Deep CNNs with Robust LBP Guiding Pooling for Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) utilize pooling layers for dimensionality reduction and computational efficiency.
• Traditional pooling layers are susceptible to noise in input images, leading to accumulated noise and degraded network performance.
• Existing methods lack effective noise handling within the pooling mechanism itself.

Purpose of the Study:

• To introduce a novel pooling mechanism, Guided Robust Local Binary Pattern (G-RLBP), designed to mitigate noise impact in CNNs.
• To enhance feature extraction robustness by simultaneously down-sampling feature maps and reducing noise.
• To improve the overall recognition accuracy of CNN models in the presence of image noise.

Main Methods:

• Proposed the G-RLBP pooling layer, integrating Local Binary Pattern (LBP) principles for noise robustness.
• Implemented a weighted averaging approach within a sliding window, assigning weights based on pixel noise probability.
• Evaluated G-RLBP's effectiveness in early layers of CNN architectures to minimize cumulative noise effects.

Main Results:

• The G-RLBP layer significantly reduces the impact of noise on feature maps compared to traditional pooling.
• CNN models incorporating G-RLBP demonstrate improved image recognition rates.
• Performance gains are particularly pronounced in scenarios with high levels of image noise.

Conclusions:

• The G-RLBP pooling mechanism offers a robust solution for noise reduction in CNNs.
• This approach enhances the reliability and accuracy of CNNs for image analysis tasks involving noisy data.
• G-RLBP represents a valuable advancement for deep learning applications requiring high recognition performance under adverse imaging conditions.