Object Extraction from Remotely Sensed Aerial Images

Abstract

The topographical map of the Earth is recorded by capturing high spatial resolution (HSR) aerial images from higher altitudes using Aircraft, Helicopters, and Unmanned Aerial Vehicles (UAVs). The object extraction from HSR remotely sensed aerial imagery data is a prerequisite in a large number of applications such as planning urban cities, accessing disasters, managing traffic congestion, and providing up-to-date road maps. The development of automated methods to extract objects accurately is highly required for the applications, as mentioned above. However, this is a notoriously challenging task in the field of remote sensing. The deep learning field has gained massive interest due to its ability to learn after the availability of high-volume data and computational resources. This thesis investigates evolutionary optimization based framework for quality enhancement of remotely sensed aerial images and various Convolutional Neural Network (CNN) based approaches of deep learning and proposes an enhancement and object segmentation techniques for HSR remotely sensed aerial imagery data. In the first part of the thesis, the contrast enhancement technique to improve the visual quality of remotely sensed aerial images is presented. The visual quality of captured aerial images is impaired due to the atmospheric effects and limitations of sensors. The visual quality of aerial images needs to improve to extract the hidden object details by increasing the pixel intensity ranges. Most of the techniques in the literature do not consider multi-objective function optimization, either computationally complex or less stable. There is a high demand to introduce a framework to find stable optimum values for multi-objective function with reduced computational complexity. The new framework is introduced to restore the visual quality of images by adjusting saturation, color values, and finally enhanced the contrast of the images through Particle Swarm Optimization (PSO). The experimental and visual quality results showed that the proposed framework outperformed over other state-of-the-art quality restoration techniques. Next, in the thesis, two deep learning-based semantic segmentation architectures are introduced to extract diversified objects from aerial images. The deep learning-based approaches achieved better results as compared with conveni tional techniques. The conventional techniques involve manual feature extraction in multiple stages, which demands to maintain the accuracy of each stage to get overall high accuracy. On the other hand, the Convolutional Neural Network (CNN) based architectures in deep learning field provide object segmentation by learning the image features from a higher volume of imagery data. Plentiful of semantic segmentation architectures are present in the literature to perform object segmentation. However, there is still scope to improve accuracy while segmenting small objects in high-resolution aerial images. In aerial images, objects appear irregularly shaped and tiny in size and also present in dominant background scenarios. Further, class-wise pixels and background pixels are in the ratio of ones-to-tens, which leads to class imbalance problem. Therefore it is very challenging to obtain better prediction accuracy and completeness by preserving the connectivity without any gap between successive objects. In this thesis, a Dense Refinement Residual Network (DRR Net) is proposed for road extraction from aerial imagery data. The DRR Net is introduced based on dense convolutions for feature learning, residual connections to guide the learning path, and provides refinement through the stacking of DRR modules in the network. In the final part of the thesis, the robust encoder and decoder architecture, namely O-SegNet for objects segmentation from high-resolution aerial imagery data, is introduced. The O-SegNet provides emphasis to relevant object details and extracts global context through the selfattention mechanism and multi-level pooling. Both of these proposed architectures are trained with composite loss function to focus more on small object instances. The proposed semantic segmentation architectures have achieved significant quantitative and qualitative results compared with the other existing semantic segmentation architectures.