Advancing Textile Damage Segmentation: A Novel RGBT Dataset and Thermal Frequency Normalization.

Area of Science:

• Computer Vision
• Materials Science
• Image Processing

Background:

• RGB-Thermal (RGBT) semantic segmentation is crucial for high dynamic range scene analysis.
• Thermal imaging offers enhanced close-range feature extraction, vital for applications like textile damage detection.

Purpose of the Study:

• Introduce RGBT-Textile, a specialized dataset for close-range textile and damage segmentation.
• Present ThermoFreq, a novel thermal frequency normalization method to mitigate temperature noise in RGBT segmentation tasks.

Main Methods:

• Developed a meticulous data collection protocol, software tools, and labeling process with textile scientists.
• Created the RGBT-Textile dataset for close-range textile and damage segmentation.
• Introduced and applied the ThermoFreq method for thermal noise reduction.

Main Results:

• Evaluated RGBT-Textile and six other RGBT datasets using state-of-the-art (SOTA) models.
• Demonstrated superior performance of SOTA models when utilizing the ThermoFreq method.
• Confirmed ThermoFreq's effectiveness in addressing noise challenges across diverse environmental conditions.

Conclusions:

• The RGBT-Textile dataset and ThermoFreq method significantly advance close-range RGBT semantic segmentation for textile applications.
• ThermoFreq effectively reduces noise, improving segmentation accuracy in challenging thermal imaging scenarios.
• Public release of the dataset aims to foster further research and collaboration in the field.