Chirag N. Paunwala

Hybrid Approach for Single Image Super Resolution Using ISEF and IBP

This paper addresses the problem of recovering a super-resolved image from a single low resolution input. This is a hybrid approach of single image super resolution. The technique is based on combining an Iterative back projection (IBP) method with the edge preserving Infinite symmetrical exponential filter (ISEF). Though IBP can minimize the reconstruction error significantly in iterative manner and gives good result, it suffers from ringing effect and chessboard effect because error is back-projected without edge guidance. ISEF provides edge-smoothing image by adding high frequency information. Proposed algorithm integrates ISEF with IBP which improves visual quality with very fine edge details. The method is applied on different type of images including face image, natural image and medical image, the performance is compared with a number of other algorithms, bilinear interpolation, nearest neighbor interpolation and Laplacian of Gaussian (LOG). The method proposed in the paper is shown to be marginally superior to the existing method in terms of visual quality and peak signal to noise ratio (PSNR).

An efficient skew detection of license plate images based on wavelet transform and principal component analysis

Character segmentation and recognition are imperative steps in the vehicle license plate recognition (VLPR) system. The skewed license plate affects badly on the accurate character segmentation and recognition. To solve the problem, an efficient approach for skew correction of license plate is proposed based on wavelet transform and principal component analysis. First, a skew feature image of the original VLP image, which preserves the license plates horizontal feature, is extracted by the two level wavelet transform. By applying a threshold, the skew feature image is then transformed in to a binary image which helps to identify the feature points which are considered as the edge of characters on the license plate with a lined regulation which reflects the slant angle of the plate useful for principal component analysis. Using principal component analysis the direction of the principal component, which gives the information about tilted angle of the plate, is achieved with the help of feature points, then the correction of the plate is accomplished. Main advantage of the proposed method is simplicity with less computationally complexity, which makes it useful for real time applications.

Statistical analysis of various kernel parameters on SVM based multimodal fusion

Biometric systems accurately recognise/authenticate an individual to access his confidential data/accounts. When multiple traits are fused together at feature/ score/ decision level, it results into highly accurate multimodal systems. This system improvise rate of recognizing an individual. Multiple biometric traits cannot be cloned simultaneously and hence it is highly secured system. The match scores of different persons are sufficient enough to recognise them and differentiate them from each other. The match scores do not require higher storage capacity as well as higher computational complexity. Hence, match score fusion is highly preferable to recognize an individual. Fusion at match score level has been carried out by several researchers with various state of arts namely weighted sum rule, product rule, majority voting rule, Support vector machine (SVM), Bayesian fusion, fuzzy rule method, etc. In this paper SVM based fusion of match scores for face and fingerprint biometric trait is implemented. Main research focus of this paper is on statistical analysis of different kernel methods namely Polynomial kernel, Radial Basis Function (RBF) kernel and Multilayer perceptron (MLP) kernel used for training SVM. The statistical analysis is based on training time required for training SVM using all three kernel methods as well upon the performance curve in terms of recognition rates i.e. Genuine Acceptance Rate (GAR) and False Acceptance Rate (FAR) of SVM fused system. SVM fusion has been implemented in MATLAB software and the results reveal that RBF kernel based SVM fused system requires the lowest training time as compared to other kernel methods. Even the recognition performance of RBF based SVM system is higher as compared to other kernel based systems i.e. GAR of RBF fused system increases and is better as compared to other kernel based SVM systems.

Improved weight assignment approach for multimodal fusion

Biometric systems are emerging trends in the world of technological revolution for accessing confidential information of an individual. Multibiometric systems are more preferable nowadays as it overcomes the problem of non-universality, increases recognition performance and spoofing faced by the unimodal system. The multimodal biometric system is built up by fusing the features/ scores/ decisions of individual model. Fusion at the match score level is more preferable as it is more efficient in terms of computational complexity and contains sufficient information to recognize an individual. In the proposed bimodal rule based method, weight assignment strategy is based upon the Equal Error Rate (EER) and Genuine Acceptance Rate (GAR) values of individual models. It has been observed that total error rate (TER) of a fused system decreases and the Genuine Acceptance Rate (GAR) of fused system increases i.e. recognition performance of fused system is better as compared to individual systems.

Exemplar-based image inpainting using ISEF for priority computation

Inpainting is an art of restoring damages or removal and filling in removed region. Exemplar-based techniques simply fill target region iteratively using best matching patch from rest of the image. An important part of this model is data term which decides order of filling procedure. In most of the methods, data term depends on strength of an isophote which is found from gradient operator. Gradient operator is a poor numerical differentiation based approximation of edges. In this paper, a new data term is proposed which is based on gradient obtained by application of Infinite size Symmetric Exponential Filter (ISEF). ISEF is a one-step based multi-edge model which detects edges by differentiation of optimal exponential and obtained edges help to preserve structure information. New ISEF based data term improves ordering of patches and helps preventing linear and curved structures. Proposed method has been applied to various images containing structure and composite structure. It can be seen that proposed method produces comparable results. Experimental results show that proposed method not only improves visual quality but also improves Peak Signal-to-Noise Ratio (PSNR) which is considered as a quantitative measure.

An improved license plate extraction technique based on gradient and prolonged Haar wavelet analysis

In a vehicle license plate extraction system, plate region detection is the key step before the final recognition. This paper presents a license plate detection algorithm from complex background based on gradient analysis and prolonged haar wavelet transform. First the license plate region is approximately detected using gradient analysis and top hat transformation of horizontal projection by choosing appropriate threshold. Accurate detection is obtained using multi resolution feature of candidates by using prolonged haar wavelet transformation. Due to the limitation of haar wavelet, we use expanded version of it to get better location of license plate without using morphological operations. Finally, accurate vertical position of license plate is detected and license plate is extracted from image. The test images taken from various scenes were employed, including diverse angles, different lightening conditions. The experiments shows that proposed method can quickly and correctly detect the region of license plate.

Learning based single frame super-resolution using Lorentzian error norm & Gabor filter

Super-resolution (SR) is the process of processing multiple low resolution (LR) images or a single low resolution image to form a high resolution (HR) image. Here, learning based approach is used to perform super-resolution on a single low resolution image. This Learning-based super-resolution algorithm synthesizes a high-resolution image based on learning patch pairs of low-resolution and high-resolution images of training set. Since a low-resolution patch is usually mapped to multiple high-resolution patches, unwanted outliers or blurring can appear in super-resolved images. Therefore, for HR patch selection from training set, we have considered Lorentzian error norm, which efficiently reject outliers which cause artifacts. Gabor filter is used to obtain prior information about the original HR image followed by optimization using iterative method. Experimental results demonstrate that the proposed algorithm can synthesize higher quality, HR images compared to the existing algorithms.

Fusion fingerprint minutiae matching system for personal identification

Fingerprint is most popular biometric because it can easily used and its features are highly reliable. Many fingerprint matching algorithm have been reported in literatures and in all most algorithms minutiae matching is used. In this paper modified fusion algorithm is proposed. It uses one technique of structure with Delaunay triangles based, formed by minutiae, and for matched candidate triangle local ridge texture information is extracted using Local Binary Pattern (LBP). Here triangle matching score is given a full weightage and minutiae matching score is given weightage as per requirement. The experimental results show that proposed algorithm is effective and reliable. Algorithm is tested on a DB1 Database of FVC 2004 competition.

An adaptive integrated rule-based algorithm for license plate localization

This paper addresses a license plate localization (LPL) algorithm for a complex background. Most of LPL algorithm works on restricted conditions, as well as on a principle of sequential elimination of blocks from image level to final LP candidate region. In most of algorithms, blocks are filtered out for not satisfying required LP features in a top-down approach and this may result in a poor efficiency in a complex scenario. The major steps of the proposed approach are adaptive edge mapping, saliency measure of edge based rules with confidence level estimation using fuzzy rules and final step for reassessment of decision by colour attributes filtering. The proposed algorithm is adaptive to across the country variations in LP standards, as well as it is tested on two data sets each one consisting of more than 700 images, set-1 being for good images while set-2 including only constrained images. The algorithm is tested for a low contrast due to overexposure or poor lighting, existence of multiple plates, variation in aspect ratio and compatible background conditions. It has been observed, that the performance degradation imposing complex condition is nominal.

Hybrid Approach for Single Image Super Resolution using ISEF and IBP: Specific Reference to License Plate

This paper addresses the problem of recovering a super-resolved image from a single low resolution input. This is a hybrid approach of single image super resolution. The technique is based on combining an Iterative back projection (IBP) method with the edge preserving Infinite symmetrical exponential filter (ISEF). Though IBP can minimize the reconstruction error significantly in iterative manner and gives good result, it suffers from ringing effect and chessboard effect because error is back-projected without edge guidance. ISEF provides edge-smoothing image by adding high frequency information. Proposed algorithm integrates ISEF with IBP which improves visual quality with very fine edge details. The method is applied on different type of images including face image, natural image and medical image, the performance is compared with a number of other algorithms, bilinear interpolation, nearest neighbor interpolation and Laplacian of Gaussian (LOG). The method proposed in the paper is shown to be marginally superior to the existing method in terms of visual quality and peak signal to noise ratio (PSNR).