A Multiple-Instance Densely-Connected ConvNet for Aerial Scene Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Aerial scenes present unique challenges for image classification due to crowded objects and the need for local semantic understanding.
• Existing Convolutional Neural Networks (ConvNets) often lose low- and mid-level features in deeper layers, hindering performance on aerial imagery.
• A gap exists in effectively analyzing local semantics within aerial scenes for accurate classification.

Purpose of the Study:

• To propose a novel Multiple-Instance Densely-connected ConvNet (MIDC-Net) for improved aerial scene classification.
• To address the limitations of existing ConvNets in preserving local features and semantics for aerial imagery.
• To develop a model that leverages multiple-instance learning for enhanced analysis of local details in aerial scenes.

Main Methods:

• Aerial scene classification framed as a multiple-instance learning problem.
• Proposed a simplified dense connection structure within an instance-level classifier to preserve multi-level features.
• Introduced a trainable attention-based multiple instance pooling layer for highlighting relevant local semantics.
• Implemented a bag-level classification layer supervised by bag labels for direct learning.

Main Results:

• The MIDC-Net achieved superior performance on three benchmark aerial scene datasets.
• The proposed method significantly outperformed state-of-the-art methods in aerial scene classification accuracy.
• MIDC-Net demonstrated efficiency by requiring substantially fewer parameters compared to existing approaches.

Conclusions:

• The MIDC-Net effectively captures local semantics crucial for aerial scene classification.
• Multiple-instance learning combined with dense connections offers a powerful framework for analyzing complex aerial imagery.
• The proposed method provides a more parameter-efficient and accurate solution for aerial scene classification tasks.