Deep learning, wireless cognitive radio networks, spectrum detection, semantic segmentation
Machine-learning algorithms have attracted much attention in a wide range of areas. Because machine-learning algorithms can extract patterns from data automatically, it is possible to model the input-output function of a system using a machine-learning algorithm. In cognitive radio networks, machine-learning-based spectrum sensing schemes depend on the positioning of the nodes. During building extraction tasks based on high-resolution aerial images, large-scale datasets are required, with customized architecture necessary to process the datasets. Moreover, when there is a domain gap between the training and test data, trained models fail to segment objects for unseen images. In this Theses, the challenges facing machine-learning-based systems and solutions are discussed. The first application is for spectrum sensing in cognitive radio networks. The hidden primary user (PU) problem, however, is a critical issue in cognitive radio networks because spectrum sensing nodes (SNs) can misclassify the degree of spectrum occupancy. To cope with this, machine-learning-based cooperative spectrum sensing schemes (CSSs) have been proposed. CSSs that do not consider node placement, however, continue to be affected by the hidden PU problem. In this paper, we present a method by which to place SNs to guarantee the performance of machine-learning-based CSSs. We verify that the hidden PU problem causes some overlap in the data distribution, which deteriorates of the spectrum sensing ability. Based on Kullback-Leibler divergence, analytical expressions for the spectrum-sensing coverage of a single SN are derived. We then propose a strategy for placing a few SNs to cover the entire area of the PU and prove the feasibility of the proposed method experimentally. The second application is related to deep-learning architecture for semantic segmentation from high-resolution aerial images. Extracting manufactured features such as buildings, roads, and water from aerial images is critical for urban planning, traffic management, and industrial development. Recently, convolutional neural networks (CNNs) have become a popular strategy to use to capture contextual features automatically. We design a multi-object segmentation system and propose an algorithm that utilizes pyramid pooling layers to improve U-Net. Test results indicate that U-Net with pyramid pooling layers, referred to as UNetPPL, learns fine-grained classification maps and outperforms other algorithms, specifically FCN and U-Net, achieving a mean intersection of union (mIOU) value of 79.52 and pixel accuracy of 87.61% for four types of objects (buildings, roads, water, and background). The final application is domain adaptation for building extraction. CNN-based semantic segmentation models garnered much attention in relation to remote sensing and achieved remarkable performance during the extraction of buildings from high -resolution aerial images. However, limited generalization for unseen images remains. When there is a domain gap between the training and test datasets, CNN-based segmentation models that are trained using a training dataset fail to segment buildings in the test dataset. In this paper, we propose domain-adaptive transfer attack (DATA)-based segmentation networks for building extraction from aerial images. The proposed system utilizes jointly both domain transfers and adversarial attacks. Based on the DATA scheme, the distribution of input images can be shifted to that of target images while maintaining the semantic spaces. This reduces the domain gap and expands the generalization of the segmentation model. From test results with two different datasets (i.e., the Inria aerial image labeling dataset and the Massachusetts building dataset), it is verified that the performance when extracting buildings is improved to 0.16% and 7.12%, respectively.
Table Of Contents
Abstract i List of contents iv List of figures vii List of tables viii
1. INTRODUCTION 1 1.1 MACHINE LEARNING ALGORITHM 1 1.2 BACKGROUND AND CHALLENGING ISSUES 1 1.2.1 COGNITIVE RADIO NETWORKS 1 1.2.2 SEMANTIC SEGMENTATION FROM HIGH-RESOLUTION AERIAL IMAGE 2 1.2.3 DOMAIN GAP BETWEEN TRAINING AND TEST DATASET 3 1.3 OVERVIEW OF Theses 3
2. SENSING COVERAGE-BASED SPECTRUM DETECTION 4 2.1 INTRODUCTION 4 2.1.1 RELATED WORK 5 2.1.2 SUMMARY AND ORGANIZATION 7 2.2 SYSTEM MODEL AND PROBLEM FORMULATION 7 2.2.1 SYSTEM MODEL OF COGNITIVE RADIO NETWORKS 7 2.2.2 MACHINE LEARNING-BASED COOPERATIVE SPECTRUM SENSING 9 2.2.3 HIDDEN PU PROBLEM AND KULLBACK-LEIBLER DIVERGENCE 11 2.3 ANALYSIS OF THE MAXIMUM DISTANCE 14 2.3.1 MAXIMUM DISTANCE BETWEEN SN AND PU 14 2.3.2 COVERAGE AREA 15 2.3.3 OUTAGE PROBABILITY FOR THE MAXIMUM DISTANCE 16 2.3.4 POSITIONING OF SPECTRUM SENSING NODES 18 2.4 EXPERIMENT SETUP AND RESULTS 19 2.5 CONCLUSION 24
3. SEMANTIC SEGMENTATION FOR GEO DATA GENERATION 26 3.1 INTRODUCTION 26 3.2 DEEP LEARNING-BASED OBJECT SEGMENTATION SYSTEMS 28 3.2.1 FULLY CONVOLUTIONAL NETWORKS 28 3.2.2 U-NET 29 3.3 U-NET WITH PYRAMID POOLING LAYERS 30 3.4 DATASET AND EXPERIMENTS 31 3.4.1 DATASET 33 3.4.2 TRAINING SETUP 34 3.5 RESULTS AND DISCUSION 36 3.6 CONCLUSION 39
4. DOMAIN ADAPTATION FOR AERIAL IMAGES 41 4.1 INTRODUCTION 41 4.2 DEEP LEARNING-BASED SEMANTIC SEGMENTATION 43 4.2.1 SEMANTIC SEGMENTATION SYSTEMS 44 4.2.2 INRIA AERIAL IMAGE LABELING DATASET 44 4.2.3 TRAINING SETUP 45 4.2.4 TEST RESULTS AND COMPARISION WITH OTHER ARCHITECTURES 47 4.3 DOMAIN GAP AND SEGMENTATION PERFORMANCE 58 4.3.1 MASSACHUSETTS BUILDING DATASET 58 4.4 DOMAIN ADAPTIVE TRANSFER ATTACK 59 4.4.1 OVERVIEW OF THE PROPOSED MODEL 60 4.4.2 OBJECTIVE FUNCTION FOR DOMAIN ADAPTATION 62 4.4.3 OBJECTIVE FUNCTION FOR ADVERSARIAL ATTACK 62 4.4.4 OBJECTIVE FUNCTION FOR DISCRIMINATOR 63 4.4.5 TRAINING ADVERSARIAL ATTACK MODEL & DISCRIMINATOR 64 4.5 DATA-BASED ADVERSARIAL TRAINING & RESULTS 66 4.5.1 ADVERSARY TRAINING SETUP 66 4.5.2 TEST RESULTS FOR ADVERSARY TRAINING 67 4.6 CONCLUSION 70