In this paper, we present a novel algorithm to classify UAV images through the image annotation which is a
semi-supervised method. During the annotation process, we first divide whole image into different sizes of blocks and
generate suitable visual words which are the K-means clustering centers or just pixels in small size image block. Then,
given a set of image blocks for each semantic concept as training data, learning is based on the Probabilistic Latent
Semantic Analysis (PLSA). The probability distributions of visual words in every document can be learned through the
PLSA model. The labeling of every document (image block) is done by computing the similarity of its feature
distribution to the distribution of the training documents with the Kullback-Leibler (K-L) divergence. Finally, the
classification of the UAV images will be done by combining all the image blocks in every block size. The UAV images
using in our experiments was acquired during Sichuan earthquake in 2008. The results show that smaller size block
image will get better classification results.
|