Design of Encoder in CNN for Effective Recognition of Odia Numerals

Deep learning is highly appreciated in the field of image processing as it follows the human brain like a training process. In general, artists look at a model and implement the structure in the canvas by observing the important features. Similarly, Autoencoders have attracted the attention of researchers as it learns the important features from the trained data to generate the structures similar to the data provided for training. Autoencoders are the basic building components of generative models. In this work, we have designed an Autoencoder (AE) to generate a large number of data to support the Generative Adversarial Network (GAN) Model applied to the IIT Bhubaneswar Odia handwritten numeral database. In this work, we have designed an encoder that generates the feature vectors by applying Convolutional layers activated by Leaky-ReLU followed by max pooling. It is verified that the decoder recognizes the features due to proper training. The generated images are quite similar to original data that validate the proposed AE is well reconstructive. To measure the performance of the model loss is calculated using mean square error. The proposed model of AE is trained with Adam Optimizer.