Search

The main contribution of this paper is proposing a novel deterministic and non-iterative clustering algorithm for automatic Farsi license plate recognition (ALPR). In fact, after discarding some regions with low probability in terms of being license plate, the edge points inside every separately remained region are considered as a cluster and a Gaussian component is estimated using expectation-maximization (E-M) algorithm, for every cluster. Candidate regions are obtained by applying application-oriented thresholds for size, aspect ratio and orientation to Gaussian components. Then the license plate regions are identified by counting regions like to numeric characters, exploiting maximally stable extremal regions (MSER) detector whereas numeric characters are extracted by using a proposed algorithm, for discarding non-character regions. So, the algorithm is able to segment the license plate numeric and alphabetic characters simultaneously. Finally for character recognition, a new simple, fast and robust algorithm, which uses feature extraction and template matching technique, is proposed. The method is evaluated for detection and recognition, by using an Iranian image database that includes access control (AC), law enforcement (LE) and road patrol (RP) applications. The method is robust to rotation, skew, and multiplicity of license plate and low quality and complex background images. [ABSTRACT FROM AUTHOR]

Synthetic Aperture Radar (SAR) images suffer of multiplicative speckle noise, which damages the radiometric resolution of SAR images and makes the data interpretation difficult. Bayesian shrinkage in a transformed domain is a well-known method based on finding threshold value to suppress the speckle noise. This paper present a new approach to obtain the optimum threshold values for Bayesian shrinkage. For this purpose, we use undecimated wavelet transform (UWT), nonsubsampled Contourlet transform (NSCT), and nonsubsampled Shearlet transform (NSST). According to our experimental results, transformed coefficients influenced by noise differently. It means that some coefficients in transformed domain belong to the specific subband are more robust against noise. We use this new found property in order to determine the optimum threshold value and developed our proposed method named weighted Bayesian Shrinkage in transformed domain. Our experimental results show that finding the optimum threshold value in Shearlet domain outperforms both Contourlet and undecimated wavelet transform. Although the weighted Bayesian Shrinkage in NSCT used before, the weighted Bayesian Shrinkage in NSST is applied in this paper for the first time. [ABSTRACT FROM AUTHOR]

In this paper, a face recognition system is implemented by using Contourlet transformation (CT) as a two dimensional transformation defined in discrete form and principal component analysis (PCA) as a subspace method to form the feature vectors, is implemented. Every input image is decomposed by CT up to three levels and the CT coefficients are obtained at three scales and 15 orientations. The obtained CT coefficients are used by PCA to form the feature vectors. At the end, the Euclidean distance is used for classification. Our experimental results on ORL data base show the appropriate performance in comparison with other approaches; Even though for each subject only one image is used for training and other 9 images are used for testing. The average accuracy of our proposed algorithm for face recognition is 96.07%. [ABSTRACT FROM AUTHOR]

In this paper, embedding a watermark image is performed in discrete Contourlet transform (CT) domain. It is known that the CT is able to capture the directional edges and contours superior to discrete wavelet transform. The proposed scheme is based on embedding watermark bits into the singular value of the selected blocks within low-pass sub-band of the original gray image CT. Two algorithms are implemented, the first uses adaptive quantization and the second uses quantization step. Our experimental results show that the method with quantization step has better fidelity in terms of peak signal to noise ratio (PSNR) and is more robust against geometrical and non-geometrical attacks in terms of normalized cross correlation (NC) in comparison with the first method, though, the method with adaptive quantization is blind in the sense not only quantization strategies but also, the original image is required. [ABSTRACT FROM AUTHOR]

Traffic road sign detection is important to a robotic vehicle that automatically drives on roads. As the colors of most traffic road signs are blue and red, in this paper, we use Hue- Saturation- Intensity (HSI) color space for color based segmentation at first. Using important geometrical features, the road signs are detected perfectly. After segmentation, it turns to classify every detected road signs. For this purpose, we employ and compare the performance of three classifiers; they are distance to border (DTB), FFT sample of signature, and code matrix. In this work, we use the code matrix as an efficient classifier for the first time. Although the achieved accuracy by code matrix is greater than the two referred classifiers in average, the main advantage is simplicity and so less computational cost. [ABSTRACT FROM AUTHOR]

In this paper a license plate detection and recognition system for Iranian private cars is implemented. The proposed license plate localization algorithm is based on region elements analysis which works properly independent of distance (how far a vehicle is), rotation (angle between camera and vehicle), and contrast (being dirty, reflected, or deformed). In addition, more than one car may exist in the image. The proposed method extracts edges and then determines the candidate regions by applying window movement. The region elements analysis includes binarization, character analysis, character continuity analysis and character parallelism analysis. After detecting license plates, we estimate the rotation angle and try to compensate it. In order to identify a detected plate, every character should be recognized. For this purpose, we present 25 features and use them as the input to an artificial neural network classifier. The experimental results show that our proposed method achieves appropriate performance for both detection and recognition of the Iranian license plates. [ABSTRACT FROM AUTHOR]