This paper introduces a prescriptive analytics approach to solving markdown-pricing optimization for an e-commerce retailer capable of price differentiation based on customer demand elasticity and the cost of delivery or other services. We consider a situation when the retailer has a limited but potentially large amount of inventory that is stored at multiple fulfillment centers and must be sold by a certain exit date. The objective is to maximize the gross profit, defined as the total revenue minus total shipping cost. We propose a model which predicts, based on historical data, the demand from each customer group as a function of price. Then we formulate the optimization using non-linear objective function and constraints and describe a so-called randomized decomposition approach to finding a near-optimal solution. Finally, we discuss the results of our computational experiments.
This research studied violence detection from less than 6-second ECG signals. Features were calculated based on the Bivariate Empirical Mode Decomposition (BEMD) and the Recurrence Quantification Analysis (RQA) applied to ECG signals from violence simulation in a primary school, involving 12 pupils from two grades. The feature sets were fed to a kNN classifier and tested using 10-fold cross validation and leave-one-subject-out (LOSO) validation in subject-dependent and subject-independent training models respectively. Features from BEMD outperformed the ones from RQA in both 10-fold cross validation, i.e. 88% vs. 73% (2nd grade pupils) and 87% vs. 81% (5th grade pupils), and LOSO validation, i.e. 77% vs. 75% (2nd grade pupils) and 80% vs. 76% (5th grade pupils), but have larger variation than the ones from RQA in both validations. Average performances for subject-specific system in 10-fold cross validation were 100% vs. 93% (2nd grade pupils) and 100% vs. 97% (5th grade pupils) for features from the BEMD and the RQA respectively. The results indicate that ECG signals as short as 6 seconds can be used successfully to detect violent events using subject-specific classifiers.
A novel procedure to quantify similarity in different images based on time-series analysis is reported. Pattern recognition and matching operation includes parameters such as mean, correlation, mutual information etc. The proposed technique consists of orderly application of various mathematical transformations on one dimensional time series obtained from a 2D image array. These transformations include array to time-series conversion, local maxima detection-joining, and calculation of cumulative angle. The final calculated parameter is a direct pointer to the image similarity. The proposed technique performed well against traditional image comparison techniques under specific circumstances. The technique can be also used to identify similar patterns in a single image. The simulation codes have been written on SCILAB platform.
Hidden Markov Models or HMMs, are a relatively recent phenomenon for Arabic handwriting recognition. They are robust and efficient in classification. In this paper, an effort has been made to further boost the recognition capability of HMM Based Arabic Optical Character Recognition Systems, by using a two-tier hybrid classification scheme. The first tier consists of Part of Arabic Word or PAW, Based HMMs, and the second tier is a k-Nearest Neighbor Classifier or KNN Classifier. The second tier receives its inputs from the first tier. A second novel Hybrid Scheme is also examined. The recognition accuracies of the proposed schemes have been compared to contemporary techniques and they show an improvement in classification accuracy. HMMs have been implemented using the HTK Toolkit. The database used has been obtained from the IFN-ENIT Database of Arabic Words.
Clustering is a popular method essentially applied to data analysis, data mining, vector quantization and data compression. The most widely used clustering algorithm, which belongs to the group of partitioning algorithms, is the k-means. In this paper, we propose an extended version of k-means where the initial cluster centers are selected based on a heuristic data based formula, in contrast to random selection adopted by the traditional k-means algorithm. In particular, a new formula for selecting the initial cluster centers, before applying the k-means algorithm for clustering of a data set, is introduced. The new extended k-means algorithm is tested on clustering a set of 2-D data points. The obtained results exhibit superiority with respect to clustering compactness of the proposed algorithm as compared to traditional k-means. The validity of the extended algorithm is assessed through a set of clustering measures (Silhouette, Davies-Bouldin), with the most prominent being the Davies-Bouldin measure, that identify how compactness and well-separated the clusters are.
The vulnerability of conventional biometrics to spoof has caused considerable concern especially in those fields that require high reliable user identification. This heightened concern leads to great interest in assessing the probability and efficiency of using eye movements in identification systems. By applying eye movements as biometrics a new approach has been taken into human identification including all the crucial attributes of previous traditional identification that may offer certain notable advantages. The most obvious is the inherent difficulty in forging them. The purpose of this article is to review examples of researches utilizing eye movements in human identification. These studies can be divided into two groups: the first group utilizes eye movement bioelectrical signals in identification purpose and another one uses eye movement tracking in human identification.
Mobile devices, such as smartphones and tablets, are frequently used for creation and transmission of private and sensitive messages and files. While personal identification numbers and passwords have been the standard for mobile device security, users tend to forget complex character combinations or reuse them for multiple accounts. These disadvantages have caused researchers to explore biometric authentication for accurate and convenient mobile device security by taking advantage of refined sensing technologies which capture environmental, positional, and interactive information. This information has been found useful in uniquely modeling physical and behavioral characteristics such that biometric recognition is feasible. In this paper, over 100 biometric approaches to mobile device security are surveyed. The advantages and research challenges associated with ten biometric modalities are provided, along with discussion of various commercial implementations and biometric template protection schemes.
Many techniques exist for describing shapes. These techniques almost exclusively consider the contour or the inside of the shape; the major problem for describing the outside of a shape, or inverse shape, being that it has an infinite extension. In this paper, we show how to adapt two shape descriptors , one region based, the Cover By Rectangles, and one transform based, the Zernike moments, to be applicable to the inverse shape. We analyze their properties, and show how to deal with the infinite extension of the inverse shape. Then, we apply these descriptors to shape classification and compare representations that use the shape, its inverse, or both. Our experiments establish that, for shape classification, a representation integrating the inverse shape often outperforms a representation restricted to the shape. This opens the path for better techniques that could use, as a rule of thumb, both the representations of a shape and its inverse for the purpose of classification.
In this paper, an efficient skew estimation technique based on iterative employment of the Piece-wise Painting Algorithm (PPA) on document images is presented. The PPA is initially employed on the document image in both horizontal and vertical directions to obtain two horizontally and vertically painted images. A small number of horizontal (vertical) regions, called horizontal (vertical) bands, are then selected from the horizontally (vertically) painted images. Top (left), middle (middle) and bottom (right) points of the horizontal (vertical) bands are identified in 6 separate lists. A linear regression and a geometric line drawing method are applied on the selected points in each list and subsequently two fit lines are drawn. A voting approach based on statistical mode of angles obtained from the fit lines is also proposed to find the best-fit line amongst all the lines. Based on the slope of the best-fit line, the skew angle of the document image is finally estimated and the document skew is corrected. This process is iteratively applied until the estimated skew is less than I. The proposed technique was tested extensively on three different datasets containing various categories of document images and encouraging results were obtained.
Reduction of feature vector dimension is a problem of selecting the most informative features from an information system. Using rough set theory (RST) we can reduce the feature vector dimension when all the attribute values are crisp or discrete. For any information system or decision system, if attributes contain real-valued data, RST cannot be applied directly. Fuzzy-rough set techniques may be applied on this kind of system to reduce the dimension. But, Fuzzy-rough set uses the concept of fuzzy-equivalence relation, which is not suitable to model approximate equality. In this paper we propose a new alternative method to reduce the dimension of feature vectors of a decision system where the attribute values may be discrete or real or even mixed in nature. To model approximate equality we first consider the intuitive relationship between distance measure and equality. Subsequently we fuzzify the distance measures to establish the degree of equality (or closeness) among feature vectors (objects or points). Finally we use the concept of a cut to obtain equivalence relation based on which dimension of feature vectors can be reduced. We also compare the performance of the present method to reduce the feature vector dimension with those of principle component analysis (PCA), Kernel Principal Component Analysis (KPCA) and independent component analysis (ICA). In most of the cases the present method performs same or even better than the other methods.
Fuzzy C-Means (FCM) algorithm is widely used for unsupervised image segmentation. However, the FCM algorithm does not take into account the local information in the image context. This makes the FCM algorithm sensitive to additive noise degrading the image pixels features. In this paper, an approach to incorporating local data context and membership information into the FCM is presented. The approach consists of adding a weighted regularization function to the standard FCM algorithm. This function is formulated to resemble the standard FCM objective function but the distance is replaced by a new one generated from the local complement or residual membership. The applied regularizing weight is a constant weight or alternatively an adaptive one. The adaptive weight is the Euclidian distance between the center prototype and the local image data mean. The regularizing function aims at smoothing out additive noise and biasing the clustered image to piecewise homogenous regions. Simulation results of clustering and segmentation of synthetic and real-world noisy images have been presented. These results have shown that the presented approach enhances the performance of the FCM algorithm in comparison with the standard FCM and several previously modified FCM algorithms.
Indian sign language interpretation is an important task to facilitate communication among Indian deaf community and other people. Dynamic hand gesture recognition among other gesture modalities is a major step towards sign language recognition as well as any human computer interaction applications. The main focus of this paper is to design and develop a new algorithm for Indian sign language sentence creation considering limitation of continuous sign language recognition. This paper explores two algorithms for word recognition. Rule based and Dynamic Time Warping-based methods for Indian sign language word recognition are developed. The Dynamic Time Warping-based method gave better accuracy for continuous word recognition than the rule-based method. The proposed new approach for Indian sign language sentence interpretation using inverted indexing overcomes the challenges of conventional continuous sentence recognition in sign language interpretation.
This paper presents a new technique for incorporating local membership information into the standard fuzzy C-means (FCM) clustering algorithm. In this technique, the objective consists of minimizing the classical FCM function with a unity fuzzification exponent plus a weighted proposed fuzzification and regularization term. The pixel to cluster-center distance is weighted using the reciprocal of the local membership average. The regularization term is formulated using the Kullback-Leibler (KL) divergence which measures the proximity between a pixel membership and the local average of this membership in the immediate neighborhood. Therefore, minimizing this KL divergence biases the cluster membership of the pixel toward the local membership average. It is also shown that the proposed weighted distance further leads to assigning a pixel to the cluster more likely existing in the immediate neighborhood. This can provide immunity against noise and results in clustered images with piecewise homogeneous regions. Results of clustering and segmentation of synthetic and real-world images are presented to compare the performance of the proposed local membership based weighted distance and KL divergence FCM (LMWD-KLFCM) and the standard FCM, a local data based information FCM (LDMFCM) and a type of local membership information based FCM (LMFCM) algorithms.
This is about frames of reference and their relation to classification. A classification is needed to establish a frame of reference and, with a frame of reference, measurements and further classifications are possible. The topic is discussed in terms of a method that chooses the "best" among alternative frames of reference. We will describe how measurements induce a fiber bundle that projects from a total space of objects onto a base space of measurement values. Local inverses of this projection, or "sections" of the fiber bundle play the role of frames of reference or ideal objects that are attached to the data, as the nearest neighbor in the fiber. In this formalism the invariant properties of personality are parameterized by the variant ones, which are measured, and classification is seen as inverse to measurement. Rather than proving theorems, the article has two goals: to provide engineers with a recipe for solving classification problems; and to bring the concept of moving frames from differential geometry into a broader discussion of classification.
A simple power-logarithm histogram modification operator is proposed to enhance digital image contrast. First a logarithm operator reduces the effect of spikes and transforms the image histogram into a smoothed one that approximates a uniform histogram while retaining the relative size ordering of the original bins. Then a power operator transforms this smoothed histogram into one that approximates the original input histogram. Contrast enhancement is then achieved by using the cumulative distribution function of the resulting histogram in a standard lookup table based histogram equalization procedure. The method is computationally efficient, independent of image content, and (unlike most existing contrast enhancement algorithms) does not suffer from the occurrence of artifacts, over-enhancement and unnatural effects in the processed images. Experimental results on a wide range of different images show that the proposed method effectively enhances image contrast while preserving overall image brightness, and yields results that are comparable or even of higher quality than those provided by previous state-of-the-art methods.
In this paper, we propose a method for recognizing human actions indoors using fuzzy rules and multi cameras. To recognize the human actions, initially, we use the background difference method to extract human area candidates. We then extract HOG features and learn to detect humans using the features and AdaBoost. Fuzzy rules are then used of detect the human actions. The detected human is determined to be stationary or not using the distance between the detected areas in consecutive frames. We also estimate the direction the human is facing using the width of detection, and finally recognize the standard action using the height of the detected region. In addition, we recognize suspicious action using duration of detection and presence of abandoned object. After experiments, recognition accuracy achieved for “walking” and “stop” actions is about 87%, for “running” action about 54%, for “sitting” about 96%, for “desk working” about 83%, and “falling” about 88%.
A 2-D shape description using the turning angle is presented 1. This descriptor is based on a scalogram obtained from a progressive filtering of a planar closed contour. At a given scale, the differential turning angle function is calculated from which, three essential points are derived: the minimum differential-turning angle (alpha-points), the maximum differential turning angle (beta-points) and the zero-crossing of the turning angle (gamma-points). For a continuum of the scale values in the filtering process, a map (called d-TASS map) is generated. As shown experimentally in a previous study, this map is invariant under rotation, translation and scale change. Moreover, it is shearing and noise resistant. The contribution of the present study is firstly, to prove theoretically that d-TASS is rotation and scale change invariant and secondly to propose a new descriptor extracted from the blocks within the scalogram. When applied to shape retrieval from commonly used image databases like MPEG-7 Core Experiments Shape-1 dataset, Multiview Curve Dataset and marines animals of SQUID dataset, experimental results yield very encouraging efficiency and effectiveness of the new analysis approach and the proposed descriptor.
Identification of body parts is an important first step for many tasks such as action recognition in automatic surveillance systems. In this paper, we present a body part segmentation system for image and video analysis. The proposed system utilises Hidden Markov Models and modified shape context features for statistical modeling of the human body shape. In our solution, we also demonstrate how a general and robust solution can be developed with the synthetically generated training data. The sequences of synthetically generated images are generated using three dimensional rendering and motion capture information. After the training phase, the model is used to segment silhouette images into four body parts; arms, legs, body and head. In experiments, the system is successfully used in body part segmentation, unusual activity detection in surveillance applications and arm swing detection in gait analysis. The advantages of the method include that the same model can be employed without any modifications of parameters after initial training.
Although Support Vector Machine (SVM) is widely used in practice, it only takes the boundary information between classes into consideration while neglects the data distribution, which seriously limits the classification efficiency. In view of this, Minimum Class Variance Support Vector Machine (MCVSVM) is proposed by Zafeiriou. Compared with SVM, MCVSVM has better generalization ability because it takes both boundary information and distribution characteristics into consideration. While the above mentioned methods SVM and MCVSVM always neglect the local characteristics of each class. Based on the above analysis, this paper presents Minimum Manifold-based Within-Class Scatter Support Vector Machine (M 2 SVM), which not only focuses on boundary information and distribution characteristics, but also preserves the manifold structure of each class. By theory analysis, M 2 SVM is equivalent to SVM and MCVSVM in a certain condition. It is believed that compared with SVM and MCVSVM, M 2 SVM has the best generalization ability. Experiments on the man-made dataset and UCI datasets verify the effectiveness of the proposed method M 2 SVM.
