Abstract
In this thesis, we proposed two new algorithms for object detection. One is the image localization algorithm for invariant or rigid objects, which are very common in automatic visual inspection. The other algorithm is used to detect variant objects. We focus on the problem of face detection. In industrial inspection, we developed a new algorithm that can localize the object under non-uniform illuminations. In this algorithm, we first apply Sobel edge detection on an image to get the gradient map, and then we compute the local maximal gradient values to normalize the gradient map. The two stages of our alignment algorithm, i.e. the learning based pattern search and the energy minimization based alignment, perform alignment on this relative gradient map. The alignment system is not only robust against non-uniform illumination changes but also very efficient for image alignment. Our face detection algorithm consists of four steps, i.e. the skin-color filtering, template matching, verification and overlap merging. We used skin-color information to extract skin-color region from input image, and then detected face on skin-color regions. Our face detection algorithm extract features from face images and classify the features into several clusters in the training process. The means of these clusters are face templates. Our system calculates the minimal distance between the feature vector extracted from an image region and the face templates to determine if it is a face candidate. We apply two verification steps on the detected face candidates. The first one is the complete face comparison verification and the second is the combination of criteria from local face region analysis. At last, we merge the overlapped face candidate regions to obtain the final results.