基於影像特徵對之顯著區域偵測技術

Abstract

在本論文中，我們提出一套偵測人眼視覺顯著區域之技術。給定一張影像，藉由我們提出的演算法可以立即判斷出哪些位置區塊會是人眼較易去注意的地方。輸入影像會先被拆解成三種通道，包含了強度跟兩個對比色彩通道。對於個別通道，會將其建構成特徵對的分布圖，並藉由分析特徵對分布圖的結果反映射回空間域去識別出視覺顯著區域。此外，為了抑制雜訊造成的影響，我們另外加上了正規化的步驟，以提高顯著區域劃分的成功率。根據實驗結果，我們發現此技術確實可以偵測出人眼視覺的顯著區，同時過濾掉較不重要的資訊。

In this thesis, we propose an algorithm for the detection of human visual saliency regions. Given an image, the proposed algorithm can automatically determine these locations where humans tend to pay more attention to. The image is first decomposed into three channels, including one intensity channel and two opponent-color channels. For each channel, a feature-pair distribution is created for saliency analysis, and the analysis result is mapped back to the spatial domain to identify visually salient regions. Beside the suppression of noise interference, a normalization stage is included to improve the performance of detection. As demonstrated in the experimental results, the proposed method can successfully identify visual saliency regions in human visual reception and, at the same time, filter out less crucial information.