A New Pooling Method for Cnn-based Deep Learning Models

Convolutional Neural Network (CNN) methods provide an effective architecture widely used in image classification tasks. The pooling method in CNN layers has a critical role in reducing the computational cost by preserving some information in the feature map. The primary objective of this study is to improve information loss in pooling methods used in the literature and enhance classification accuracy. The Turhan pooling method offers a weighting, balancing, and adjustment capability beyond traditional max-pooling and average-pooling methods. This method allows tuning the parameters of the two features with the highest signal that can generate action potentials in the pooling mechanism similar to biological neurons. The method enables to optimize pooling for specific datasets or tasks. The results demonstrate that the Turhan pooling method is effective and competes with different architectures such as CNN, AlexNet, U-Net, and ResNet-18 on the Cifar10 dataset, improving classification performance.