3D ORDER STATISTICS FILTERS IN PROCESSING OF VIDEO SEQUENCES Volodymyr Ponomaryov, Alberto Rosales, Francisco Gallegos-Funes National Polytechnic Institute of Mexico, ESIME-Culhuacan Av. Santa Ana, 1000, Col. San Fco. Culhuacan, 04430, Mexico-city Mexico vponomar@mail.ru ; vponomar@ipn.mx ; fgallegosf@ipn.mx ABSTRACT Several novel algorithms used for impulsive noise suppression are applied in 3D color imaging. The analysis has shown that some of them are presented good performances in preservation of fine details, edges, and chromaticity. Robust algorithms that use the order statistics, vector directional and adaptive methods are developed applying 3D video processing, and realizing noise suppression. The results of the filtering in the 3D mode for video processing by proposed Video Adaptive Vector Directional filter have shown that it outperforms the video versions of Median M-type K-Nearest Neighbour, Vector Median, Generalized Vector Directional, K-Nearest Neighbour, . -trimmed Mean, and Median filters. The PSNR, MAE and NCD criteria and visual subjective analysis have been applied during simulation to evaluate performances of the proposed and existed techniques. KEY WORDS Denoising, video processing, color, and 3D processing. 1. Introduction Several applications, such as remote sensing, medical imaging, multimedia services, television, and movies make use the digital video. All these resources are affected by the noise of different nature during a broadcast stage during channel communication. This noise is usually modelled as impulsive one. Video denoising is usually realized by temporal- only or spatial-temporal filtering. It is generally agreed that in the case of low noise corruption, which is important in many real video applications, spatial- temporal filtering performs better than temporal filtering [1, 2]. However, in the case of spatial-temporal filtering there is a danger of significantly reducing the effective resolution of video, i.e. spatial blurring [3]. Algorithms to suppress noise, which employ both spatial and time correlations among pixels are called “spatial-temporal” or “three-dimensional (3D)” algorithms. Some of the investigated here algorithms are ones: Median [4], Vector Median [4], Generalized Vector Directional [5, 6], Median M-type K-Nearest Neighbour [7, 8] . that we extended from 2D to 3D space, 3D Alpha- trimmed Mean [9], and 3D K-Nearest Neighbour [9]. These filters have demonstrated good ability to remove impulsive noise, preserve fine details, and provide chromaticity properties in multichannel processing applications for 2D field [4-8, 10, 11]. 3D extensions of some of these algorithms outperform characteristics in video denoising, but they fail in preservation of the image properties due to increasing pixels quantity taken into account during the processing. In here we propose a novel method, which involves preservation of characteristics, such as fine details, edges, and chromaticity properties. The results of filtered images are evaluated using different criteria: Pick Signal to Noise Ratio to characterize the preservation of the details and edges, the Mean Absolute Error, and Normalized Color Difference criteria [8] to quantify the chromaticity properties. Adaptive [12], vector directional [5, 6. 8] and order statistics [7, 8] techniques are used in development of this proposal. Figure 1 presents a diagram illustrating procedure of the proposed method. Firstly, the directional processing is used to reject all the pixels with the biggest values in angle among others ones. These values are ordered in order statistics rule, so a set of pixels obtained on the first step that present the lowest probability of corruption are used to calculate an average value to determine: if central pixel is highly corrupted using an adaptive method, if not, the central pixel is the output’s pixel, otherwise, and finally, 3D magnitude filter is employed. Extensive simulations presented in the paper demonstrate that the proposed filter consistently outperforms well known ones in the different noise levels. We filtered the video sequences corrupted by impulsive noise of different intensity using several filters. “Miss America” and “Flowers” color sequences were treated because of their different nature in pixel distributions and colors. The sequences are processed using RGB color space, they are presented in QCIF format usually used in video conferences with 176x144 pixels for each a frame. 554-134 45