AI-driven feature descriptor design using CNNs and transfer learning for enhanced image retrieval accuracy and...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Content-Based Image Retrieval (CBIR) systems require effective feature descriptors.
• Traditional descriptors like SIFT and SURF have limitations in large-scale, variable datasets.
• Need for robust, semantically rich features for modern CBIR.

Purpose of the Study:

• To propose an AI-driven feature descriptor framework for CBIR.
• To leverage Convolutional Neural Networks (CNNs) and transfer learning for improved feature representation.
• To enhance retrieval accuracy, efficiency, and adaptability in CBIR systems.

Main Methods:

• Utilized pre-trained CNN architectures (ResNet, VGG) for feature extraction.
• Integrated Principal Component Analysis (PCA) for dimensionality reduction.
• Employed attention-based pooling to generate compact, semantically rich features.

Main Results:

• Achieved a mean Average Precision (mAP) of 85.9% on benchmark datasets (Oxford5k, Paris6k, Holidays).
• Outperformed traditional handcrafted descriptors and baseline CNN features.
• Compressed high-dimensional features (2048-d to 128-d) using PCA, improving storage and speed.

Conclusions:

• The proposed AI framework offers a superior balance of accuracy, efficiency, and adaptability for CBIR.
• Demonstrates effectiveness for large-scale, real-world CBIR applications.
• Highlights the potential of CNNs and dimensionality reduction in advancing image retrieval technology.