Middle-shallow feature aggregation in multimodality for face anti-spoofing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current face anti-spoofing algorithms often overlook shallow features, leading to a lack of fine-grained information and reduced model accuracy and robustness.
• Advanced methods rely on complex deep network characteristics, potentially missing crucial subtle details present in initial feature layers.

Purpose of the Study:

• To improve face anti-spoofing performance by integrating shallow network features for enhanced fine-grained detail.
• To develop a more accurate and robust algorithm for distinguishing live faces from deceptive attempts.

Main Methods:

• Shallow features are integrated into middle layers using a "shortcut" structure to preserve detailed information.
• Network initialization uses pre-trained parameters from unbalanced samples, followed by training on balanced samples.
• Depthwise separable convolution-based RS Blocks replace traditional residual modules, significantly reducing model parameters and computations.

Main Results:

• The proposed algorithm achieves an average classification error rate (ACER) of 0.0008 on the CASIA-SURF dataset.
• Achieved a True Positive Rate (TPR) of 0.9990 at a False Positive Rate (FPR) of 10^-4.
• Demonstrated superior performance compared to existing face anti-spoofing methods.

Conclusions:

• Integrating shallow features significantly enhances the fine-grained detail representation in face anti-spoofing models.
• The novel approach improves both the accuracy and robustness of liveness detection.
• The optimized network architecture offers substantial reductions in model size and computational cost while maintaining high performance.