A Method for Human Facial Image Annotation on Low Power Consumption Autonomous Devices

Area of Science:

• Computer Vision
• Machine Learning
• Embedded Systems

Background:

• Facial feature annotation is crucial for many applications.
• State-of-the-art methods often require significant computational resources, limiting their use in autonomous systems.
• There is a need for efficient algorithms deployable on low-power, resource-constrained devices.

Purpose of the Study:

• To propose and evaluate a novel facial feature annotation classifier for autonomous microcomputer systems.
• To demonstrate comparable accuracy to deep neural networks (DNNs) with significantly reduced computational requirements.
• To ensure portability and ease of implementation on devices with limited RAM and no coprocessors.

Main Methods:

• The proposed method combines a Histogram of Oriented Gradients (HOG) face detector with a set of neural networks.
• The system was trained and evaluated on a large dataset of 200,000 facial images.
• Performance was benchmarked against state-of-the-art deep neural network (DNN) architectures.

Main Results:

• The HOG-based classifier achieved accuracy comparable to state-of-the-art DNNs in facial feature annotation.
• The method demonstrated effective performance on microcomputer systems with limited RAM and no coprocessors.
• The algorithm successfully performed facial image attribute classification on real-time video data from an RGB camera.

Conclusions:

• The developed facial annotation algorithm offers a viable, low-power alternative to computationally intensive DNNs for autonomous systems.
• The portability and ease of implementation across various hardware and operating systems represent a significant contribution.
• The availability of the dataset and source code facilitates reproducibility and further research in edge computing for computer vision.