Transfer Learning Based Flat Tire Detection by Using RGB Images

Traffic accidents have caused many casualties through the years. Most of the time those accidents are driver-related but there are also many non-negligible vehicle-related ones. A report indicates that 35% of vehicle-related accidents are directly related to tires. This emphasizes the importance of inspecting tires carefully to avoid undesired situations. Many manufacturers have been using computers for automated inspection for a long time. Computer-based inspections of tires are also common practice. The majority of tire fault detection research in the literature focuses on manufacturing line inspection and employs X-ray imaging. Tire inspection must be ongoing till the tire is changed due to a necessity at some point. In most countries, however, it is part of the mandatory periodic vehicle inspection. An automated tire inspection system, which uses RGB images to detect flaws on tires and is mounted at inspection stations, has the potential to reduce inspection time. To begin, we updated the dataset using various image manipulation techniques in order to improve generalization during the training phase. Second, we utilized conventional CNN, DenseNet201, InceptionResNetV2, and Xception in this work to identify flat tires in an RGB image data set that included flat tires, normal tires, and non-tire situations. As a result we obtain 80%, 98%, 97%, and 100% F1-Score from CNN, DenseNet201, InceptionResNetV2, and Xception respectively.