With the advent of personal computers, humans have always wanted to communicate with them in either their natural language or by using gestures. This gave birth to the field of Human Computer Interaction and its subfield Automatic Sign Language Recognition. This paper proposes the method of automatic feature extraction of the images of hand. These extracted features are then used to train the Softmax classifier to classify them into 20 classes. Five stacked Denoising Sparse Autoencoders (DSAE) trained in unsupervised fashion are used to extract features from image. The proposed architecture is trained and tested on a standard dataset [1] which was extended by adding random jitters such as rotation and Gaussian noise. The performance of the proposed architecture is 83% which is better than shallow Neural Network trained on manual hand-engineered features called Principal Components which is used as a benchmark.