Naked people retrieval based on Adaboost learning

Presents a learning scheme for judging whether there are any naked people in an image. First, learning vector quantization is used to build several classifiers based on the low-level features, such as color histogram, region shape, texture and etc., which are extracted from the images. The best classifier performs a recognition ratio of 81.3%, so the Adaboost learning method is applied to combine these classifiers to form a stronger one, and the final classification achieves a result of 86.0% on the test set. Adaboost's ability in multi-feature analysis is emphasized, and the proposed algorithm is important both for the blue-picture-filter in the WWW and for semantic image indexing in content-based image retrieval. In experiments, the groundtruth is made of 1,200 images and the test set is independent from the training set.