Signature recognition

About: Signature recognition is a research topic. Over the lifetime, 2138 publications have been published within this topic receiving 37605 citations.

Non-Intrusive Load Monitoring: A Multi-Agent Architecture and Results

Abstract: Non-intrusive Load Monitoring (NILM) is a technology that allows the identification of individual electrical loads from a single aggregated measurement of voltage/current, hence, useful for diagnostic of the consumption of electrical energy. This is performed by means of load detection and disaggregation techniques, as there are several different power signatures from the active loads. This paper proposes a multi-agent architecture and evaluates its performance. Four detection methods were selected: Discrete Wavelet Transform (DWT); Kalman Filter; Derivatives, and Half Cycle Active Power. For the power signature recognition agents, different feature extractors and machine learning methods were evaluated: Discrete Fourier Transform, DFT with Exponential Damping, V-I Trajectories, Wavelet and Power Envelope; these were combined with four classifiers: k-Nearest Neighbors, Ensemble Method, Support Vector Machine and Decision Tree. The detection, feature extractors and classifiers methods were tested using waveforms sampled in real situations of the electric network available at the COOL and LIT datasets. The results for the proposed multi-agent architecture indicates improvements in the event detection, lower occurrences of false positives, better feature extractions and better classification results.

Flexible enhanced fuzzy min–max neural network model for pattern classification problems

Abstract: In the attempts of building an efficient classifier model, various hybrid computational intelligence models have been introduced. Among these, the enhanced fuzzy min-max (EFMM) model was one of the most recent models coming with many essential features like the ability to provide online learning processes and handling the forgetting problem. Although EFMM has been proven to be one of the most premier models for undertaking the pattern classification problems, issues related to its learning process, concerning the overlap between the hyperboxes, random expansion coefficient value (user-defined) and hyperbox contraction remain unsolved. Therefore, two stages of improvements are introduced in this research to overcome the current limitations and improve classification performance in terms of accuracy and complexity. In the first stage, a new flexible enhanced fuzzy min-max (FEFMM) model is proposed to overcome limitations related to accuracy issue. Hence, four new procedures are introduced. First, a new training strategy to avoid generating unnecessary overlapped regions. Second, a new flexible expansion procedure to replace the expansion coefficient user-defined parameter with a self-adaptive value to produce more accurate decision boundaries. Third, a new overlap test rule is applied during the testing phase to identify any possible containment overlap case and activate the contraction process (if necessary). Fourth, a new contraction procedure to overcome the containment overlap and avoiding the data distortion problem (missing hyperbox information). In the second stage, a new pruning strategy is proposed to further enhance the performance of the proposed model in regards to overcome the network complexity problem. Hence, the resulting model is known as FEFMM-based pruning strategy (FEFMM-PS). The usefulness of both stages is evaluated systematically using a series of experiments using several benchmark datasets. Sixteen data sets are used in the evaluation process. These data sets are obtained from the UCI machine learning repository and the selection of these data sets is related to cover examples of different levels of difficulties, input and output classes, features, and a number of instances. The performance of FEFMM-PS in these experiments are then quantified using statistical measures where the bootstrap and k-fold cross-validation methods have been adopted. The results demonstrate the efficiency of FEFMM in handling pattern classification problems and providing a superior performance of classification accuracy as compared to the other network structures from the same variants such as EFMM, FMM variants and also non-FMM related models. Concerning the FEFMM-PS, the finding reveals that the model (FEFMM-PS) is able to solve network complexity problem and presents better classification accuracy as compared to FEFMM and other models from the literature. The proposed models FEFMM and FEFMM-PS can be applied in several application areas to further assess their applicability, such as face recognition, speaker recognition, signature recognition, and text classification.

Estimating acoustic-labial weights in connected speech recognition systems based on HMM

Abstract: Describes an approach for weighting the contribution of the acoustic and visual sources of information in a bimodal connected speech recognition system. We consider that a different acoustic-labial weight is attached to each recognition unit. The values of the weighting vector are optimised in order to minimise the error rate on a learning set. Experiments are performed on a two-speakers audiovisual database, composed of connected letters, with two different acoustic-labial speech recognition systems. For both speakers and both systems, the weights optimisation allows us to increase the recognition rate of our bimodal system.

Determining valuable ranges of handwritten signature using fuzzy approach and window method

Abstract: The paper proposes possible improvements in signature recognition approach based on window method. The analysis focuses on a stage of window preprocessing using fuzzy sets in order to choose significant ranges of each signature. Proposed extension allows the solution to improve in two areas. First of all minimizing a number of processed windows significantly reduces computation time. Secondly, filtered signatures with valuable information about significant ranges allow the system to recognize signatures of a poor or good quality. Developed method of signature quality assessment can be used in any signature recognition system, regardless of used method of analysis. Merging the information about signature quality and choosing only important signature ranges should also improve the overall detection results, however, more examinations are needed to confirm this statement.