Salient building detection based on SVM

This paper focuses on detecting salient buildings in a scenery image. A method based on bottom-up attention mechanism is proposed to detect salient buildings. Firstly, Haar wavelet decomposition is used to obtain the enhanced image which is the sum of the square of LH sub-image and HL sub-image. Secondly, the enhanced image is projected in the vertical direction to obtain the projection profile, and building candidates are separated from the background based on multi-level thresholding. Thirdly, the structure statistic features of buildings are extracted based on Sobel operator. The feature vector is formed by the number of long horizontal edges and that of vertical edges. Finally, linear support vector machines are used to classify buildings and the others. The proposed approach has been experimented on many real-world images with promising results.