PhD ceremony Mr. G. Azzopardi: Cosfire. Combination of shifted filter responses. A trainable filter approach to visual pattern recognition
|Fr 26-04-2013 at 14:30
|Academiegebouw, Broerstraat 5, Groningen
PhD ceremony: Mr. G. Azzopardi, 14.30 uur, Academiegebouw, Broerstraat 5, Groningen
Dissertation: Cosfire. Combination of shifted filter responses. A trainable filter approach to visual pattern recognition
Promotor(s): prof. N. Petkov
Faculty: Mathematics and Natural Sciences
This thesis proposes an innovative trainable detection approach to visual pattern recognition, called COSFIRE (Combination of Shifted Filter Responses). It is inspired by neurophysiological evidence about the visual processing in the ventral stream of the brain for the recognition of objects.
The proposed COSFIRE approach is trainable, in that a user speciﬁes a pattern of interest from an image and in an automatic way a detector is conﬁgured, which is then able to locate and recognize the same and similar patterns in other images. The patterns of interest may vary from simple edges or lines, to objects with complex shapes.
COSFIRE detectors use the shape properties (geometrical arrangement of contour parts) to spot patterns of interest. They are robust to contrast variations and to the presence of noise or texture around such patterns. Moreover, they are able to detect patterns of different orientations and sizes as well as reﬂected versions of the pattern that is speciﬁed by the user.
The high effectiveness of the COSFIRE approach is demonstrated in various applications, namely contour detection in images of natural scenes, detection of vascular bifurcations in retinal fundus images, detection and recognition of trafﬁc signs in outdoor environments, recognition of handwritten digits and letters, together with recognition and localization of deformable objects in complex settings.
This work contributes to the continuing trend of simulating biological vision to design more effective and robust computer vision solutions. The proposed detectors provide a foundation of innovative visual pattern recognition to many computer vision applications.