Using AAEHS-Net as an Attention-Based Auxiliary Extraction and Hybrid Subsampled Network for Semantic Segmentation

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Deep learning has advanced semantic segmentation, but struggles with shallow feature extraction, causing inaccuracies.
• Existing methods often neglect shallow features, limiting segmentation precision, especially at object edges.

Purpose of the Study:

• To propose an attention-based auxiliary extraction and hybrid subsampled network (AAEHS-Net) to improve semantic segmentation accuracy.
• To enhance the extraction of both deep and shallow image features for better segmentation results.

Main Methods:

• Introduced a complementary and enhanced extraction module (CEEM) for improved deep and shallow feature extraction.
• Implemented a hybrid subsampled module (HSM) to minimize feature loss during processing.
• Designed a global max pool and global avg pool module (GAGM) as an attention mechanism to integrate global and salient feature information.

Main Results:

• AAEHS-Net achieved higher accuracy than U-Net on three datasets: 90.12%, 96.23%, and 95.15%.
• Demonstrated improved edge segmentation and feature continuity compared to baseline models.
• Outperformed popular algorithms across all evaluation metrics on aerial drone, Massachusetts roads, and Massachusetts buildings datasets.

Conclusions:

• The AAEHS-Net effectively addresses the limitations of shallow feature neglect in deep learning-based semantic segmentation.
• The proposed network modules (CEEM, HSM, GAGM) contribute to enhanced feature representation and segmentation accuracy.
• AAEHS-Net represents a significant advancement in semantic segmentation for various image analysis applications.