Textile Retrieval Based on Image Content from CDC and Webcam Cameras in Indoor Environments

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Textile-based image retrieval is valuable for identifying related indoor scenes, particularly for law enforcement evidence matching.
• Existing methods may lack efficiency or accuracy in complex indoor environments.

Purpose of the Study:

• To develop a novel pipeline for searching and retrieving textile images within indoor scenes.
• To establish a robust method for matching textile regions using feature descriptors and a new dataset.

Main Methods:

• Utilizing MSER on high-pass filtered RGB, HSV, and Hue channels to detect textile regions.
• Employing a combination of Histogram of Oriented Gradients (HOG) and Color Histogram (HCLOSIB) with correlation distance for feature description and matching.
• Introducing the TextilTube dataset with 1913 labeled textile regions across 67 classes.

Main Results:

• The proposed pipeline achieved 84.94% success in 40 nearest neighbor coincidences.
• A precision of 37.44% was obtained for the first coincidence, surpassing current deep learning methods.
• Experimental validation confirms the pipeline's effectiveness for indoor textile image retrieval.

Conclusions:

• The developed pipeline offers an effective solution for textile-based image retrieval in indoor settings.
• The combination of MSER, HOG, HCLOSIB, and the TextilTube dataset provides a strong foundation for future research and applications.