Abstract: Prior contour detectors based on single scale input images usually extract multi-scale low-level cues to increase the detection accuracy, however, this kind of methods do not improve accuracy much due to the extracted noisy boundaries. The reason behind this is the lack of object-level features. To address this problem, this paper proposes a boundary fusion method which is capable of combining low-level and object-level cues on multiple scales. For a single scale image, at the beginning, we use Trimaps and an image matting method to generate object-level features, then we combine the object-level features with extracted low-level cues. For multi-scale images, we fuse the boundary maps extracted from each single scale image to reduce noisy boundaries. Experiments show that our method outperforms the single-scale boundary detector on the BSDS500 dataset.

Key words: computer application, boundary detection, multi-scale, image matting, image processing, computer vision