Face Video Retrieval Based on the Deep CNN With RBF Loss.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face video retrieval requires discriminative features despite variations in angle, illumination, and expression.
• Existing CNN-based binary hashing and metric learning methods face limitations in information loss and storage inefficiency.

Purpose of the Study:

• To develop a novel framework for extracting compact and discriminative features for face video retrieval.
• To address the limitations of existing methods in terms of information loss and storage efficiency.

Main Methods:

• A novel Radial Basis Function kernel (RBF Loss) based loss function trains neural networks for compact, high-level feature generation.
• Optimized Logistic Quantization converts real-valued features to 1-byte integers with minimal information loss.

Main Results:

• The proposed framework significantly outperforms state-of-the-art feature extraction methods in face video retrieval on the ICT-TV dataset.
• The RBF loss demonstrates effectiveness in image classification and retrieval tasks on CIFAR-10 and Fashion-MNIST datasets.

Conclusions:

• The novel framework offers an effective solution for compact and discriminative feature extraction in face video retrieval.
• The proposed RBF Loss and Logistic Quantization contribute to improved performance and efficiency in retrieval tasks.