基於機器學習方法的顯著性偵測

Abstract

在計算機視覺中，顯著性偵測模擬了人類的視覺系統並且廣泛運用在許多影像 處理的運用上，例如適應性影像壓縮、物件辨識、影像搜尋及物件偵測。 在過去，顯著性偵測使用了基於超像素的方法搭配上低階特徵與啟發式的法則 來解決。而在過去幾年，卷積神經網絡(CNN)在電腦視覺上呈為主流，然而，因 為卷積神經網絡需要網格狀的輸入而超像素在形狀上基本上是不規則的，要將超像素與卷積神經網絡做結合是非常困難的。 在本作中，我們將基於超像素的方法與學習式演算法以兩種方式做結合，第一 種是基於支持向量機(SVM)的方式，此方法使用了基於超像素的特徵以及使用 深度合併模型(DMMSS)來進行超像素的合併，最後使用了支持向量機來做為 分類器。第二種方法我們提出了新的顯著性偵測網路叫做圖顯著網路(GSN)， 我們使用了圖卷積網絡(GCN)做為我們主要的架構，並且搭配了跳躍知識網路 (JKNet)做為骨幹。 本作最主要的貢獻在於以兩種方式結合了學習性演算法以及基於超像素的特 徵，實驗上也顯示我們的結合性方法可以在顯著性偵測上達到非常好的表現。

In computer vision, saliency detection simulates human perception and is useful for several image-processing applications, such as adaptive compression, object recognition, image retrieval, and object detection. In the past, saliency detection problem was solved by superpixel-based method cooperate with low-level features and heuristic rules. Recently, the Convolutional Neural Networks (CNN)-based methods have been thrived in computer vision area. However, it was difficult to integrate CNN with superpixel-based method since CNN required grid-like input while superpixel generally has irregular shape. In this work, we integrate the superpixel-based method with learning algorithm in two different ways. The first one is Support Vector Machine (SVM) Based method. We revisit superpixel-based feature with the help of Deep Merging Model for Superpixel Segmentation (DMMSS) on merging superpixel and combining the feature with the SVM classifier. For the second method, we introduced a novel saliency detection neural network model called the Graph Saliency Network (GSN), which use the Graph Convolutional Network(GCN) as main architecture and the Jumping Knowledge Network(JKNet) as our backbone. The main contribution of this work is that we combine learning algorithm with the superpixel-based features in two ways. Experiment has shown that our integration methods can achieve high performance for the saliency detection problem.