Low-resolution Face Recognition in the Wild via Selective Knowledge Distillation (1811.09998v2)

Abstract: Typically, the deployment of face recognition models in the wild needs to identify low-resolution faces with extremely low computational cost. To address this problem, a feasible solution is compressing a complex face model to achieve higher speed and lower memory at the cost of minimal performance drop. Inspired by that, this paper proposes a learning approach to recognize low-resolution faces via selective knowledge distillation. In this approach, a two-stream convolutional neural network (CNN) is first initialized to recognize high-resolution faces and resolution-degraded faces with a teacher stream and a student stream, respectively. The teacher stream is represented by a complex CNN for high-accuracy recognition, and the student stream is represented by a much simpler CNN for low-complexity recognition. To avoid significant performance drop at the student stream, we then selectively distil the most informative facial features from the teacher stream by solving a sparse graph optimization problem, which are then used to regularize the fine-tuning process of the student stream. In this way, the student stream is actually trained by simultaneously handling two tasks with limited computational resources: approximating the most informative facial cues via feature regression, and recovering the missing facial cues via low-resolution face classification. Experimental results show that the student stream performs impressively in recognizing low-resolution faces and costs only 0.15MB memory and runs at 418 faces per second on CPU and 9,433 faces per second on GPU.

• The paper introduces a dual-stream model that selectively distills key facial features from a high-resolution teacher to a resource-efficient low-resolution student network.
• The methodology employs sparse graph optimization to enhance feature extraction, achieving speeds of 418 faces/sec on CPU and 9,433 faces/sec on GPU with only 0.15MB memory usage.
• This approach significantly advances practical face recognition, enabling robust AI deployment in mobile, surveillance, and other low-resource environments.

Selective Knowledge Distillation for Low-Resolution Face Recognition

The paper "Low-resolution Face Recognition in the Wild via Selective Knowledge Distillation" by Ge et al. explores a novel approach to tackle the demanding task of recognizing low-resolution faces in environments with limited computational resources. In the context of growing applications on mobile and embedded devices, where both storage and processing power can be limited, achieving efficient and accurate face recognition is paramount.

The authors propose an architecture that integrates two primary components: a high-resolution teacher stream and a low-resolution student stream. The teacher stream employs complex Convolutional Neural Networks (CNNs) providing high-accuracy recognition capabilities. On the other hand, the student stream is designed to be computationally lean, targeting low-resolution face images to meet demands for speed and memory efficiency.

A distinctive aspect of the proposed method is its adoption of Selective Knowledge Distillation—a process wherein only the most salient aspects of facial features extracted by the teacher stream are transferred to the student stream. The approach employs a sparse graph optimization model to selectively distil knowledge, effectively regularizing the student stream’s training process.

The experimental results highlight several impressive outcomes:

• The student stream operates efficiently with minimal memory requirements of just 0.15MB.
• The student stream achieves processing speeds of 418 faces per second on a CPU and 9,433 faces per second on a GPU.

These results underscore the practical significance of the method, especially for deployment in resource-constrained environments. The capacity to distil knowledge selectively mitigates the loss of accuracy typically associated with compressing more robust models.

While the paper focuses heavily on the implications for real-world applications, particularly in scenarios involving mobile devices and surveillance systems, there is also an implicit acknowledgement of broader theoretical advancements in AI:

• The integration of sparse graph optimization represents a critical advancement in the efficient transfer of rich informative features amidst the constraints posed by low-resolution images.
• The paper contributes to a growing body of literature on knowledge distillation that emphasizes the importance of selective feature extraction.

Moving forward, further research may explore refining the distillation process with advanced machine learning techniques or extending the model to incorporate facial attributes such as age, and emotion, which might offer additional discriminative power and insights into complex low-resolution facial recognition challenges. Continued exploration of leveraging recurrent mechanisms for error handling in teacher networks could further enhance the robustness of the student networks when deployed in increasingly varied environments.

The paper by Ge et al. provides an important step toward optimizing face recognition models for operational environments with limited resources, making significant contributions to the ongoing discourse on efficient AI model deployment.