Classification of national flags using convolutional neural network (CNN)

Abstract

Various countries are recognized for their unique symbols and a combination of special colors. Flag classification using image processing is a challenging task due to a variety of colors and textures. Sometimes, the colors and layouts of some flags are very similar. Therefore, we build a custom dataset which is a collection of flag images for 110 countries. Furthermore, we have applied different augmentation techniques to increase the number of images in our dataset. The augmentations are 10° rotation, 20° rotation, 30° rotation, 40° rotation, 50° rotation, 60° rotation, 70° rotation, 80° rotation, 90° rotation, adding ‘salt & pepper’ noise, ‘Gaussian’ noise and translation (20, 20). Each class contains 13 images after applying image augmentation. In total, our custom dataset contains 1430 images. We have proposed a flag classification technique based on Convolutional Neural Network (CNN). The proposed CNN contains 19 layers. The proposed model's validation accuracy and testing accuracy are 98.18% and 97.58%, respectively. We have calculated class wise accuracy. It has been observed that 97 classes out of 110 classes have been detected with 100% accuracy, whereas the previous classification approaches of flag using Gray-Level Co-Occurrence Matrix (GLCM) features (Contrast, Correlation, Energy, Homogeneity) and Support Vector Machine(SVM) classifier showed 19.70% overall accuracy. The Discriminant analysis classifier with the same GLCM features showed 32.73% accuracy. The Decision Tree classifier showed 32.42% accuracy, and the Naïve Bayes classifier showed 32.42% accuracy with the same GLCM features.