Lightweight Facial Hash Retrieval: A Novel Approach Combining Knowledge Distillation and Optimal Transport

Proposing a Facial Hash Retrieval Algorithm Based on Knowledge Distillation and Optimal Transport Technology

Addressing the challenges of excessive model parameters, high computational complexity, and inaccurate fine-grained retrieval inherent in visual neural networks, this article introduces a novel lightweight facial hash retrieval algorithm grounded in the principles of knowledge distillation and optimal transport. The algorithm's key contributions are:

1. Attention-Based Triplet Knowledge Distillation: The algorithm leverages a triplet knowledge distillation framework incorporating an attention mechanism. The loss function is a composite of attention loss, Kullback-Leibler loss, and identity loss. This approach effectively focuses on salient facial regions while concurrently reducing network parameters.

2. Optimal Transport-Based Hash Quantization: The proposed algorithm employs a hash quantization scheme built upon optimal transport technology. Class centers are computed to partition the facial feature space, facilitating binary quantization through optimal transport. This strategy significantly enhances hash retrieval performance.

3. Alternating Training Strategy: The algorithm employs an alternating training strategy to fine-tune the parameters of the lightweight hash network. This approach optimizes the network for efficient and accurate retrieval.

Experimental evaluations on two benchmark facial datasets demonstrate that this lightweight hash retrieval algorithm surpasses the performance of several well-established hash retrieval methods. These findings highlight the algorithm's efficacy in addressing the limitations of traditional approaches while achieving superior retrieval accuracy.