| Abstrakt: |
Deep learning with convolutional neural networks (CNNs) is very powerful for image analysis. Of late, deep architectures in machine learning and artificial intelligence (AI) have attracted industry and academia significantly. CNNs could produce impressive results across diversified set of problems in different fields. From the literature, CNNs are found to be ideal candidates for learning of data that is represented in images. Similarly, they are used to exploit the content of Synthetic Aperture Radar (SAR) images with their fundamental building blocks known as convolutions. In this paper, CNN based architecture is implemented for SAR image classification. Multiple convolutional layers are used to form a deep neural network and extract features at different levels. The novelty of the research is to have efficient multi-labelling. Multiple convolutional layers are used to form a deep neural network and extract features at different levels. Unlike many conventional machine learning algorithms, the proposed method learns multiple discriminative features from the training set without the need for prior specification from human expert. This property is efficiently used with a classifier that can learn from previously unknown trends. Besides, it finds latent dependencies between labels and use this knowhow while performing multi-label classification. Thus it can improve performance of SAR image retrieval as part of Content Based Image Retrieval (CBIR) system. SAR data is collected from Amazon Basin dataset which is available to public. Performance of deep learning models such as ResNet-50, VGG-16 and VGG-19 are evaluated. ResNet-50 showed better performance over VGG-16 and VGG-19. Unlike many conventional machine learning algorithms, the proposed method learns multiple discriminative features from the training set without the need for prior specification from human expert. This property is efficiently used with a classifier that can learn from previously unknown trends. Besides, it finds latent dependencies between labels and use this knowhow while performing multi-label classification. Thus it can improve performance of SAR image retrieval as part of CBIR system. With deep learning, the proposed architecture is found to be efficient than many competitive methods found in the literature. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink