Classification of man-made objects and nature-made objects is an essential task in region segmentation of natural images. We studied this task using both independent component analysis (ICA) and orientation maps. ICA bases were computed from image blocks of man-made objects and nature-made objects. Since an image block can be reconstructed by ICA bases and their weight coefficients, we used the weights coefficients computed by both image blocks of man-made objects and image blocks of nature-made objects. A support vector machine (SVM) was used to learn the weight coefficients for classification. Cross-validation was used to determine the optimal parameters for the SVM. We also carried out experiments using the SVM to classify blocks in a natural image in which there were both man-made objects and nature-made objects. The ground truth was made by visual inspection. In comparison with ground truth, an average F-value of 0.55 (man-made objets) was obtained as classification accuracy. Orientation maps of the blocks of images were also learned by the SVM to classify man-made objects and nature-made objects. The experimental results showed that an orientation map is not effective for classifying man-made objects and nature-made objects.
NETL
