表題番号:2016S-117 日付:2017/04/10
研究課題多チャンネル画像のエッジ保存平滑化のための高速な定数時間フィルタの実現
研究者所属(当時) 資格 氏名
(代表者) 理工学術院 情報生産システム研究センター 助手 杉本 憲治郎
研究成果概要
Edge-preserving smoothing filter is a fundamental tool for many applications in image processing and computer vision. This research targeted the bilateral filter (BF), which is one of major edge-preserving smoothing filters. The naive BF has the drawback that it requires much higher computational complexity than traditional linear filter such as Gaussian filter. For the last decade, a variety of accelerated algorithms of the BF has been proposed and they have focused on grayscale images, but not on color or hyper-spectral images. We therefore discussed fast algorithms of the BF for multi-channel images.

This research presented an efficient BF algorithm for multi-channel images using soft-assignment coding, which is an essential tool in machine learning. In order to accelerated the BF, existing state-of-the-art methods attempted to roughly describe the color distribution of an image by vector quantization or interpolation technique. The proposed method can provide the advantages of both approaches by using soft-assignment coding and achieved more efficient performance. Experiments using the test image set “Kodak Photo CD” showed that the proposed method can run 3.5~4.0x faster than the state-of-the-art methods without a loss in approximate accuracy. We won an IE award since this research outcome was highly evaluated.