MFPI-Net: A Multi-Scale Feature Perception and Interaction Network for Semantic Segmentation of Urban Remote Sensing...

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Complex urban remote sensing images present challenges such as multi-scale object distribution, class similarity, and small object omission.
• Existing semantic segmentation networks struggle to effectively address these challenges, leading to suboptimal performance.

Purpose of the Study:

• To propose MFPI-Net, an advanced encoder-decoder semantic segmentation network tailored for complex urban remote sensing imagery.
• To enhance the performance of semantic segmentation by effectively handling multi-scale objects, class similarities, and small object detection.

Main Methods:

• MFPI-Net integrates a Swin Transformer backbone encoder for global semantic feature extraction.
• It incorporates a diverse dilation rates attention shuffle decoder (DDRASD) for multi-scale contextual awareness and resolution enhancement.
• The network also features a multi-scale convolutional feature enhancement module (MCFEM) for local feature modeling and a cross-path residual fusion module (CPRFM) for improved feature interaction.

Main Results:

• MFPI-Net achieved superior performance on the ISPRS Vaihingen and Potsdam datasets compared to mainstream methods.
• The proposed network attained mean Intersection over Union (mIoU) scores of 82.57% and 88.49% on the respective datasets.
• Experimental results validate the effectiveness of MFPI-Net in improving semantic segmentation for urban remote sensing.

Conclusions:

• MFPI-Net demonstrates significant improvements in semantic segmentation accuracy for complex urban remote sensing images.
• The network's architecture effectively addresses challenges related to multi-scale objects, class similarity, and small object recognition.
• MFPI-Net represents a substantial advancement in the field of remote sensing image analysis.