A CNN based framework for Translation Invariant Image Classification

The contribution of this paper is to classify the images with transformation present in it. Transformations in the images are translation, rotation, scale and noise. The accuracy of classification decreases because of translations in the images. The proposed framework uses Convolutional Neural Network (CNN) for classification of transformed images. It consists of two convolutional neural networks, first CNN1, it is the standard CNN trained by images formed by data augmentation approach. Then its performance is tested for dataset of transformed images. The second one CNN2 is trained on only translated images. In the first stage, the images are classified using CNN1. And in the second stage, images misclassified by CNN1 are fed to CNN2 for further classification. The proposed framework is evaluated on hand written digits dataset- MNIST and alphabets dataset by NIST. The result shows that framework proposed outperforms single CNN.