Semi-supervised edge learning for building detection in aerial images

In this paper, a new building detection scheme using semi-supervised edge learning is proposed. This scheme utilizes a feature based on edge flow to delineate the patterns of sharp contrast at the edges of building. The contrast patterns with their distribution in the features space based on similarity metric provide discriminative evidences for the building detection. By the extended kernelBoosting, the semi-supervised edge learning, a number of Gaussian Mixture Models (GMMs) are computed and optimized to model the local distribution of contrast patterns according to their similarity. The 'weak kernel' hypotheses are then generated from these optimized Gaussian Mixture Models. The final kernel is defined by accumulating a weighted linear combination of such "weak kernel". The kernel function can then be used for classification with kernel SVM. Experiments show that this scheme is capable of achieving both low false positive rate and low false negative rate with only a few training examples and that this method can be generalized to many object classes.