|
ABSTRACT Several transfer learning methods based on deep Convolutional Neural Networks （CNNs） have been developed so far but obtaining satisfactory classification accuracy still remains a challenging task. However, effective classification performance mainly depends on the feature set of input data to be classified. During transfer learning, features are extracted from multiple layers or multiple models of pre-trained convolutional neural networks to form feature sets of input data, which may contain some redundant features thus affecting the classification results. To tackle this issue, a feature selection method must be applied to transfer features before being fed to the classifier. Hence, a feature selection method is proposed using the idea of mutual information theory to remove redundant features for effective transfer of the learning task. The proposed feature selection method is the combination of filter and embedded approaches that are cascaded one after another to form the final feature set for classification. The proposed method proves its effectiveness when compared to some existing methods on three aerial scene datasets.
|
