The Impact of Filter Size and Number of Filters on Classification Accuracy in CNN
Wafaa Shihab Ahmed, Abdul Amir Abdullah Karim
Abstract
Convolution Neural Networks (CNNs) have received considerable attention due to their ability to learn directly from data classification features. CNNs used for human motion classification, where predefined and fixed convolutional filter size used. In this paper, different sizes and numbers of filters were used with CNN to determine their effect on accuracy of human motion classification. This work has been done through series of experiments; in each experiment, different filter size and number of filters have been applied. The best performance has been obtained when using 4 convolution layers and 2 pooling layers, whereas has been used the large filter size with upper convolution layer and with each layer the size of filter decreased and number of filters increased, so that, the maximum value of the accuracy classification was 98.98%.