TGV Upsampling: A Making-Up Operation for Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning has advanced semantic segmentation, but convolution and pooling cause information loss.
• Symmetric network architectures with encoding and decoding stages are common.
• Upsampling operations in the decoding stage aim to mitigate information loss.

Purpose of the Study:

• Analyze and compare existing upsampling operations in neural networks.
• Propose a novel upsampling method to address information loss.
• Improve the preservation of detailed textures and edges in feature maps.

Main Methods:

• Detailed analysis of current upsampling operations.
• Integration of image restoration knowledge.
• Development and implementation of the TGV upsampling algorithm.
• Replacement of existing upsampling layers with the TGV method.

Main Results:

• The TGV upsampling algorithm better preserves detailed textures and edges.
• Models using the TGV method achieved an average accuracy improvement of 1.4-2.3% compared to original models.
• Demonstrated effectiveness in semantic segmentation tasks.

Conclusions:

• The TGV upsampling algorithm is an effective solution for information loss in deep learning.
• This novel approach enhances feature map quality and segmentation accuracy.
• Future work could explore broader applications of the TGV upsampling algorithm.