Correlation Filters For Pattern Recognition Research Articles

Design of correlation filters for pattern recognition using a noisy reference

We present the design of correlation filters for detection of a target in a noisy input scene when the object of interest is given in a noisy reference image. The target signal, shape and location in the reference image are assumed to be unknown. Two signal models are considered for the input scene: additive and nonoverlapping. The design of the filters consists of automated estimation of needed parameters from a noisy reference image and maximization of the peak-to-output energy ratio criterion. Two filter variants are proposed. The matching error metric is used to determine the regions of the parameter space where each filter variant performs better. Computer simulation results obtained with the proposed filters are presented and evaluated in terms of discrimination capability, location errors, and tolerance to input noise.

Design of correlation filters for pattern recognition with disjoint reference image

Correlation filters for pattern recognition are commonly designed under the assumption that the shape and appearance of an object of interest are explicitly known. In this paper, we consider a signal model in which an object of interest is given at unknown coordinates in a cluttered reference image and corrupted by additive noise. The reference image is used to design filters for detecting a target in scenes with a nonoverlapping background and additive noise. An optimum correlation filter with respect to peak-to-output energy for object detection is derived. The shape and appearance of the target are estimated from the reference image. Two methods to estimate the frequency response of the derived filter are used. Computer simulation results obtained with the proposed filters are presented and discussed. The performance of the filters is evaluated in terms of discrimination capability and location accuracy for different statistics of the backgrounds and noise processes present in the signal model.

Pattern recognition in nonoverlapping background with noisy target image

In the design of conventional correlation filters for pattern recognition, the appearance and the shape of a target are assumed to be known. In this work, the target is assumed to have unknown coordinates in a noise-corrupted reference image. We obtain a filter th at is optimal in terms of the ratio of the correlation peak to the energy of the correlation plane. Computer simulation is used to compare the performance of the conventional and the developed filters with regard to the detection of targets.

Adaptive correlation filters for pattern recognition

Adaptive correlation filters based on synthetic discriminant functions (SDFs) for reliable pattern recognition are proposed. A given value of discrimination capability can be achieved by adapting a SDF filter to the input scene. This can be done by iterative training. Computer simulation results obtained with the proposed filters are compared with those of various correlation filters in terms of recognition performance.

Quad-phase correlation filters for pattern recognition

A quad-phase-only filter (QPOF) is introduced. It is shown that the QPOF can have improved correlation response and signal-to-noise ratio with respect to the Hartley binary phase-only filter for some objects.