employing a novel search window on the preliminary boundary, provides a single-pixel-width-connected boundary. The new method does not need a priori knowledge about the image characteristics and is completely automatic. It reduces the computation time by 36% when compared to conventional threshold selection methods and facilitates high-speed response for real-time navigation of a vision-guided micro robotic endoscope. A new adaptive threshold selection method for skin extraction from the normalized r-g histogram of a face image is also presented in this thesis. It is observed from the spatial analysis of the r-g histogram that skin regions appear as clusters having negative binomial distribution characteristics, thus enabling accurate skin extraction from the background which is random or uniform. The textures of the candidate skin region are further analyzed using wavelet packet decomposition to obtain a set of feature vectors. Classification of the regions as face or non-face is done by evaluating the Bhattacharya distance of each feature vector to a prototype vector. The adaptive method has an improved accuracy of 9% when compared to traditional face detection techniques with images in varying lighting conditions. Research is progressing towards development of a relative variance based lighting compensation technique to segment static color images in normalized color space.