Erol Seke

Super resolution using variable size block-matching motion estimation with rotation

Motion estimation (ME) for image sequences is one of the important processes in image restoration. Block matching algorithms can be used for local ME. Recently, the variable block size motion estimation (VBS-ME) has been received attention in order to improve the performance of the block matching algorithms. In this paper, we have evaluated variable-sized-block-matching ME with rotation in the registration step. Results of proposed method are compared with exhaustive search (ES) that does not consider rotation and static block size (BS) angle search (AS) that considers rotation. Both peak-signal-to-noise-ratio (psnr) and visual quality are used in comparison of all three methods.

Image denoising using common vector approach

Common vector approach (CVA) is an increasingly popular classification method in recognition problems where probability of having the dimensionality of the problem higher than the number of data items is not zero. In CVA, common component of the members of classes is separated from the discriminating difference parts and used to determine whether a given vector (a block of data) belongs to the class in question, or to find out the class it belongs to. In this study, overlapping image blocks near the current pixel to be denoised are used as input data and a class is constructed per pixel position. Denoised image block is then constructed with the sum of common vector of the class and difference vector of the centre block denoised by linear minimum mean square error estimation technique. Since the classes are formed using similar blocks, the edges are preserved while denoising the image.

Least Squares Sub-pixel Registration Refinement Using Area Sampler Model

Super-resolution applications require sub-pixel registrations of low resolution images to be almost exact due to the deterioration caused by inaccurate image registration. A linear-least-squares technique is proposed to refine sub-pixel translation parameters, which can be employed when the images are registered but just where there is not enough sub-pixel accuracy. In the technique, it is assumed that low resolution pixels are obtained by area sampling high resolution pixel field which have twice the density of their low resolution correspondents. Using this downsampling schema, a set of equations is formed. Assumed geometry and layout provide a constraint set to be used with the equation set. The sub-pixel translations are then found using least-squares-solution-with-equality-constraints. The method is shown to improve the registration accuracy.

Denoising speech signal using common vector approach

In this paper, we proposed a speech signal denoising method using common vector approach (CVA) for very noisy signals. Based on CVA, classes are constructed from windowed speech samples/vectors according to their characteristics and denoising is performed on difference vectors. Speech signal is then reconstructed using denoised patches. Size of windows and amount of overlapping are important parameters in this method, affecting the performance/noise reduction ratio. CVA results are compared against the results of Geometric Approach and Magnitude Squared Spectrum methods according to well-known performance measures. Experimental results show that the proposed common vector based method is superior to other techniques for very noisy speech signals. Method is open to enhancement with different parameters and noise estimation methods.

Direct generation of upsampled FIR filter response a simple extension to filters with distributed arithmetic

A memory based upsampling/interpolating FIR filter modification/extension to distributed arithmetic (DA) based FIR filters is proposed that can be used for any filter coefficient set. Use of minimum or no multiplier is a desired design property when signal processing is performed using FPGAs since multipliers are scarce/expensive resources within FPGAs whereas registers and such are abundant. Upsampling a digital stream is usually performed by inserting zeros between original samples followed by a low pass filter to reject images. Compared to basic distributed arithmetic based filter designs where partial products/sums are stored in memory blocks, our design stores interpolation values. These samples are output sequentially using a simple counter, eliminating zero insertions and saving circuit elements. As an example FIR filter, we have designed a raised-cosine band-limiting filter with example roll-off factor and upsampling values. Successful implementation using VHDL+FPGA with ease has proven that the approach is a simple and effective compared to input upsampling.

Frame rate up-conversion with nonlinear temporal interpolation

Motion compensated frame rate up-conversion (MCFRUC) partly resolves some of the issues in video interpolation such as blurred images resulting from incorrect linear motion interpolation and motion jerkiness caused by frame repetition. Most of the block artifacts and incorrect interpolated frames are caused by linear-motion-based FRUC. In this paper, we evaluated a motion-compensation approach based on nonlinear spatio-temporal motion assumption which is more general and realistic than linear assumption. New motion model is tested against linear counterpart, proving clear superiority in generating better reconstructed and closer to the ground truth intermediate frames.

Determining illumination sources in frontal face images using salient samples from a face depth map

An efficient method is presented for determining multiple light sources illuminating faces using a general face depth map for comparable image generation. A 3D face depth map is artificially illuminated by known sources from several known directions and a set of samples from each resulting image is collected. These and corresponding samples from probed face image are used to find the light source vectors that generates an artificial face image whose samples are illuminated optimally similar to the probed face. Since the artificial data are pre-calculated and stored, illumination calculation is quite fast for a given frontal face image.

Real-time Motion Region Detection and Contouring System

Many real world applications require real-time motion detection. Performance of a motion detection system should be fast enough so that moving objects in video can be detected and processed in real time. Once motion region in a video is detected, object tracking, image data mining, semantic meaning extraction, and other video/image processing techniques can be performed. There is a lot of research in realtime motion detection however most of these studies are either computationally expensive or not robust enough against real world image overheads like background clutter and pixel noise. In this paper, we propose a robust real-time motion detection system with low computational complexity. The proposed system finds the difference images from successive frames, applies filtering and the pixel clustering operation on them. Then, sub-contours of the observed motion regions are found using either minimum bounding coarse rectangles, minimum bounding rectangles using moments or minimum bounding convex hull methods. Finally, our contour merging algorithm merges sub-contours to identify moving object regions. This novel approach, detects moving objects in a video frame with tight enclosing boundary contours without sacrificing robustness. Since no constraint is attached to the technique, it is a good candidate to be used as motion detection of choice in wide range of applications. We obtained promising results in the experiments with real time data.

A new speech signal denoising algorithm using common vector approach

Speech denoising may improve intelligibility of speech and hearing comfort in voice communication/recognition applications in noisy environments. It can also be used to enhance old recordings. Most speech enhancement methods are intrusive and cause some loss in the signal component while removing noise. In this paper, we propose a method based on common vector approach (CVA) for reducing losses in single-channel enhancement algorithms. In the proposed technique, overlapping speech sample frames are collected in classes according to their similarity and common and difference vectors of the classes are separated using CVA. Since the noise component is uncorrelated and therefore presumably concentrated in the difference part, difference vectors are denoised using a common denoising technique and sample frames are reconstructed by combining the common and the denoised difference parts. This operation does not affect the common vector and somewhat secures improvement even for highly noised data. Compared to the state-of-the-art, highly promising results are obtained in terms of several speech quality measures.

Image enhancement with outlier detection

In this study, a new method is proposed to achieve a high resolution (HR) image from low resolution (LR) video frames, which include camera motions as well as local motions. The best matching block in the motion estimation is obtained by using the variable size block matching (VSBM) algorithm. Furthermore, before merging the pixels, the outlier pixels were cleaned using statistical outlier detection method. Peak-signal-to-noise ratio (PSNR) and structural similarity index measurement (SSIM) were used to compare the experimental results of the proposed method.