Glaucoma is an eye disease. In this study, we investigated Retinal Nerve Fiber Layer (RNFL) reflectance as a biomarker for glaucoma diagnostics, noting that RNFL reflectance decreases in glaucoma due to damage. The research evaluated various normalization techniques for RNFL reflectance, as normalization was used to eliminate variability from unrelated factors and improve sensitivity to subtle changes indicative of glaucoma. The study highlighted a new metric called Post-NFL-Bright. It employed two clinical datasets to assess the repeatability, reproducibility, and diagnostic accuracy of RNFL reflectance metrics, aiming to establish a reliable method for glaucoma assessment. [Paper Link]

In my PhD project, I focused on drone infrared (IR) image processing to detect leakage in underground pipes within district heating networks. To effectively detect leakage with IR, capturing high-quality images with the drone during acquisition is essential. To address this, I developed no-reference (NR) quality estimation methods that allow the drone to adjust its parameters when the image quality significantly decreases, ensuring that high-quality images are maintained. Furthermore, I created a leakage detection method utilizing a convolutional neural network (CNN).  

We developed an automatic solar panel fault detection system by evaluating electroluminescence images using two machine learning classifiers: Support Vector Machine (SVM) and Random Forest (RF).  We trained and evaluated our models on four types of faults, namely finger failures and three types of cracks based on their severity levels, as well as on cells with no faults. We also evaluated which classifier performed better for detecting the defect types. [Paper Link] 

In my master’s project, I developed a human tracking algorithm using particle filtering (PF). Particle filters represent an object's state with a set of particles, each weighted by its match to observed data. As new observations are received, the particles are updated through prediction and correction, allowing effective tracking despite variations in appearance and occlusions. A significant challenge with PF-based models is that the traditional appearance template combines the human target with its background, making it difficult to track when the background changes. To address this, I created local patches around the target object to detect background changes, while the inner patches assist in estimating weights. I also proposed an observation model based on phase correlation.