Identifying Defective IR Cameras through a Machine Learning Approach to Image Artifact Detection

Title: Identifying Defective IR Cameras through a Machine Learning Approach to Image Artifact Detection This report is about methods for automatic detection of subtle image artifacts. Such methods could be a part of the image quality control procedure in the production process of infrared cameras. Currently image quality control requires manual inspection. The main goal of the investigation described in this report was to suggest and evaluate different measures that can be used to detect image artifacts and asses image quality. In many cases these measures are attempts to find structure in what should be random white noise and will assume that test images are acquired against a uniform radiation. Often single-valued measures are sufficient for achieving good detection accuracy, but in some cases more complex, vector valued measures (termed features) in combination with a machine learning classifier are needed. This report conclusively shows that automatic detection of subtle image artifacts with high accuracy is possible and in many cases achievable using single-valued measures. Sammanfattning Titel: Identifiering av defekta IR-kameror med ett lärande system för detektion av bildartefakter Denna rapport handlar om metoder för automatisk detektion av diffusa bildartefakter. Sådana metoder kan bli en del av den bildkvalitetskontroll som utförs vid produktion av värmekameror. I nuläget krävs manuell granskning av varje kamera. Den utredning som beskrivs har fokuserats på att föreslå och utvärdera olika mått som kan användas för att detektera artefakter i bilder och för att mäta bildkvalité. I många fall är dessa mått försök att finna struktur i bilder som borde bestå av enbart vitt brus och kräver att testbilden tas mot en likformig strålare. Ofta räcker envärda mått för att detektera artefakter med god precision, men i vissa fall krävs mer komplexa vektorvärda mått (som kommer att kallas features) i kombination med en lärande klassificerare. Denna rapport visar slutgiltigt att automatisk detektion av artefakter med hög precision är möjlig och ofta kan uppnås med envärda mått. Preface This thesis project was suggested and sponsored by FLIR Systems AB in Danderyd, Sweden. I would like to thank my supervisors at FLIR, Malin Ingerhed and Lars-Åke Tunell, for the great support and encouragement that I have received during my work on this thesis. I also thank my other colleagues at FLIR for their good spirit and enthusiasm and for their interest in this project. The academic supervision of the thesis was provided by Örjan Ekeberg and the Computational Biology and Neurocomputing (CBN) group at the School of Computer Science and Communication, KTH. I thank him for his support. Oscar Danielsson, December 2006, Danderyd, Sweden