Geographical landmark recognition: Using CNN-Based approach to improve accuracy

Abstract

Landmarks in a digital image are the visibly distinguishable objects that can be properly classified by human brain. In the field of computer vision, landmark recognition is an important topic since it involves the basic ability of machines to identify objects in an image. Training a machine to identify all the things in an image will involve training it on the most basic landmarks first. But landmark recognition is not an easy task to accomplish. The widely available datasets in the field of landmark recognition are not only very large to adversely impact the training, but they are also highly noisy and complex. Even the best of pictures clicked tend to contain a lot of noise in the processed data. To tackle the problem, we trained a Convolutional Neural Network (CNN) on a publicly available geographical landmarks dataset and found that both, training and test accuracies of CNNs (98% and 73% respectively),were higher than the classical machine learning techniques. Also, it was seen that overfitting of models was lesser in CNNs.