Mobile Generalized Category Discovery for Road Damage Detection

In recent years, there has been a considerable amount of research conducted on the topic of road damage detection using deep learning techniques, with the aim of supporting the safe driving of mobility vehicles. To date, convolutional neural networks and supervised learning with large datasets have been employed for road damage detection, demonstrating high classification accuracy. However, all images in the dataset must be manually labeled. Moreover, images cannot be classified into classes that are not included in the training dataset. The real world presents a vast array of damage types, and the unlabeled input images encompass both known and unknown classes. This renders the system unsuitable for use in providing safe driving support for mobility. Considering this background, Generalized Category Discovery (GCD) has been put forth as a potential solution, aiming to achieve accurate classification in a comprehensive setting where both known and unknown classes are present in unlabeled data. The purpose of this study is to propose a novel model for GCD with the aim of providing safe driving assistance for mobility. The model is based on the pre-trained MobileViT. To validate its performance, this paper examines its performance on the general CIFAR-10 and CIFAR-100 datasets and on a road damage-specific dataset, and discusses the challenges encountered.