A New Pooling Approach Based on Zeckendorf's Theorem for Texture Transfer Information.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Pooling layers are crucial for invariance in convolutional neural networks (CNNs).
• Traditional pooling methods like max pooling have limitations in handling data variations and geometric arrangements.
• Image segmentation tasks require robust feature extraction and invariance to transformations.

Purpose of the Study:

• To propose and evaluate a new pooling method, Z pooling, based on Zeckendorf's number series.
• To demonstrate the advantages of Z pooling over traditional pooling functions for image segmentation tasks.
• To explore the parameterless receptive field expansion and rotation tolerance properties of Z pooling.

Main Methods:

• Replacing standard maximum pooling layers with Z pooling layers in deep learning architectures.
• Utilizing Zeckendorf's number series as the basis for the Z pooling mechanism.
• Evaluating the proposed method on traditional image segmentation and dense labeling tasks.

Main Results:

• Z pooling exhibits superior performance in image segmentation tasks compared to other pooling functions.
• The Z pooling layer enhances rotation tolerance and parameterlessly increases the receptive field.
• Different Z pooling combinations can emphasize image frequencies and extract ultrametric contours.

Conclusions:

• Z pooling offers a promising alternative to traditional pooling methods in CNNs, particularly for segmentation.
• The method's independence from geometric arrangements provides significant advantages for image analysis.
• Z pooling's unique properties enable advanced image processing capabilities like frequency emphasis and contour extraction.