Data Augmentation Based on Multiscale Radon Transform for Seven Segment Display Recognition

Abstract

To alleviate the problem of limited data in creating rotation, scale, perspective, and illumination invariant of the neural network training sets, the multiscale Radon transform is proposed in this study to enhance the data augmentation for seven segment display recognition. Resizing, smoothing, and coefficient shifting generate the desired invariant effects for the training model. The accuracy rates from the experiment demonstrate the superiority of the proposed method over other data augmentation techniques with the best overall accuracy performance of 87.05% - outperforming other data augmentation techniques by 6-13%. The convolutional neural network model generated from the proposed multiscale Radon transform data augmentation is suitable for seven segment display recognition and could become beneficial to other type of self-luminous type of images.