A contrario edge detection with edgelets

Abstract

Abstract-Edge detection remains an active problem in the image processing community, because of the high complexity of natural images. In the last decade, Desolneux et al. proposed a novel detection approach, parameter free, based on the Helmhotz principle. Applied to the edge detection field, this means that observing a true edge in random and independent conditions is very unlikely, and then considered as meaningful. However, overdetection may occur, partly due to the use of a single pixelwise feature. In this paper, we propose to introduce higher level information in the a contrario framework, by computing several features along a set of connected pixels (an edgelet). Among the features, we introduce a shape prior, learned on a database. We propose to estimate online the a contrario distributions of the two other features, namely the gradient and the texture, by a Monte-Carlo simulation approach. Experiments show that our method improves the original one, by decreasing the number of non relevant edges while preserving the true ones.

Evaluating perceptual maps of asymmetries for gait symmetry quantification and pathology detection

“…In this last example, we say that the feature response is meaningful, conveying the idea that significant events, according to the human perception, are rare. This is the idea of the Helmholtz principle, and has notably been used in an a contrario framework [40], [41]. Using this contextual information, these J distributions will be exploited to decide whenever some particles need to be reassigned to new objects, and when to stop the detection algorithm.…”

A biologically inspired framework for contour detection

“…The prior one is a Gibbs distribution and includes a smoothing parameter and all the internal parameters of the model are learned on-line using an iterative conditional estimation procedure. In [23], the a contrario framework allows the authors to model the Helmholtz principle which enunciates that relevant geometric structures such as contours have a very low probability to occur in a random context. Finally, the particle filter framework proposed in [15] allows both to express and learn the prior (and transition) distribution of pieces of contour shapes in such a way that the distribution automatically embeds efficiently a semi-local information of the contour structure via every possible contour variation and then to locally track with a particle filter, these pieces of contour shapes along contours.…”

Multiscale Particle Filter through Contour Detection Framework

“…A connected pixel set is the atomic element of the proposed method, and is called an edgelet. This term shall not be mistaken with the edgelet transform (an image representation method), however, our definition is similar to the ones in [1], [2]. The structure of an edgelet is learned offline using a shape database.…”