چكيده به لاتين
The Transfer Learning Technique has been adopted in recent years to obtain high precision classification in the field of aerial imagery. In this technique, the descriptor (a pre-trained ConvNet) is employed to acquire high-level features, which can be adapted to the given application by replacing the pre-trained classifier with a network. To achieve a good performance, more data is required if the network is being trained on an arbitrary set of data from scratch. To overcome this issue, in this paper the Neural Architecture Search Network Mobile (NasNet Mobile), trained with ImageNet images, is proposed as the descriptor. Moreover, an MLP network is designed in the form of a classifier. In addition, a new loss function is proposed in this paper to further improve the results by adding a new term to the categorical-crossentropy (CE). In the proposed loss function, in contrast with the conventional CE, the probability difference between the misclassified category and the category correctly classified is utilized to enhance the network focus on complex images. As the results indicate, our proposed method achieves 99.54, 98.10, and 93.56 percent (in terms of Overall Accuracy) on UC Merced land-use, AID, and NWPU-RESISC45 dataset, respectively. This indicates a relative improvement of 0.62%, 1.11%, and 1.75% compared to the best results reported in the literature, and thus sets a new state-of-the-art result. In addition, unlike the networks such as VGG (14 million parameters), ResNet50 (23 million parameters), etc., the proposed descriptor network possess a considerably fewer number of parameters (about 4 million), which improves the implementation at least 3.5 times in terms of memory.
Keywords: Aerial imagery, NASNet Mobile architecture, Transfer Learning strategy, MLP and categorical-crossentropy
