Éditeur(s) : HAL CCSD
Résumé : International audience
In this paper, we present a novel method for recognition of handwritten sketches. Unlike previous approaches, we focus on online retrieval and ability to build our model incrementally, thus we do not need to know all the data in advance and we can achieve very good recognition results after as few as 15 samples. The method is composed of two main parts: feature representation and learning and recognition. In feature representation part, we utilize SIFT-like feature descriptors in combination with soft response Bag-of-Words techniques. Descriptors are extracted locally using our novel sketch-specific sampling strategy and for support regions we follow patch-based approach. For learning and recognition, we use a novel technique based on fuzzy-neural networks, which has shown good performance in incremental learning. The experiments on state-of-the-art benchmarks have shown promising results.
17th Biennial Conference of the International Graphonomics Society
Pointe-à-Pitre, Guadeloupe

hal-01166497
https://hal.univ-antilles.fr/hal-01166497
https://hal.univ-antilles.fr/hal-01166497/document
https://hal.univ-antilles.fr/hal-01166497/file/IGS_2015_submission_37.pdf

Éditeur(s) : HAL CCSD
Résumé : International audience
We present a novel approach towards signature recognition from spatio-temporal data. The data is obtained by recording gyroscope and accelerometer measurements from an embedded pen device. The idea of Universum learning was previously presented by Vapnik and recently popularized in machine learning community. It assumes that the decision boundary of a classifier lies close to data with high uncertainty. The quality of the final classifier strongly depends on a way how to choose the Universum data and also on the representation of original data. In our paper we use a novel approach of Universum learning to classify signature data, also we present our novel idea how to sample the Universum data. At last, we also find more effective representation of the signature data itself compared to the baseline method. These three novelties allow us to outperform previously published results by 4.89% / 5.58%.
17th Biennial Conference of the International Graphonomics Society
Pointe-à-Pitre, Guadeloupe

hal-01165925
https://hal.univ-antilles.fr/hal-01165925
https://hal.univ-antilles.fr/hal-01165925/document
https://hal.univ-antilles.fr/hal-01165925/file/IGS_2015_submission_38.pdf