Advancing semantic segmentation: Enhanced UNet algorithm with attention mechanism and deformable convolution

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Semantic segmentation is crucial for understanding image content.
• Existing UNet architectures can be enhanced for improved feature representation and contextual information capture.

Purpose of the Study:

• To develop a novel semantic segmentation method using an enhanced UNet architecture.
• To investigate the impact of attention mechanisms in both the encoder and decoder paths of the UNet.

Main Methods:

• Utilized an enhanced UNet with a ResNet50 backbone, incorporating deformable convolution.
• Implemented ECA-ASPP (Attention Spatial Pyramid Pooling) in the encoder path.
• Explored Efficient Channel Attention (ECA) and Large Kernel Attention (LKA) in the decoder path after feature concatenation.

Main Results:

• Attention mechanisms after feature concatenation in the decoder improved segmentation accuracy.
• The Large Kernel Attention (LKA) module demonstrated superior performance over the Efficient Channel Attention (ECA) module.
• Achieved state-of-the-art results on Stanford (mIoU 85.79) and Cityscapes (mIoU 82.25) datasets.

Conclusions:

• The proposed enhanced UNet with LKA attention achieves state-of-the-art semantic segmentation performance.
• Attention mechanisms, particularly LKA, are vital for capturing multi-scale context and improving segmentation accuracy.
• The method shows strong generalization across diverse datasets.