link to homepage

JuBrain: Juelich Brain Model

Navigation and service

IAS Seminar "Brain-inspired computing for machine vision: Trainable COSFIRE filters for keypoint detection and pattern recognition"

18 Jul 2012 14:00
18 Jul 2012 15:00
Jülich Supercomputing Centre, Hörsaal (Lecture Hall), buildng 16.3, room 006


George Azzopardi, Johann Bernoulli Institute for Mathematics and Computer Science, University of Groningen


The further understanding of the visual system of the brain can provide insights to design effective computer vision algorithms. The Gabor function model, e.g., was inspired by the properties of orientation-selective neurons in areas V1 and V2 of visual cortex. However, information about the properties of subsequent cortical areas in the ventral stream of visual cortex, namely V4 and TEO, has not been sufficiently employed yet in computer vision.

We propose a novel trainable keypoint detector, which we call COSFIRE (Combination Of Shifted FIlter REsponses) filter, that is inspired by the properties of shape-selective neurons in area V4 of visual cortex. It is automatically configured to be selective for a local contour prototype pattern that is specified during a single-step training phase. The configuration comprises selecting given channels of a bank of Gabor filters and determining certain blur and shift parameters. The area of support of the resulting COSFIRE filter is adaptive as it is composed of the support of a number of orientation-selective filters whose geometrical arrangement around a point of interest is learned from a single prototype pattern. A COSFIRE filter response is computed as the weighted geometric mean of the blurred and shifted responses of the selected Gabor filters.
We demonstrate the effectiveness of the proposed COSFIRE filters in three applications: the detection of retinal vascular bifurcations, the recognition of handwritten digits, and the detection and recognition of traffic signs in 48 complex scenes.
The proposed COSFIRE filters are conceptually simple and easy to implement: they involve convolutions, blurring, shifting, and pixel-wise function evaluation. They are versatile keypoint detectors and are highly effective in practical computer vision applications.