Automated Car Damage Detection: A Novel Approach : Anomaly Detection Models for Car Damage Identification Without Pre-existing Damaged Car Datasets
Ali, Haider (2024)
Ali, Haider
2024
Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-10-14
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202408178151
https://urn.fi/URN:NBN:fi:tuni-202408178151
Tiivistelmä
Car damage detection is a highly demanded task in the real world. Collecting damaged car data is a costly process and might not cover all kinds of damages. Anomaly Detection (AD) neural network models can be used to detect car damage. Most of the models are trained with the low variant industrial dataset, it is hypothesized that they will not work with the car dataset. This claim is answered by creating a car damage dataset modeled similarly to traditional anomaly detection datasets making it testable with traditional state-of-the-art AD networks. Distance histograms with mean squared error and structural similarity measure index are used to show the variance difference in the industrial and car damage datasets. Latent Space representations with dimension reduction algorithms i.e. Prinicipal Component Analysis and t-distributed Stochastic Neighbor Embedding are used to show the feature-wise difference between the traditional and car datasets. State-of-the-art models’ results are used to present their limited performance on the car damage dataset. The dataset is further categorized into multiple categories which is useful for testing which anomaly categories are easier to generalize than other categories. DRAEM[56] is used for further experimentations with a few additions to the existing model to achieve better efficiency.