Face Recognition Using the SR-CNN Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Face recognition systems struggle with environmental variations like illumination changes, object rotation, and occlusion.
• These variations lead to imprecise target positioning and reduced recognition accuracy.
• Existing algorithms often lack robustness in complex, real-world scenarios.

Purpose of the Study:

• To propose a novel, robust face recognition algorithm capable of handling complex environmental conditions.
• To enhance the precision and reliability of face matching through multi-feature fusion.
• To optimize computational performance for practical applications.

Main Methods:

• A multi-feature fusion approach combining rotation-invariant texture feature (RITF) vectors, scale-invariant feature transform (SIFT) vectors, and convolution neural networks (CNNs).
• Development of a robust face-matching method named SR-CNN.
• Parallelization of the SR-CNN model using a graphics processing unit (GPU) for accelerated computation.

Main Results:

• The SR-CNN method demonstrated significant improvements in true positive rates on both the Labeled Faces in the Wild (LFW) database (10.97–13.24% increase) and a self-collected database (12.65–15.31% increase).
• GPU parallelization resulted in substantial speedups, with acceleration ratios of 5–6 times for LFW and 6–7 times for the self-collected database compared to CPU.
• The algorithm effectively addresses challenges posed by illumination variations, rotation, and occlusion.

Conclusions:

• The proposed multi-feature fusion SR-CNN algorithm offers a robust and efficient solution for face recognition in complex environments.
• The integration of RITF, SIFT, and CNN, coupled with GPU acceleration, significantly enhances both accuracy and computational performance.
• This approach shows considerable promise for improving the reliability of face recognition systems in diverse real-world applications.