Deep Fourier Ranking Quantization for Semi-Supervised Image Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Deep Learning

Background:

• Supervised product quantization methods exhibit extreme label dependence.
• Semi-supervised learning uses unlabeled data to regularize deep networks, but existing methods overlook individual instance variances.
• This oversight limits exploration of local neighborhood structures, leading to overfitting.

Purpose of the Study:

• To introduce a Fourier-based perspective to explore semantic relations in unlabeled data for improved semi-supervised learning.
• To address the limitations of existing methods that focus on overall distribution consistency and ignore individual instance variances.
• To enhance the robustness and generalization of deep learning models in semi-supervised settings.

Main Methods:

• A novel Phase Mixing (PM) strategy based on Fourier Transform is proposed to manipulate phase components for controlling semantic information proportion.
• This strategy naturally constructs multi-level similarity neighbors for unlabeled data.
• A ranking quantization loss is formulated to perceive multi-level semantic variances within neighbor instances.

Main Results:

• The proposed method successfully explores semantic relations between unlabeled instances in a self-supervised manner.
• It constructs multi-level similarity neighbors and perceives multi-level semantic variances, improving model robustness and generalization.
• Experiments across three semi-supervised settings demonstrated an average improvement of 3.95% over state-of-the-art methods on four datasets.

Conclusions:

• The Fourier-based approach effectively alleviates label dependence in semi-supervised learning by leveraging unlabeled data more effectively.
• The Phase Mixing strategy and ranking quantization loss enhance model performance by capturing subtle semantic variations.
• This work offers a promising direction for improving deep network regularization and generalization in semi-supervised learning.