LbpRIPooling: A novel rotation-invariant pooling method for CNNs

Abstract

Pooling layers in convolutional neural networks (CNNs) are crucial for reducing computational complexity and facilitating deeper model training by downsampling image/feature maps. While max and average pooling are commonly used for their efficiency, they fail to retain pixel position information, leading to potential information-loss. LbpPooling addresses this by considering pixel neighborhoods and preserving local correlations but generates different values when the input image is rotated, affecting performance. This paper introduces LbpRIPooling, a novel method that provides rotation-invariant values by dynamically assigning weights based on the magnitude of values in the pooling area. LbpRIPooling is evaluated against LbpPooling, max pooling, average pooling, and several advanced pooling techniques across diverse datasets, consistently demonstrating superior performance, particularly on challenging datasets like CIFAR10 and CIFAR100, and Pins Face Recognition. Additionally,these methods are integrated into transfer learning models such as DenseNet121, VGG16, ResNet50, InceptionV3, and the recent transformer-based MaxViT architecture. DenseNet121, with LbpRIPooling, surpasses LbpPooling by 0.78% on CIFAR10 and outperforms the original model by 2.33%. On CIFAR100, LbpRIPooling shows a 9.16% improvement over LbpPooling and a 15.78% improvement over the original model. To demonstrate real-world applicability, LbpRIPooling was tested on the Pins Face dataset, confirming its robustness to rotation in practical scenarios. The method achieved the highest accuracy and F1-score among compared pooling strategies, validating its effectiveness in real-world face recognition tasks.This method's ability to consistently preserve meaningful information highlights its potential for significant advancements in CNN-basedtasks.