Recurrently exploring class-wise attention in a hybrid convolutional and bidirectional LSTM network for multi-label...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Aerial image classification is crucial for remote sensing.
• Existing methods often focus on single-label classification, neglecting real-world multi-label scenarios.
• Understanding object co-occurrence in aerial images is vital for in-depth analysis.

Purpose of the Study:

• To address the limitation of underexplored class dependency in existing aerial image multi-label classification methods.
• To propose a novel end-to-end network for improved multi-label classification.
• To enhance the comprehensive understanding of high-resolution aerial images.

Main Methods:

• Developed a class-wise attention-based convolutional and bidirectional LSTM network (CA-Conv-BiLSTM).
• Employed a feature extraction module for fine-grained semantic features.
• Utilized a class attention learning layer for discriminative class-specific features.
• Incorporated a bidirectional LSTM sub-network to model class dependency.

Main Results:

• The proposed CA-Conv-BiLSTM network effectively extracts features and models class dependencies.
• Experimental results on UCM and DFC15 multi-label datasets demonstrate significant improvements.
• The model achieved strong quantitative and qualitative validation for aerial image multi-label classification.

Conclusions:

• The CA-Conv-BiLSTM network offers a robust solution for aerial image multi-label classification.
• Modeling class dependency is essential for accurate and comprehensive aerial image analysis.
• This approach advances the field of remote sensing by providing more insightful object-level labeling.