Enhanced Feature Fusion from Dual Attention Paths Using Feature Gating Mechanism for Scene Categorization of Aerial Images

Abstract

The widespread use of remote sensing (RS) aerial images in scene categorization has drawn the attention of numerous researchers. As the consequence of this advancement, the approaches for RS image scene categorization that use convolutional neural networks (CNNs), have progressed enormously. Majority of the layouts in the works already published only took the account of scene's global information. However, a lot of replicated areas in the global information reduce the effectiveness of classification and disregard local contextual details, which reflect finer spatial characteristics of local objects. Additionally, the majority of CNN-based approaches give the identical weights on each feature vector, which prevents the mode from differentiating the important features. To overcome the above limitations, we propose a model called enhanced feature fusion from dual attention paths using feature gating mechanism (EFFDA-FG). This model employs Resnext50 baseline model, with channel spatial attention for deriving global and local features along with feature fusion using a gating mechanism. Our suggested model was implemented on three well-known datasets and it proves to give accuracy equal to or higher than that of many state-of-art CNN architectures while using comparatively lesser parameters.