DCGAN: Deep Convolutional GAN with Attention Module for Remote View Classification

Abstract

In recent times, the development of Deep Learning Techniques for Image Classification has increased. The deep learning module uses an unsupervised learning technique. The supervised learning requires an adequate and outsized dataset with labels to train a machine. This paper proposes a unique Unsupervised Deep Feature Learning Method called Deep Convolutional GAN (DCGAN) with Attention Module for Remote Scene Classification. The Attention module is integrated with DCGAN to optimize the power of feature extraction. To extract the contextual information, a feature fusion architecture is proposed and it is integrated with Discriminator. The proposed module optimizes the discriminator and generator losses. The extracted features are given as input to SVM for the classification. The proposed module DCGAN with Attention module is implemented with publicly available remote sensing scene UC-Merced dataset which have 21 different scene classes and the RSSCN7 dataset which have 7 different scene classes. The experimental results obtained by the proposed model DCGAN with Attention module are better than the state-of-art machines or methods results. The proposed model achieves an accuracy of 91.67% and 84.29 % for the RSSCN7 and UC-Merced datasets respectively.