Resumen
Method and system for verifying dynamic handwriting and signatures by means of deep learning. The method comprises the following steps:extraction of temporal functions (110) from a set of samples consisting of at least one recorded sample (S<sub>i</sub>) belonging to a certain identity and a doubted test sample (S<sub>T</sub>), obtaining at least one recorded pattern (TF<sub>i</sub>) and a test pattern (TF<sub>T</sub>);temporal alignment (120) of the at least one recorded pattern (TF<sub>i</sub>) with the test pattern (TF<sub>T</sub>), obtaining at least one aligned recorded pattern (TAF<sub>i</sub>) and an aligned test pattern (TAF<sub>T</sub>),comparison of the aligned patterns (TAF<sub>i</sub>, TAF<sub>T</sub>) by using a previously trained recurrent neural network (130), obtaining a similarity measure (132) between the at least one recorded sample (S<sub>i</sub>) and the test sample (S<sub>T</sub>); andverification of the identity (140) of the test sample (S<sub>T</sub>) based on a comparison of the similarity measure (132) with a threshold value.
Reivindicaciones
1. A method for verifying dynamic handwriting and signatures, characterised in that it comprises the following steps: extraction of temporal functions (110) from a set of samples consisting of at least one recorded sample (S<i>) belonging to a certain identity and a doubted test sample (S<T>), obtaining at least one recorded pattern (TF<i>) and a test pattern (TF<T>); temporal alignment (120) of the at least one recorded pattern (TF<i>) with the test pattern (TF<T>), obtaining at least one aligned recorded pattern (TAF<i>) and an aligned test pattern (TAF<T>); comparison of the aligned patterns (TAF<i>, TAF<T>) by using a previously trained recurrent neural network (130), obtaining a similarity measure (132) between the at least one recorded sample (S<i>) and the test sample (S<T>); and verification of the identity (140) of the test sample (S<T>) based on a comparison of the similarity measure (132) with a threshold value. 2. The method according to claim 1, wherein the comparison of the aligned patterns (TAF<i>, TAF<T>) comprises: for each recorded sample (S<i>), comparing the aligned recorded pattern (TAF<i>) corresponding to said recorded sample (S<i>) with the aligned test pattern (TF<T>) by using a recurrent neural network (130) with two inputs; obtaining a similarity measure (132) by means of computing a value using the comparisons performed. 3. The method according to claim 1, wherein there is a plurality N of recorded samples (Si), and wherein the comparison of the aligned patterns (TAF<i>, TAF<T>) comprises comparing a plurality N of aligned recorded patterns (TAF<i>) corresponding to the recorded samples (Si) with the aligned test pattern (TAF<T>) by using a Siamese recurrent neural network (130) with N+1 inputs. 4. The method according to any of the preceding claims, comprising a previous step of training the recurrent neural network (130) iteratively by minimising a cost function, using a plurality of training sets (202) formed from a sample database (200) which stores a set of recorded samples (S<i>) and test samples (S<T>) corresponding to different identities, wherein each training set (202) comprises one or more recorded samples (S<i>) corresponding to one same identity I and a test sample (S<T>) corresponding to said identity I or to a different identity. 5. The method according to any of the preceding claims, wherein the extraction of temporal functions (110) is performed starting from dynamic information acquired for each sample, wherein the dynamic information includes at least temporal samplings of the X and Y coordinates of the sample. 6. The method according to claim 5, wherein the basic dynamic information includes temporal samplings of the pressure (P) of the sample. 7. The method according to any of the preceding claims, wherein the temporal alignment (120) is performed by using a temporal alignment algorithm. 8. The method according to any of the preceding claims, wherein the verification of the identity (140) of the test sample (S<T>) comprises: determining that the identity of the test sample (S<T>) is the same as the identity of the at least one recorded sample (S<i>) if the similarity measure (132) exceeds the threshold value; determining that the identity of the test sample (S<T>) is different than the identity of the at least one recorded sample (S<i>) if the similarity measure (132) does not exceed the threshold value. 9. The method according to any of claims 1 to 8, wherein the samples are dynamic handwritten signatures (302). 10. The method according to any of claims 1 to 8, wherein the samples are dynamic handwriting (304). 11. A system for verifying dynamic handwriting and signatures, comprising at least one memory, at least one programme stored in memory, and one or more processors configured to execute at least one programme, characterised in that the at least one programme includes instructions for carrying out the method of claims 1 to 10. 12. A programme product for verifying dynamic handwriting and signatures, comprising programme instruction means for carrying out the method defined in any of claims 1 to 10 when the programme is executed on one or more processors. 13. The programme product according to claim 12, stored on a programme support medium.
