Abstract: Convolutional Neural Networks (CNN’s) are widely being used for various image processing applications such as denoising, classification, de-hazing and super-resolution. In this paper, CNN image classifier to classify the digits is designed with the convolutional units, batch normalization units and the rectified linear units. The classification accuracy variation by changing different CNN parameters such as learning rate, convolutional filters, convolutional layers and training images is being analyzed. The accuracy saturates or degrades with the increment in number of convolutional layers. Selection of number of filters and learning rate are important hyper parameters impacting classifier accuracy.