Lukas Tencer

Online handwritten gesture recognition based on Takagi-Sugeno fuzzy models

In this paper, we present a new method for incremental online handwritten gesture recognition based on fuzzy rules. This approach allows starting from a scratch with no previously learned classes and adding new ones lifelong. Unlike methods based on evolving mountain clustering, our approach suits incremental concept better. We introduce a new method for evolving clustering and usage of incremental density measurement for determining the membership function which significantly improves the results. Density measurement as membership function allows using only few parameters instead of the costly covariance matrices and does not require any estimating by averaging and thus preventing from information lost. We also introduce a new set of features based on a shape of gestures. Combination of these new system characteristics thus lowers memory and computational requirements while significantly increasing recognition rate.

ARTIST: ART-2A Driven Generation of Fuzzy Rules for Online Handwritten Gesture Recognition

Incremental learning, especially when learning from a scratch, has a lot of interest for online gesture recognition. However the lack of learning examplers combined to low computational cost suggests building robust and efficient learning machines. In this paper we introduce a hybrid model of ART-2A neural network combined to Takagi-Sugeno (TS) neuro-fuzzy network. The latter model is applied for online handwritten gesture recognition, when the learning is starting from scratch and no class information, such as gesture type or number of classes, is predefined. Moreover, using ART-2A neural network and our novel distance measure, the computational complexity of the whole model decreases while preserving high accuracy. Furthermore, we exploit the forgetting dilemma of online learning by introducing a competitive Recursive Least Squares method for TS models. Together, all the modeling has shown promising results.

TITS-FM

Graphical abstractDisplay Omitted HighlightsWe developed transductive inference model for Takagi-Sugeno fuzzy models.We introduce a novel model for transductive similarity.Transductive similarity improves the accuracy of TS fuzzy models for classification.Developed transductive similarity model is not limited only to TS fuzzy models.Developed model improves the precision of real world recognition tasks. In this paper we present a novel model originating from Takagi-Sugeno fuzzy models. It is based on a concept of transductive similarity, where unlike a simple inductive similarity, it considers also local neighborhood of a given element. Transductive property of a local space is used in an inference process, what allows the technique to be used also in incremental settings. Since incremental model construction brings new challenges, we are unable to use the offline transductive approach as some of the previous works did. The key idea of our model is to adjust activation properties of each rule, based on cross-rule similarities. Our method is capable of using the transductive property for any metric. Besides the final model, we also present several improvements to the transductive similarity technique itself, where we alternate the similarity metric in several ways to better exploit the influence of local neighborhood in the final metric. At the end, we demonstrate a superior performance of our technique over the state-of-the-art techniques build on TS fuzzy models on several machine learning datasets.

Sketch-Based Retrieval of Document Illustrations and Regions of Interest

In this paper we present a novel approach towards retrieval of documents with pictorial data. Many prior works focused on word-spotting and text-based retrieval, but none of these techniques handled the retrieval of pictorial part of documents. In this paper, we present a new method that allows users to retrieve any visual data from documents, based on a sketch example. It mainly emphasizes all three main aspects of a visual retrieval system: feature representation, indexing and retrieval. Especially we focus on the design of salient descriptors, capable of capturing unique mapping from sketched images to document illustrations. We evaluate several approaches towards feature representation and indexing, with the aim to maximize the performance of our method. Our proposed technique is highly useful to complement word-spotting technique, when the indexed documents are composed of mixture of visual and textual data. This technique has shown promising results, both on pictorial data automatically extracted from documents as well for those selected by users as regions of interest.

A new framework for online sketch-based image retrieval in web environment

We present a novel framework for retrieval of images, based on the users sketches performed in web environment. Unlike previous approaches, our method is capable of online retrieval and provides balanced trade-off between computational cost and robustness, while still preserving local properties. In our approach novel combination of features is introduced, which describes properties of desired results and query images. Based on the extracted features we use nearest neighbor recognizer, trained on dynamic neighborhoods, in combination with k-means and k-D tree technique for further speedup. A novel technique for online retrieval, based on sequential input processing and partial hierarchical score evaluation, is introduced, allowing us to suggest on the fly entries based on in-progress sketch. We tested our method on small and large scale databases and achieved promising results. The solution itself is implemented in collaborative environment, what allows users to produce queries cooperatively.

Incremental Similarity for real-time on-line incremental learning systems

New Incremental Similarity for incremental learning problems.Handling the processing time vs. accuracy issue, the learning on the fly as well as the lack of data at the beginning of the learning.Incorporation of Incremental Similarity into various classification models.Extensive comparison and evaluation of various models using our incremental learning framework. The expectation of higher accuracy in recognition systems brings the problem of higher complexity. In this paper we introduce a novel Incremental Similarity (IS) that maintains high accuracy while preserving low complexity. We apply IS to on-line and incremental learning tasks, where the need of low complexity is of significant need. Using IS enables the system to directly compute with the samples themselves and update only few parameters in an incremental manner. We empirically prove its efficiency on several evolving models and show that by using IS they achieve competitive results and outperform the baseline models. We also consider the problem of incremental learning used to handle fast growing datasets. We present a very detailed comparison for not only evolving models, but also for the well-known batch models, showing the robustness of our proposal. We perform the evaluation on various classification problems to show the wide application of evolving models and our proposed IS.

SO-ARTIST

Graphical abstractDisplay Omitted HighlightsWe developed self-organized Takagi-Sugeno model for online learning, starting from the scratch, adding new classes any time and independent on database knowledge, not knowing the number of classes.We introduce the solutions for merging and splitting of antecedent and consequent parameters.Our merging and splitting solutions are limited for our novel incremental distance measurement (antecedent part) and recursive least squares (consequent part).We introduce the automated organization of rule and decision making behind it. In this paper we introduce a novel online self-organized clustering method based on the ART-2A network for Takagi-Sugeno fuzzy models. To accomplish the self-organization, we introduce an automatic decision algorithm along with solutions for merging and splitting of rules as well as the parameters they operate with, such as our novel incremental distance measurement and competitive recursive least squares. We emphasize the learning algorithms having an impact for initial as well as long-term learning capabilities. We also emphasize the challenge for online learning, where examples are incoming in real-time and thus are unknown before they can be learned. Therefore, we solve parameter fixing by introducing a parameter free method. We show the performance of our method on various machine learning benchmarks as a highly accurate and low time-consuming method capable of adapting to different databases without the need for fixing any of its parameters according to the database.

Forgetting of unused classes in missing data environment using automatically generated data: Application to on-line handwritten gesture command recognition

Abstract In this paper we exploit the use of synthetic data for on-line handwritten gesture commands recognition with an emphasis on the problem of forgetting unused classes. For on-line learning, one of the most crucial moments of the processing is the initialization. In some applications the data is available and these can be fed to the learning model. However, in applications such as user-friendly handwritten gesture recognition, this scenario is not possible. Since from the user perspective it is better to let the user define his own symbols, the learning model is lacking in the amount of data at the initialization. Some strategies have been proposed to acquire synthetic handwritten gesture commands and use these for on-line learning. In this paper we exploit this technique further and focus on the forgetting of unused classes by applying a random buffer and Elastic Memory Learning (EML) to avoid this from happening. In the experiments we search for the proper amount of synthetic data produced for each sample as well as exploit the most appropriate time to stop the generation of synthetic data for learning purposes. We also investigate the influence of synthetic data on forgetting when using the proposed EML. We base the generation of synthetic gesture commands on Kinematic Theory.

RETRACTED: UFuzzy: Fuzzy Models with Universum

This article is retracted at the request of the author, as they inadvertently provided the wrong source file for publication. The Authors wish to apologize for this error. The full Elsevier Policy on Article Withdrawal can be found at http://www.elsevier.com/locate/withdrawalpolicy .

Summit-Training

Graphical abstractDisplay Omitted HighlightsWe propose a novel hybrid Semi-Supervised technique (Summit-Training) for classification tasks.Summit-Training combines elements of Semi-Supervised training and Active Learning.We evaluated it on six datasets and applied to 13 classifiers resulting in 78 test cases.It improves performance of all classifiers, outperforms other baseline Semi-Supervised techniques.We use Summit-Training in gesture recognition which is deployed in real-world HCI system. With an explosion in the amount of data available for classification problems it becomes more and more complicated to obtain labels for the whole dataset. Because of this we are frequently presented with only a fraction of labeled data from the whole dataset. To leverage the presence of unlabeled examples Semi-Supervised methods could be used to increase the precision of used classifiers. Therefore, we propose a novel hybrid technique which extends the concept of Self-Training and Help-Training used in Semi-Supervised techniques by incorporating the Active Learning approach for determining the confidence of the classifier in the testing set samples. Specifically we employ the Query-by-Committee (QbC) approach and we call the final method Summit-Training. We apply this method to a range of classifiers (generative and discriminative) and evaluate it on several datasets and real world problems. The formulation of the Summit-Training method especially allows us to use Semi-Supervised approaches for purely discriminative classifiers in which no probabilistic representation of the evaluated classes exists. Compared to other Semi-Supervised techniques (Self-Training and Help-Training), the proposed new method achieves superior performance. It also has better generalization properties since it reduces the number of hyper-parameters and relaxes the conditions for classifiers on which it could be applied.