Efficient deep feature selection for remote sensing image recognition with fused deep learning architectures

摘要

Convolutional neural networks (CNNs) have recently emerged as a popular topic for machine learning in various academic and industrial fields. It is often an important problem to obtain a dataset with an appropriate size for CNN training. However, the lack of training data in the case of remote image research leads to poor performance due to the overfitting problem. In addition, the back-propagation algorithm used in CNN training is usually very slow and thus requires tuning different hyper-parameters. In order to overcome these drawbacks, a new approach fully based on machine learning algorithm to learn useful CNN features from Alexnet, VGG16, VGG19, GoogleNet, ResNet and SqueezeNet CNN architectures is proposed in the present study. This method performs a fast and accurate classification suitable for recognition systems. Alexnet, VGG16, VGG19, GoogleNet, ResNet and SqueezeNet pretrained architectures were used as feature extractors. The proposed method obtains features from the last fully connected layers of each architecture and applies the ReliefF feature selection algorithm to obtain efficient features. Then, selected features are given to the support vector machine classifier with the CNN-learned features instead of the FC layers of CNN to obtain excellent results. The effectiveness of the proposed method was tested on the UC-Merced dataset. Experimental results demonstrate that the proposed classification method achieved an accuracy rate of 98.76% and 99.29% in 50% and 80% training experiment, respectively.