A Novel Fused Random Pooling Method for Convolutional Neural Network to Improve Image Classification Accuracy

Abstract

A convolutional neural network (CNN) is an artificial neural network type used in image recognition and processing. The layers of CNN comprise of an input layer, an output layer, and a hidden layer that contains multiple convolutional layers, pooling layers, and fully connected layers. However, CNN faces issues such as overfitting, which occurs when a model learns the detail and noise in training data to the extent that it negatively affects the model's performance on new data. In this study, a new technique of pooling called fused random pooling replaces the deterministic pooling of CNNs based on a random approach in choosing activation to create better pooled feature maps, thereby decreasing overfitting by performing regularization technique and hyper-parameter tuning by defining the network's best parameters such as the number of layers, filter size, and optimization algorithms. Decrease in the training error and increase in accuracy have been realized in the CIFAR-10, CIFAR-100, and Street View House Numbers (SVHN) datasets as compared to studies in various CNN pooling methods. In the CIFAR-10, the test error and accuracy attained 4.11 percent and 92.37 percent, respectively. In CIFAR-100, a 17.76 percent training error and 69.19 percent accuracy was realized. In the SVHN dataset, fused random pooling achieved a training error of 3.10 percent and accuracy of 96.90 percent. Therefore, fused random pooling proved to be efficient in terms of the decrease in training error and increase in accuracy.