A plug and play attention block for accurate multi scenario remote sensing image segmentation

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• High-resolution remote sensing images present challenges like object complexity, occlusion, and rich semantic information.
• Existing methods may require large datasets and parameters, limiting their efficiency.

Purpose of the Study:

• To propose SegRSNet, a novel network for effective feature extraction from remote sensing images.
• To address the challenges of object variability, occlusion, and multilevel semantic information in remote sensing data.
• To develop a computationally efficient model that achieves state-of-the-art performance.

Main Methods:

• Designed a plug-and-play SegRS Block incorporating channel feature aggregation, spatial location capture, cross-layer feature fusion, and channel dimension modeling.
• Utilized a combination of convolutional neural networks and a specially designed attention module.
• Evaluated SegRSNet on multiple benchmark datasets for building and road feature extraction.

Main Results:

• SegRSNet achieved state-of-the-art (SOTA) performance on benchmark datasets for building and road feature extraction.
• The proposed method outperforms existing approaches with similar parameter scales.
• Demonstrated high adaptability and accuracy across various remote sensing image analysis tasks.

Conclusions:

• SegRSNet offers an effective solution for remote sensing image segmentation, particularly for feature extraction tasks.
• The combination of CNNs and attention mechanisms reduces training costs while enhancing the receptive field.
• The model provides detailed and accurate recognition and parsing of multi-scenario remote sensing data.