Unadorned Gabor based Convolutional Neural Network Overrides Transfer Learning Concept
R. Leela Jyothi, Abdul Rahiman M., Anilkumar A. · International Journal of Applied Engineering Research · 2021
The efficiency of Convolutional Neural Networks (CNN) is highly influenced by the size of dataset.To train CNN systems from the scratch, dataset of very large size is essential.Transfer learning concept was introduced to overcome this deficiency of CNN.Even though in recent years transfer learning networks has attained high popularity, the adaption of transfer learning networks to entirely different dataset would produce very low recognition results.In this work a Gabor based CNN network is introduced which works highly efficient compared to transfer learning networks.Another deficiency of CNN is that it is not robust to rotation.Even though the notion of Gabor filter being induced in CNN has been suggested earlier, this work introduces an entirely different and very simple Gabor based CNN which produces high recognition efficiency in dataset of very small size and works invariant to rotation.This work would shed new light in deep Learning research where researchers are forced to focus on and build highly complex CNN network structures.Small subset of MNIST dataset with rotated images is used to learn the proposed architecture.