|
ABSTRACT Feature selection plays a vital role in image classification task. With the advent of deep learning, significant efforts have been made in developing deep architectures with the aim of getting relevant features. Moreover, deep architectures require large number of training data and also the training time is very high. Due to scarcity of high-dimensional remote-sensing images, many researchers adopted pre-trained deep architectures as feature extractors. When transferring these large number of features to train the classification layer, the network still overfits, yielding inaccurate classification results. To overcome this issue to some extent, we proposed two feature ranking methods to further refining the features transferred from pre-trained inception architecture for scene classification task. Another important fact is that down sampling large images to the input size of the pre-trained architecture causes loss of some informative pixels, so in our framework, we fed image patches of size equal to the size of the input layer of the pre-trained inception and then the features extracted from all the patches are combined. Finally, we compared our framework with some existing methods and obtained acceptable results.
|
