Nemanja Petrovic

Universal Impulse Noise Filter Based on Genetic Programming

In this paper, we present a novel method for impulse noise filter construction, based on the switching scheme with two cascaded detectors and two corresponding estimators. Genetic programming as a supervised learning algorithm is employed for building two detectors with complementary characteristics. The first detector identifies the majority of noisy pixels. The second detector searches for the remaining noise missed by the first detector, usually hidden in image details or with amplitudes close to its local neighborhood. Both detectors are based on the robust estimators of location and scale-median and MAD. The filter made by the proposed method is capable of effectively suppressing all kinds of impulse noise, in contrast to many existing filters which are specialized only for a particular noise model. In addition, we propose the usage of a new impulse noise model-the mixed impulse noise, which is more realistic and harder to treat than existing impulse noise models. The proposed model is the combination of commonly used noise models: salt-and-pepper and uniform impulse noise models. Simulation results show that the proposed two-stage GP filter produces excellent results and outperforms existing state-of-the-art filters.

Impulse noise detection based on robust statistics and genetic programming

A new impulse detector design method for image impulse noise is presented. Robust statistics of local pixel neighborhood present features in a binary classification scheme. Classifier is developed through the evolutionary process realized by genetic programming. The proposed filter shows very good results in suppressing both fixed-valued and random-valued impulse noise, for any noise probability, and on all test images.

Content adaptive wavelet based method for joint denoising of depth and luminance images

In this paper we present a new method for joint denoising of depth and luminance images produced by time-of-flight camera. Here we assume that the sequence does not contain outlier points which can be present in the depth images. Our method first performs estimation of noise and signal covariance matrices and then performs vector denoising. Luminance image is segmented into similar contexts using k-means algorithm, which are used for calculation of covariance matrices. Denoising results are compared with the ground truth images obtained by averaging of the multiple frames of the still scene.

Efficient video segmentation using temporally updated mean shift clustering

This paper presents a new method for unsupervised video segmentation based on mean shift clustering in spatio-temporal domain. The main novelties of the proposed approach are dynamic temporal adaptation of clusters due to which the segmentation evolves quickly and smoothly over time. The proposed method consists of a short initialization phase and an update phase. The proposed method significantly reduce the computation load for the mean shift clustering. In the update phase only the positions of relatively small number of cluster centers are updated and new frames are segmented based on the segmentation of previous frames. The method segments video in real-time and tracks video objects effectively.

Watershed data aggregation for mean-shift video segmentation

Object segmentation is considered as an important step in video analysis and has a wide range of practical applications. In this paper we propose a novel video segmentation method, based on a combination of watershed segmentation and mean-shift clustering. The proposed method segments video by clustering spatio-temporal data in a six-dimensional feature space, where the features are spatio-temporal coordinates and spectral attributes. The main novelty is an efficient data aggregation method employing watershed segmentation and local feature averaging. The experimental results show that the proposed algorithm significantly reduces the processing time by mean-shift algorithm and results in superior video segmentation where video objects are well defined and tracked throughout the time.