Event Detection and Automatic Summarization in Soccer Video
In this paper we propose a new complex method for automatic football video summarization, the method we have provided here does the summarizing by tow other methods, one of them do that by detecting the events and other one do that without detecting the events. One of the tools the second method used to do this was distinguishing between the views of the goal and the field-center and the first method used the slow motion features. Experimental results show the Complex method is more accurate than each of the used methods
We present an approach to detect eyes in color images. First of all, RGB facial image is converted to YCbCr one. According to YCbCr facial image, the proposed algorithm constructs two EyeMaps, one map from luminance component (EyeMapL) and the other from chrominance components (EyeMapC). When the two separate EyeMaps are constructed, we combine them to make final EyeMap. We use final EyeMap to generate potential eye candidates and then perform an extra phase on these candidates to determine suitable eye pair. This extra phase consists of flexible thresholding and geometrical tests. We test our approach on CVL and Iranian databases. Simulation results showed this phase improved the correct detection rate by about 12% and reach 98% success rate on the average
Image Denoising with a Mixture of Gaussian Distributionswith Local Parameters in Wavelet Domain
The proposed model for noise-free data distribution play an important role for maximum a posteriori (MAP) estimator. Thus, in the wavelet based image denoising, it is necessary to select a proper model for distribution of wavelet coefficients. This paper presents a new image denoising algorithm based on the modeling of wavelet coefficients in each subband with a mixture of Gaussian probability density functions (pdfs) that parameters of mixture model are local. The mixture pdf is able to model the long tailed property of wavelet coefficients and the local parameters can model the empirically observed correlation between the coefficient amplitudes. Therefore, the statistical properties of wavelet coefficients are better modelled by using this new pdf. Within this framework, we describe a new image denoising algorithm based on designing a MAP estimator, which use the mixture distributions with high local correlation. The simulation results show that our proposed technique achieves better performance than several published methods such as denoising based on mixture pdfs without local parameters both visually and in terms of peak signal-to-noise ratio (PSNR).
Image Noise Reduction Using a Wavelet Thresholding Method Based on Fuzzy Clustering
In this paper, a new method is presented for reducing the image noise by wavelet transform. Wavelet thresholding is a standard method of reducing the signal noise in which the small coefficients are replace by zero and the big ones are either remain unchanged (hard thresholding) or reduced to the level of the threshold (soft thresholding). In the proposed method, for the first time, fuzzy kmeans clustering in each sub-band is used for choosing the threshold in soft thresholding method. Using fuzzy clustering, the coefficients in each sub-band are divided into three clusters, and then the noise cluster is obtained regarding the decomposition level and the maximum coefficient in each level. The upper and lower limit of the noisy cluster is an appropriate threshold for soft thresholding. This method is more efficient for reducing Gaussian and salt and pepper noises in comparison to methods that model the noise. In other words, the proposed method is not dependent on statistical noise or data driven is the manifest feature of the proposed approach relative to other methods and the threshold is selected based on type of images without each assumption on probability density function of noise. The experiments performed on basis images, show a higher performance of the proposed algorithm relative to the statistical method and the generalized cross validation method
Image fusion is a process of combining two or more images into an image. It can extract features from source images, and provide more information than one image can. In this research, we propose a novel method for multimodality medical image fusion. Low spatial resolution limits the diagnostic potential of brain positron emission tomography (PET) imaging. As a possible remedy for this problem we propose a technique for the fusion of PET and MR images, which requires for a given patient the PET data and the T1- weighted MR image. Basically, after the registration steps, the high-frequency part of the MR, which would be unrecoverable by the set PET acquisition system is extracted and added to the PET image. This paper introduces new application of the human vision system model in multispectral medical image fusion. The methodological approaches proposed in this paper result in merged images with improved quality with respect to those obtained by HSI, DWT, wavelet à trous algorithm and wavelet based sharpening methods. Results show proposed method preserves more spectral features with less spatial distortion.
Mosaicking Images with High Motion Parallax with Application to Video Compression
Image mosaicking has been a focus of attention of many researchers in recent years. Mosaicking methods which exist today are merely unable to construct mosaics from images taken with large motion parallax. The idea is resembles some kind of layered mosaicking. The proposed method uniformly distributes the parallax mismatch between consecutive frames in the whole mosaic. In this paper, we have shown the results of the developed algorithm on two consecutive frames, as well as the result obtained on a set of frames. Although the resulted mosaic gets a bit blurry by using a set of frames with high motion parallax, it can be completely useable for video compression where the mosaic is used to compute the residuals. The proposed algorithm is much faster than the existing methods which tend to have computation time of several seconds to several minutes. Besides, a multi-resolution version of the algorithm is introduced will lessens the computation time considerably. The obtained results show the efficiency of the proposed algorithm
Multi-Features and Multi-Stages RBF Neural Network Classifier with Fuzzy Integral in Face Recognition
This paper presents a high accuracy human face recognition system using multi-feature extractors and multi-stages classifiers (MFMC), which are fused together through fuzzy integral. The classifiers used in this paper are Radial Basis Function (RBF) neural network while feature vectors are generated by applying PZM, PCA and DCT to the face images separately. Each of the feature vectors are sent to an RBF neural network classifiers and the output of these classifiers are fused to obtain better recognition rate. Experimental results on the ORL and Yale database yield excellent recognition rate
Semantic Image Segmentation Based on the Global Precedence Effect and Deformable Templates
In this paper a knowledge-based automatic“object-of-interest” extraction algorithm based on the image’s partition information and deformable template matching is proposed. The proposed algorithm is based on the similarity between the template of the “object-of-interest” and a region formed by potential fusion of image segments. By simulating the “Global Precedence Effect” (forest before trees) of the human visual system (HVS), the global/large size objects are found at lower resolutions with significantly lower omputational complexity. By using deformable templates, a generic template can be used for an object in different examples/ situations. 2D Deformable templates are modelled by some connected primitive regions and some application dependent flexibilities in angel, scale, etc.
Vehicle Velocity Detection System Based on Real-Time Motion Tracking
Intelligent transportation systems (ITS) use novel technology relying on computer vision to provide traffic parameters such as average speed in the road, lane changes, vehicles’ accelerations/decelerations, vehicles classification, etc. In this paper a mixture algorithm is proposed for velocity detection. The algorithm is based on object tracking, and relies on a perspective transformation. A novel heuristic is proposed to detect cars and trace them. The algorithm is fast enough to run real-time on a normal laptop which makes it efficient in practice (e.g. to be used by police force). There are still some open issues such as shadow detection and cancellation, and occlusion handling that will be considered in future works