Effect of parameters on classification performance in convolutional neural networks for image classification

Nowadays, great progress has been made in the field of image processing and pattern recognition, and convolutional neural networks play a major role behind the success in this field. Convolutional neural networks designed with inspiration from effectiveness of biological neural networks successfully implement processors such as object recognition, thinking, and feature extraction on image data. However, there are some difficulties encountered in the design and implementation of convolutional neural networks. Optimization of hyperparameters according to the convolutional neural network model, data set and hardware used is one of these difficulties. In this study, the effects of some hyper parameters on classification performance in convolutional neural networks models were investigated using two different datasets. Hyperparameters examined in the study; the number of epochs, the number of neurons, batch size, activation functions, optimization algorithms and learning rate. In the convolutional neural network models thosetake place in the Keras library, applications were made with the NASNetMobile and DenseNet201 models, which showed the best performance as a result of the tests made in the dataset. 65 different training tasks planned with the values of the hyperparameters in different intervals were applied and the results were obtained. As a result of the studies, it was observed that the accuracy rates increased by 6.5% compared to the initial value in the NASNetMobile model and by 11.55% compared to the initial value in the DenseNet201 model.