基於超像素並運用全卷積網路及識別特徵的多層影像切割演算法

Abstract

在電腦視覺這個領域裡，影像分割是一個研究許多年的重要課題。影像分割指的是將影像細分為多個影像物件的過程，影像物件內的相素會具備某些相同的特徵，因此在本次的研究中，我們使用到多種影像識別特徵，包含：顏色、梯度、材質、亮度、銳度及顯著程度等資訊。 我們提出一個多層的超像素群聚演算法，將具備相似特徵值的超像素進行合併，第一階段是由RGB色彩空間、超像素邊界圖及邊緣偵測圖訓練全卷積網路模型，並由此模型決定相鄰的兩個超像素，其邊界是否該被保留；第二階段則透過評分法及支持向量機決定剩餘的超像素是否需再進一步合併，其中考量的條件包含Lab色彩空間、銳度、顯著圖、紋理感知基元等。整體而言，從模擬及評比數據上，我們的方法皆呈現高度影像分割正確性。

In the field of computer vision, image segmentation is an important task which has been explored for many years. Image segmentation is a process to partition an image into multiple segments. Pixels within the same segments share certain characteristics. Therefore, in this thesis, discriminative features are considered, including color, gradient, texture, brightness, sharpness, saliency and etc. We propose a multi-stage superpixel-clustering algorithm to merge superpixels of similar characteristics. In the first stage, the fully convolutional network is applied to decide whether the boundary of two adjacent superpixels should be kept or not. The model is trained basing on color, superpixel boundary, and edge of the image. In the second stage, scoring method and SVM classification model are used to further decide whether the rest superpixels should be merged or not. We consider up to 14 factors to further improve the performance. Overall, simulations and evaluation metrics show that our algorithm has highly accurate segmentation results.