Enhancing Deep Training of Image Landmarking with Image CAPTCHA
Osama Hosam, Nashwa Abousamra · 2022
Image landmark classification is the ability to define the geolocation of an image using a landmark inside it. It is needed in many applications such as storing social images and image sharing. The landmark classification has exceptionally low accuracy in literature. Without proper dataset augmentation, classification comes with about 20% accuracy. When using proper dataset augmentation and weight initialization, the accuracy is increased to about 44%. This low accuracy occurs because of the nature of the problem. In this paper we will introduce a new methodology to overcome the accuracy problem. Image CAPTCHA will be used to add more correctly classified images to the training dataset. A few experiments using the proposed system show the ability of the new model to overcome overfitting. The accuracy is increased to about 70%. In addition, the performance of the model during training and validation phases is greatly improved.