Martin Lillholm

Scale Space Methods in Computer Vision

In an ordinary 2D image the critical points and the isophotes through the saddle points provide sufficient information for classifying the image into distinct regions belonging to the extrema (i.e. a collection of bright and dark blobs), together with their nesting due to the saddle isophotes. For scale space images, obtained by convolution of the image with a Gaussian filter at a continuous range of widths for the Gaussian, things are more complicated. Here only scale space saddle points occur. They are related to spatial saddle points and spatial extrema and can thus provide a scale space based segmentation and hierarchy. However, a spatial extremum can be related to multiple scale space saddles. The key to solve this ambiguity is the investigation of both the scale space saddles and the iso-intensity manifolds (the extension of isophotes in scale space) through them. I will describe the different situations that one can encounter in this investigation, which scale space saddles are relevant, give examples and show the difference between selecting the relevant and the non-relevant (“void”) scale space saddles.

Feature-Based Image Analysis

According to Marrs paradigm of computational vision the first process is an extraction of relevant features. The goal of this paper is to quantify and characterize the information carried by features using image-structure measured at feature-points to reconstruct images. In this way, we indirectly evaluate the concept of feature-based image analysis. The main conclusions are that (i) a reasonably low number of features characterize the image to such a high degree, that visually appealing reconstructions are possible, (ii) different feature-types complement each other and all carry important information. The strategy is to define metamery classes of images and examine the information content of a canonical least informative representative of this class. Algorithms for identifying these are given. Finally, feature detectors localizing the most informative points relative to different complexity measures derived from models of natural image statistics, are given.

What Do Features Tell about Images

According to the Marr paradigm [10], visual processing is performed by low-level feature detection followed by higher level task dependent processing. In this case, any two images exhibiting identical features will yield the same result of the visual processing. The set of images exhibiting identical features form an equivalence class: a metameric class [7]. We choose from this class the (in some precise sense) simplest image as a representative. The complexity of this simplest image may in turn be used for analyzing the information content of features. We show examples of images reconstructed from various scale-space features, and show that a low number of simple differential features carries suficient information for reconstructing images close to identical to the human observer. The paper presents direct methods for reconstruction of minimal variance representatives, and variational methods for computation of maximum entropy and maximum a posteriori representatives based on priors for natural images. Finally, conclusions on the information content in blobs and edges are indicated.

Single gaze gestures

This paper examines gaze gestures and their applicability as a generic selection method for gaze-only controlled interfaces. The method explored here is the Single Gaze Gesture (SGG), i.e. gestures consisting of a single point-to-point eye movement. Horizontal and vertical, long and short SGGs were evaluated on two eye tracking devices (Tobii/QuickGlance (QG)). The main findings show that there is a significant difference in selection times between long and short SGGs, between vertical and horizontal selections, as well as between the different tracking systems.

Breast Tissue Segmentation and Mammographic Risk Scoring Using Deep Learning

Mammographic scoring of density and texture are established methods to relate to the risk of breast cancer. We present a method that learns descriptive features from unlabeled mammograms and, using these learned features as the input to a simple classifier, address the following tasks: i) breast tissue segmentation ii) scoring of percentage mammographic density (PMD), and iii) scoring of mammographic texture (MT). Our results suggest that the learned PMD scores correlate well to manual ones, and that the learned MT scores are more related to future cancer risk than both manual and automatic PMD scores.

On image reconstruction from multiscale top points

Image reconstruction from a fiducial collection of scale space interest points and attributes (e.g. in terms of image derivatives) can be used to make the amount of information contained in them explicit. Previous work by various authors includes both linear and non-linear image reconstruction schemes. In this paper, the authors present new results on image reconstruction using a top point representation of an image. A hierarchical ordering of top points based on a stability measure is presented, comparable to feature strength presented in various other works. By taking this into account our results show improved reconstructions from top points compared to previous work. The proposed top point representation is compared with previously proposed representations based on alternative feature sets, such as blobs using two reconstruction schemes (one linear, one non-linear). The stability of the reconstruction from the proposed top point representation under noise is also considered.

Single stroke gaze gestures

This paper introduces and explains the concept of single stroke gaze gestures. Some preliminary results are presented which indicate the potential efficiency of this interaction method and we show how the method could be implemented for the benefit of disabled users and generally how it could be integrated with gaze dwell to create a new dimension in gaze controlled interfaces.

Natural image profiles are most likely to be step edges

We introduce Geometric Texton Theory (GTT), a theory of categorical visual feature classification that arises through consideration of the metamerism that affects families of co-localised linear receptive-field operators. A refinement of GTT that uses maximum likelihood (ML) to resolve this metamerism is presented. We describe a method for discovering the ML element of a metamery class by analysing a database of natural images. We apply the method to the simplest case--the ML element of a canonical metamery class defined by co-registering the location and orientation of profiles from images, and affinely scaling their intensities so that they have identical responses to 1-D, zeroth- and first-order, derivative of Gaussian operators. We find that a step edge is the ML profile. This result is consistent with our proposed theory of feature classification.

Basic Image Features (BIFs) Arising from Approximate Symmetry Type

We consider detection of local image symmetry using linear filters. We prove a simple criterion for determining if a filter is sensitive to a group of symmetries. We show that derivative-of-Gaussian (DtG) filters are excellent at detecting local image symmetry. Building on this, we propose a very simple algorithm that, based on the responses of a bank of six DtG filters, classifies each location of an image into one of seven Basic Image Features (BIFs). This effectively and efficiently realizes Marrs proposal for an image primal sketch. We summarize results on the use of BIFs for texture classification, object category detection, and pixel classification.

Diagnosis of osteoarthritis and prognosis of tibial cartilage loss by quantification of tibia trabecular bone from MRI.

A longitudinal study was used to investigate the quantification of osteoarthritis and prediction of tibial cartilage loss by analysis of the tibia trabecular bone from magnetic resonance images of knees. The Kellgren Lawrence (KL) grades were determined by radiologists and the levels of cartilage loss were assessed by a segmentation process. Aiming to quantify and potentially capture the structure of the trabecular bone anatomy, a machine learning approach used a set of texture features for training a classifier to recognize the trabecular bone of a knee with radiographic osteoarthritis. Using cross‐validation, the bone structure marker was used to estimate for each knee both the probability of having radiographic osteoarthritis (KL >1) and the probability of rapid cartilage volume loss. The diagnostic ability reached a median area under the receiver‐operator‐characteristics curve of 0.92 (P < 0.0001), and the prognosis had odds ratio of 3.9 (95% confidence interval: 2.4–6.5). The medians of cartilage loss of the subjects classified as slow and rapid progressors were 1.1% and 4.9% per year, respectively. A preliminary radiological reading of the high and low risk knees put forward an hypothesis of which pathologies the bone marker could be capturing to define the prognosis of cartilage loss. Magn Reson Med 70:568–575, 2013.