Feature fusion network based on strip pooling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Contextual information is crucial for accurate semantic segmentation.
• Existing self-attention methods capture context but are computationally expensive.
• There is a need for efficient contextual information extraction in semantic segmentation.

Purpose of the Study:

• To develop an efficient self-attention network (FFANet) for semantic segmentation.
• To reduce the computational cost associated with capturing contextual information.
• To improve the performance of semantic segmentation models.

Main Methods:

• Introduced FFANet, a novel self-attention network.
• Utilized strip pooling and linear layers to decrease computational load.
• Proposed a feature fusion (FF) module to compute an affinity matrix.
• Calculated pixel relationships using the affinity matrix to re-weight feature maps.

Main Results:

• Achieved Mean Intersection over Union (IoU) scores of 74.5% on PASCAL VOC2012, 70.3% on CityScapes, and 63.9% on DLRSD.
• Demonstrated superior performance compared to existing semantic segmentation algorithms.
• FFANet effectively captures contextual information and enhances segmentation accuracy.

Conclusions:

• FFANet offers an efficient and effective solution for semantic segmentation.
• The proposed feature fusion module and computational reduction strategies are key to its success.
• The method shows significant potential for various applications requiring precise image segmentation.