利用多階段階級式融合架構的影像分割技術

Abstract

近年來，許多影像分割的方法使用超像素來減少運算成本，而有了深度學習的幫助，監督式學習的方法往往正確率比傳統的方法還要高。在這篇論文中，我們提出一個基於階層式超像素融合架構的新的影像分割方法。利用迭代的方式融合超像素同時更新結果，我們將影像分割轉換成一連串的超像素融合決策問題。與其他方法相比，我們從原影像提取大量特徵，利用特徵讓分類器學習如何分辨兩個超像素是否該合併。我們採用相同的階層式架構，把原本的超像素融合決策步驟改為用標準答案計算的標籤來產生訓練樣本。即使訓練集中只有一些訓練圖片，仍然可以得到大量的訓練樣本。而且，使用初始不同數量的超像素可以更進一步增加訓練樣本的數量，讓分類器達到更高的準確率與更好的穩定性。而多階級的結構用來實現可適性的超像素融合標準。在柏克萊分割資料集的實驗結果顯示我們的方法勝過其他的方法並且達到很高的影像分割準確率。

Recently, many image segmentation approaches are superpixel-based in order to reduce computation costs, and with the aid of deep learning, supervised methods achieve higher accuracy than traditional methods. In this thesis, we proposed a novel image segmentation method based on the hierarchical superpixel merging architecture for superpixels. With iterative operations of superpixel merging and result update, we transform image segmentation into a series of decision problems about superpixel merging. Compared with other superpixel-based methods, we extract a lot of features from original images, and decide whether two superpixels should be merged or not by learning classifiers with these features. We adopt the same hierarchical architecture for training sample generation by replacing superpixel merging decisions during testing with computing labels using ground truth. Even if there are only a few training images in the training set, a lot of training samples can be obtained. Furthermore, with different number of superpixels, we can get much larger number of training samples to make classifiers more accurate and robust. The multi-stage structure is used for implementing adaptive merging criteria for superpixels. Experiment results on the Berkeley segmentation dataset show that our proposed method outperforms other state-of-the-art methods and achieves high performance in image segmentation.