Semantic Segmentation for Object Extraction in Digital Imagery

  • Rasha Alshehhi

Student thesis: Doctoral Thesis

Abstract

Automated extraction of objects from imagery data is a problem of great interest in several applications such as medical imaging, remote sensing, robotics, etc. This is because of the complexity of imagery structures due to geometric occlusion and occurrence of objects in multiple scales, leading to higher intra-class variations and lower inter-class variations. Object-based image segmentation is an effective method for automated extraction of objects from imagery data. This method is an integral part of Object-Based Image Analysis methodology. It aims to create objects with semantic interpretation by partition of an image into homogeneous and contiguous regions based on spectral (e.g., illumination information) and spatial (e.g., geometric and context information) characteristics. In this thesis, a coherent hierarchical segmentation architecture for automated extraction of objects is proposed. This framework is based on two main methods: feature extraction and feature learning. To extract features, hierarchical graph-based segmentation is used based on merging and splitting of spectral and spatial features, considering contextual information between adjacent segments. To learn features, Convolution Neural Network is designed to encode spectral and spatial features hierarchically from image data. The goal of this thesis is studying the performance of segmentation framework in extraction of objects based on extracting and learning features. In this thesis, we consider three scenarios. First, extraction of a relatively homogeneous foreground from a slightly heterogeneous background. An example is the extraction of blood vessels from retinal images by employing hierarchical graph-based segmentation. Second, the extraction of heterogeneous foreground objects from a highly heterogeneous background objects. An example is extraction of roads from high-resolution satellite images by applying an effective hierarchical graph-based segmentation. Third, the extraction of multiple classes which are heterogeneous with a highly heterogeneous background. An example is extraction of roads and buildings from high-resolution satellite images by employing Convolution Neural Network and graph-based segmentation. The segmentation framework is carried out on different datasets. The experimental results demonstrate that the proposed method is able to correctly extract complete and regular objects with semantic interpretation from sophisticated backgrounds, compared with state-of-the-art methods.
Date of AwardDec 2016
Original languageAmerican English
SupervisorPrashanth Marpu (Supervisor)

Keywords

  • Digital Imaging
  • Image Segmentation
  • Feature Interpretation
  • Feature Learning. Neural Networks

Cite this

'