Single image main objects extraction via stacked sparse auto-encoders using sharpness information

Extracting main object from photos is prerequisite for image processing and semantic image understanding in many areas especially in multimedia signal processing at internet. So far, either human interaction in single image or sequence image frames are required for the extraction and most of them still rely on hand-crafted features. In contrast, the proposed work cast the human boundary detection in the daily photos as a patch-level binary classification, where the features are learned directly and automatically from the raw pictures and corresponding sharpness information via a stacked sparse autoencoder model. The experiments on the figure database from Baidu have demonstrated that the proposed method is able to extract the human boundary in single image without any human interaction, even in the various complex backgrounds.