Convolutional neural networks (CNNs) have recently witnessed a notable interest due to their superior performance demonstrated in computer vision applications; including image retrieval. This paper introduces an optimized bilinear-CNN architecture applied in the context of remote sensing image retrieval, which investigates the capability of deep neural networks in learning transfer from general data to domain-specific application, i.e. remote sensing image retrieval. The proposed deep learning model involves two parallel feature extractors to formulate image representations from local patches at deep convolutional layers. The extracted features are approximated into low-dimensional features by a polynomial kernel projection. Each single geographic image is represented by a discriminating compact descriptor using a modified compact pooling scheme followed by feature normalization. An end-to-end deep learning is performed to generate the final fine-tuned network model. The model performance is evaluated on the standard UCMerced land-use/land-cover (LULC) dataset with high-resolution aerial imagery. The conducted experiments on the proposed model show high performance in extracting and learning complex image features, which affirms the superiority of deep bilinear features in the context of remote sensing image retrieval.
|Number of pages||8|
|Journal||IAENG International Journal of Computer Science|
|Early online date||20 Nov 2019|
|Publication status||Published - 30 Nov 2019|
- Deep learning
- Image retrieval
- Remote sensing