Learned image representations for visual recognition

Anders Boesen Lindbo Larsen

Research output: Book/ReportPh.D. thesis

636 Downloads (Pure)


This thesis addresses the problem of extracting image structures for representing images effectively in order to solve visual recognition tasks. Problems from diverse research areas (medical imaging, material science and food processing) have motivated large parts of the methodological development. The solutions are inspired by and extend state-of-the-art techniques for describing and learning image content.

More specifically, the thesis explores two approaches to constructing image representations, namely feature engineering and feature learning. In the feature engineering approach, we devise a new image representation for texture-like patterns based on count statistics of second-order image structure. We demonstrate the discriminative capabilities of this representation on medical images and perform both cell classification and mitosis detection. Moreover, we develop an object identification method based on vector quantized local image descriptors allowing us to distinguish individual meat cuts along a production line and trace them in a non-intrusive manner. In the feature learning approach, we propose to solve the task of segmenting scanning electron microscopy images of calcite crystals by learning a meaningful pixel description to facilitate the actual segmentation. Finally, we present a new unsupervised generative image model addressing the problem of pixel-based similarity measures for images. We propose a scheme for employing feature-based similarity measures and demonstrate how this improves the ability to learn high-level concepts in images of faces.

The thesis argues in favor of learning features and presents new methods for domains with limited amounts of labeled data allowing feature learning to be applied more broadly.
Original languageEnglish
Place of PublicationKgs. Lyngby
PublisherTechnical University of Denmark
Number of pages34
Publication statusPublished - 2016
SeriesDTU Compute PHD-2016


Dive into the research topics of 'Learned image representations for visual recognition'. Together they form a unique fingerprint.

Cite this