Correlation Filters for Pattern Recognition Using a Noisy Reference

Abstract

Abstract. Correlation filters for recognition of a target in overlapping background noise are proposed. The object to be recognized is given implicitly; that is, it is placed in a noisy reference image at unknown coordinates. For the filters design two performance criteria are used: signalto-noise ratio and peak-to-output energy. Computer simulations results obtained with the proposed filters are discussed and compared with those of classical correlation filters in terms of discrimination capability.

Accurate three-dimensional pose recognition from monocular images using template matched filtering

“…1 Pose recognition consists in estimating geometric descriptors of a target such as location coordinates and orientation angles, while the target is embedded in a 3-D space, by processing two-dimensional (2-D) image projections of the observed scene. [4][5][6] A particular combination of pose parameters results in a unique view of the target in the observed scene. 2,3 Additionally, the appearance of the target in a scene can be degraded by several factors, such as the presence of noise and background clutter, influence of incident light sources, and partial occlusions of the target, among others.…”

Design of Correlation Filters for Pattern Recognition Using a Noisy Training Image

“…Since the introduction of the matched filter [1], correlation filters have been extensively used for pattern recognition [2][3][4][5][6][7][8][9][10][11][12][13][14][15]. Two tasks of interest in pattern recognition are detection of a target and the estimation of its location in an observed scene.…”

“…In practical situations, the target may be given in a noisy reference image with a cluttered background. Recently [14,15], a signal model was introduced that accounts for additive noise in the image used for filter design. In this paper, we extend that work to account for the presence of a nonoverlapping background in a training image that is corrupted by additive noise.…”

Correlation Pattern Recognition in Nonoverlapping Scene Using a Noisy Reference

“…However, in real-life situations the target is often given by a reference image, which contains the reference object at unknown coordinates, as well as a noisy background. In a recent paper [12] a signal model was proposed that takes into account additive noise in the reference image to design filters for detecting a target in overlapping noise. The considered signal model is close to practical situations, in which observed and reference images are inevitably corrupted by noise owing to the image formation process.…”

“…Since the pioneering work by VanderLugt [1], correlation filters have been extensively studied for the purpose of pattern recognition [2][3][4][5][6][7][8][9][10][11][12][13][14][15]. Within the context of pattern recognition, detection and location estimation are two very important tasks.…”