Aerial scene understanding in the wild: Multi-scene recognition via prototype-based memory networks.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Aerial scene recognition is crucial for understanding Earth's surface.
• Current methods primarily focus on single-label classification, limiting real-world applicability.
• Identifying multiple scenes in a single aerial image presents a significant challenge.

Purpose of the Study:

• To develop a method for multi-scene recognition in single aerial images.
• To address the time and labor costs associated with manual annotation for this task.
• To leverage large datasets of single-scene images for multi-scene recognition.

Main Methods:

• A prototype-based memory network is proposed.
• Key components include a prototype learning module and an external memory.
• A multi-head attention-based memory retrieval module is used for scene prototype identification.

Main Results:

• The network effectively recognizes multiple scenes in single aerial images.
• It significantly reduces the need for annotated multi-scene data during training.
• Experimental results validate the proposed network's effectiveness.

Conclusions:

• The developed prototype-based memory network offers an efficient solution for multi-scene aerial image recognition.
• The approach alleviates the burden of extensive manual annotation.
• The creation of a new multi-scene aerial image (MAI) dataset facilitates future research.