Locating Target Regions for Image Retrieval in an Unsupervised Manner

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Supervised learning for image retrieval requires extensive annotated data, which is costly and impractical.
• Accurate localization of target regions is crucial for effective image retrieval.
• Unsupervised methods for precise target region localization remain a significant challenge.

Purpose of the Study:

• To propose a novel unsupervised image retrieval method, unsupervised target region localization (UTRL) descriptors, for precise target localization without manual annotation.
• To enhance the localization capabilities of pre-trained Convolutional Neural Network (CNN) models in an unsupervised setting.
• To improve the overall performance of image retrieval systems by enabling accurate feature extraction and dimensionality reduction.

Main Methods:

• Developed a zero-label transfer learning approach to improve co-localization of target regions using pre-trained CNNs.
• Implemented a multiscale attention accumulation method with local Gaussian weights for extracting distinctive target features.
• Introduced twice-PCA-whitening (TPW) for effective vector dimensionality reduction, minimizing performance degradation.

Main Results:

• The proposed UTRL descriptors accurately locate target regions without requiring any supervisory information.
• The multiscale attention mechanism effectively extracts distinguishable features, enhancing localization precision.
• The TPW method robustly reduces feature dimensionality, preserving retrieval performance.
• Achieved approximately 7% higher mean average precision (mAP) compared to state-of-the-art unsupervised methods across six benchmark datasets.

Conclusions:

• The UTRL descriptor method offers a powerful unsupervised solution for accurate target region localization in image retrieval.
• The proposed techniques, particularly zero-label transfer learning and TPW, significantly advance unsupervised image retrieval capabilities.
• This work paves the way for developing efficient image retrieval systems utilizing short vector features and unsupervised learning.