Logo PTB

Machine learning for automatic object localization in digital images

Kolloquium der Abteilung 6

Machine learning and especially deep learning has gained enormous popularity in many application areas, e.g. computer vision, due to significant performance improvements compared to traditional approaches. In this talk, I will outline various machine learning approaches which have been developed in recent Ph.D. projects in our working group to improve object localization and image analysis performance in selected applications of industrial and medical image analysis. After summarizing key concepts in machine learning and deep learning, I will first describe an algorithm for automatic pedestrian and car localization in digital images based on the Discriminative Generalized Hough Transform and Convolutional Neural Networks. The second project focusses on a generic framework to localize a set of spatially correlated key points in medical images. Our approach generates localization hypotheses using ensembles of regression trees or alternatively convolutional neural networks, and then applies prior knowledge on the spatial configuration of the key points which is modeled in a conditional random field. In the last project, we extend the localization approach towards a pipeline for automatic vertebral fracture detection in computed tomography images using convolutional neural networks. Here, we also apply methods of explainable artificial intelligence in form of attribution methods.