Effect of Number of Filters on Convolutional Neural Networks’ Performance

Convolutional Neural Networks are one of the most successful methods used in image classification. In convolutional neural networks, the image is processed by passing it through successive layers of convolution and pooling. For feature extraction, the convolution process is applied to the image using filters in convolution layers. In this study, the effect of the number of filters in the VGG16 model on classification performance was examined. CIFAR10 and Fashion-MNIST datasets were used in the experiments which shown that increasing the number of filters, especially in the first layers, increased the classification accuracy. In the experiments performed with the CIFAR10 dataset the accuracy is increased from 85.20% to 88.58 when the number of filters is increased from minimum to maximum. The improvement with Fashion- MNIST database is from 91.74% to % 92.62, with all the remaining parameters the same. Depending on the increasing number of filters, it is seen that the number of transactions, and consequently the duration, is increased in training.