Pooling layer is the most important layer in convolutional neural network (CNN), it is used to down sampling the feature size of the convolutional layer to increase the computation efficiency and reduce the size of the network as well as increase robustness of the parameters .So different research and studies are proposed to design pooling approaches, some of these method are based on deterministic such as max pooling and average pooling and other depended on probability such as sophisticated method. In this paper, we have proposed a new method based on wavelet transform, our methods are relies on exploitation the features of the wavelet transform and find another form of data more suitable depending on the neighbored interpolation, this can reduce the contrast between adjacent element in the pool size, then using another one or average of two of determinists method with wavelet transform to sample data. Using the wavelet without sub sampling to overcome the shortcoming of max pooling to reduce the elimination of important information of details data while enjoying with most important features. Three different database were used to test the proposed method and the experimental results show that our proposed methods performs or outperforms comparative with other method such as max and average pooling.
Volume 12 | Issue 4
Pages: 76-85
DOI: 10.5373/JARDCS/V12I4/20201420