The CNNs-based model has been proven to achieve impressive performance on a wide range of classification tasks. However, the convoluted results will only retain the local features and discard the global information when using max-pooling is performed with decreasing resolutions. some features of similar data are always diluted during several convolutions, so the decision will be more difficult after max pooling. In this work, we propose a novel pooling layer called Chebyshev Pooling. It makes use of Chebyshev's inequality to produce results about the probability distributions within the kernel which contains the functions of maximum and average pooling. In addition, the proposed layer can ensure that its output is in the range of (0. 0, 1. 0), which is more stable for subsequent processing. Experiments illustrate that our proposed pooling layer can improve the classification performance of various data sets. Moreover, the design and implementation can be easily deployed in some type of CNNs-based classification systems.