International Journal of Applied Pattern Recognition (7 papers in press)

Regular Issues

• Zebra Crossing Detection Based on HSV Colour Model and Projective Invariant for Driver Assistance System  
  by Muhammad Kamal Hossen, Mohammad Obaidur Rahman, Anika Sadia, Md. Shahnur Azad Chowdhury 
  Abstract: This paper implements a colour and shape based Zebra crossings detection model. As Zebra crossings are white coloured, a robust binarization based on the V channel of the input image is used to find the large rectangle sized white stripes of Zebra crossing. After the morphological open operation, boundary tracing is done using Moore-neighbourhood tracing algorithm along with Jacobs stopping criterion. For every region, the bounding quadrilateral with the minimum area is computed and the candidate regions are then filtered out using some criteria e.g., aspect ratio. The y-coordinates of the lower vertices of the white stripes are considered as cluster and the optimal number of clusters is determined using the Calinski-Harabasz criterion. These clusters are now checked to determine whether they represent crosswalks and then validate using projective invariant. Various Zebra crossing images are used to test the proposed framework and the results are presented to prove its effectiveness.
  Keywords: Zebra crossing; crosswalks; intelligent transport systems; HSV; region of interest; Moore-neighbourhood; bounding quadrilateral; K-means clustering; Calinski-Harabasz; projective invariant.
  
  
• Using Empirical Mode Decomposition and KullbackLeibler distance for online handwritten signature verification  
  by Toufik Hafs, Layachi Bennacer, Hanene Brahmia, Mohamed Boughazi, Amine Naitali 
  Abstract: The signature is one of the most accepted biometric modalities in the world. In this paper, we present a new method for online handwritten signature based on the empirical mode decomposition (EMD).After extracting each of the signature coordinates, a phase of pre-processing and normalisation is carried out. Then, the features of the signatures are extracted by using the EMD. After that, three similarity measures are used to match the signatures between them. The database used in our work is the one used in Signature Verification Competition (SVC 2004). Experimental results confirm the effectiveness of our approach and show the level of its reliability. Finally, the proposed method gives an EER of 2.03 % and allows high rates of recognition compared to other approaches.
  Keywords: Online signature; Verification; Biometrics; Kullback–Leibler divergence; Empirical Mode Decomposition (EMD); Hilbert transform.
  
  
• Influence of noise, light and shadows on image segmentation algorithms  
  by Pedro Furtado, Pedro Martins, Jose Cecílio 
  Abstract: Image segmentation is a required step in many image processing and object recognition tasks, where regions obtained by segmentation are analyzed to classify as specific objects. In all those procedures, image segmentation tech- niques are used to divide the image into multiple regions or clusters. Each image pixel will be assigned to one of the clusters using different metrics such as pixel color value, gray-scale intensity, edges, shapes, among others. The most diverse image segmentation algorithms have already been proposed and are currently used. It is important to analyze the robustness of those algorithms, taking into consideration diverse types of noise and difficult con- ditions. Both the types of noise and difficult conditions and the number of algorithms makes this a difficult task. In this paper we choose a small set of frequently-used algorithms and analyze their behavior in those robustness- testing conditions. In order to do that, we inject difficulties (noise, shadows, various degrees of illumination) and compare the quality of the segmentation with those algorithms against a ground truth. The objective is to analyze how differences in illumination, shadows or noise influence the output of the algorithms, and how they compare on those metrics. Based on those results we conclude about the quality of the approaches tested.
  Keywords: Image segmentation; image clustering; algorithms.
  
  
• Deep neural network based phoneme classification of standard Khasi dialect in continuous speech  
  by Bronson Syiem, L. Joyprakash Singh 
  Abstract: In this paper, a deep neural network (DNN) is used to classify phonemes of the standard Khasi dialect which is one of the commonly used dialects in the state of Meghalaya. For this, clean speech data were recorded in the laboratory from native speakers. In the proposed system, a monophone and a triphone hidden Markov models (HMMs) were also built to compare the results obtained. It was found that DNN outperformed the classification over the other two models with a classification accuracy of 89.70 %.
  Keywords: Acoustic model (AM); DNN; Gaussian mixture model (GMM); HMM; Language model (LM); Mel frequency cepstral coefficient (MFCC); Phone error rate (PER); Voice activity detection (VAD); Word error rate (WER).
  
  
• Player skill estimation for soccer match prediction  
  by Juan Pablo Maldonado Lopez, Vojtech Jindra 
  Abstract: In this paper we propose two algorithms for assessing skill of soccer players. We introduce an adaptation of the Elo rating algorithm, which is widely used in chess, to handle teams. A different approach is proposed to assess player skill as a function of the proportion of matches won. Since it is hard to measure skill directly, to estimate the performance of our approach we propose as a proxy the number of correctly predicted match outcomes of a classification algorithm that takes as input our algorithms' output. Numerical experiments suggest that our approach is competitive with bookkeeper's estimates, even though we rely only on historical match outcome data.
  Keywords: Elo ranking; team skill; skill-based ranking.
  
  
• In-Line Grading System for Mango Fruits Using GLCM Feature Extraction and Soft-Computing Techniques  
  by Ebenezer Olaniyi, Oyebade Oyedotun, Clement Ogunlade, Adnan Khashman 
  Abstract: In the fruit production industries and supermarkets, mature (ripe) fruits are demanded for consumption by the consumers and also for production in fruit processing industries. Therefore, there is an urgent need for an in-line grading system in such industry to aid the grading of mango fruit; in order to enhance the use of ripe and mature mangoes for production. Also, such in-line grading systems will speed up the production in these industries since machines are faster which gives a better and standard result as compared with human operators. In this work, we have implemented an in-line grading system using GLCM feature extraction and soft computing techniques. Two models have been implemented to classify the mango fruits into mature (ripe) and immature (unripe) fruits. These models are the feed-forward network trained with back-propagation neural network and the radial basis function network. These models are compared with each other and also with the result of other proposed systems using the same data base to ascertain the best result required in such industry.
  Keywords: Mature mango; Unripe Mango; Food Quality Control; Radial basis function; Neural Network.
  
  
• A Robust Multi-level Sparse Classifier with Multi-modal Feature Extraction for Face Recognition  
  by Virendra P. Vishwakarma, Gargi Mishra 
  Abstract: In the past few years, face recognition based on sparse representation is providing satisfactory classification accuracy. Face images involved in real life applications usually exhibit considerable pose, lighting, and expression variations, resulting in significant performance degradation of traditional sparse based algorithms. In this paper, a novel face recognition method is developed as multi-level sparse (MLS) classifier with multi-modal feature extraction, which integrates benefits of sparse representation manifolds. In MLS classifier, sparse representation based classification is performed at multiple levels to extract the hierarchical relationship information between training and testing images, which not only improves classification accuracy but also makes the system scalable. Also, the use of multi-modal feature in MLS classifier makes it discriminative to face changes while robust to intra-personal variations. To highlight the competency of proposed method, results are compared with sparse representation and other existing state-of-art methods in terms of mean classification error. An investigation on classification accuracy is performed to showcase the reliability of proposed method.
  Keywords: Multilevel sparse; Face recognition; Sparse representation; Pattern recognition; Nearest neighbours.