Since the ImageNet Large Scale Visual Recognition Challenge has been run annually from 2010 to present, researchers have designed lots of brilliant deep convolutional neural networks(D-CNNs). However, most of the existing deep convolutional neural networks are trained with large datasets. It is rare for small datasets to take advantage of deep convolutional neural networks because of overfitting when implementing those models. In this report, I propose a modified deep neural network and use this model to fit a small size dataset. The goal of my work is to show that a proper modified very deep model pre-trained on ImageNet for image classification can be used to fit very small dataset without severe overfitting.