Convolutional neural network performance in image classification with small amounts of data resources

In this research, an analysis on convolutional neural network performance in image classification with small amounts of image data resources was performed. To train a convolutional neural network a sufficiently large number of labeled images is required, however not every task has enough data collected for such network training. During the experiments it was noticed that convolutional networks, which had a high number of nodes in each layer had better performance than networks which had increased layer count. Three overfitting prevention methods that were used: Weight decay, Dropout, DisturbLabel produced results which have shown that regardless of small training image data amount an improved accuracy can be achieved using any of these methods. The best classification accuracy improvement was achieved using weight decay. By combining a convolutional network with large node count in each layer and weight decay overfitting prevention method, a network with an even more increased classification accuracy was trained