利用水平與垂直方向上顏色差異實現影像顯著性偵測

Abstract

圖片的顯著性偵測(Saliency Detection)在電腦視覺(Computer Vision)中是一個基礎的問題，對於許多的應用像是物體分割(Object Segmentation)、自適性壓縮(Adaptive Compression)、物體辨識(Object Recognition)能有所幫助。過往的研究結果顯示出獨特性(Distinctness)在顯著性偵測中扮演著主導的因子，而這些研究則是利用各種不同的演算法去計算獨特性，如顏色、方向、樣式等低階的信息或是其他如臉部的先驗等高階信息。本篇論文設計了一個新的低階特徵來協助偵測。首先使用圖片分割(Image segmentation)將圖片切割成數個同質的區域並取出全域與局域的顏色特徵以及背景先驗。接著由水平向與垂直向的顏色差異來計算其在方向與顏色上的特殊性。此外我們建立連通元件(Connected-Component)來強調一個完整的物件。最後，經由臉部偵測加入臉部先驗的特徵，並整合這些特徵以產生偵測的結果。實驗結果顯示本論文之方法能夠找出圖片中顯著性較高的區域及物件，並且能夠保有同一物件內部擁有較完整的顯著性。

In computer vision, saliency detection of image is a fundamental problem, it’s helpful for applications like object segmentation, adaptive compression, and object recognition. Previous works demonstrate that distinctness is the dominating factor, and these works try to use various algorithms to compute distinctness, low-level cues like color, orientation, and pattern or high-level cues like face prior. This thesis proposed a new low-level feature to assist the detection. First, we use image segmentation to segment the image into homogeneous regions then extract global and local color features and combined with background prior. Second, the particularity of color on horizontal direction and vertical direction are calculated. In addition, we create connected-component to emphasis the completeness of an object. Finally, we obtain high-level cue of face prior by face detection and then integrate these features to generate the detected result. The experiment results show that the proposed method can find the region and object with higher saliency, and retain the completeness inside of the object.