| 1. IntroductionWith the social reform and opening up and the rapid development of commodity economy, economic crimes have increased year by year. In the field of Document Examination ,off-line signature verification plays an increasing role in the case of the detection process, part of it is maliciously imitating other people's signature individually. Part of it has been fully practised and high-quality handwriting imitation,so it poses a challenge to the Document Examination workers.It appears somewhat unable to use the traditional character contrast method to test and identify the increasingly'realistic'imitation handwriting.Therefor,proposing a new method of off-line signature verification and how to use computers for accurate, rapid identification of the off-line signature has become urgent problems.As for the off-line signature verification, how to choose the characteristics of good and describe the characteristics of the choice to give the descriped characteristics a relative stability ,and through such a characterization model ,reflecting the bilolgical characeristics of the writer rather than the Chinese itself ,is the first problem to solve for the off-line signature verification.2. Entropy-based pattern selection methodThis article first reviewed the pattern recognition unified information entropy theory. The theory uses the definition of information entropy to conduct an overall introduction on the pattern recognition system, point out the importance of a good feature embodiment selection to the pattern recognition system clearly and indicate qualitatively the information entropy embodiment of characteristics, and it has a guiding significance on how to carry out feature selection. However, in this article, there is no detailed, quantitative introduction in practical applications, how to use the concept of information entropy to select characteristics .Therefore, this paper puts forward a kind of pattern selection method based on information entropy , by the proposed solutions , solving feature selection problem in the following several cases.(1)Different characteristic collections provide different system mutual information.(2)Several characteristic collections provide the same system mutual information. It can be classified as:①Conditions characteristic entropies are different from each other.②The smallest conditions characteristic entropies of several characteristic collections are the same.At the situation②,there shows the math expression of the characterization ability of a single character to a model category and the one of the average mutual information provided for the whole model category of collections by a single character. The aim is to give an further analysis of the several characteristic collections which have the same characterization ability in general, then get the expression of math expression of the characterization ability of a single character to a model category.3. The description method for off-line signatureSignature is divided into on-line signature and off-line signature in general. Compared with on-line signature, the off-line signature has lost the dynamic information, such as the order of strokes, pen pressure and so on. So it's more difficult to identify of static off-line signature than online one. Recently, to verify the writer of the Chinese off-line signature, we always extract a number of features of the handwritten signature, and then select the appropriate pattern recognition approach according to the different characteristics. However, handwritten signature is such a special problem of pattern recognition that any feature couldn't play the role independently, for only the inherent contacts among the characteristics can reflect the biological characteristics of the signer naturally. Based on this idea, this paper offers a way to describe the characterization of off-line signature which helps us obtain the signature contained in the intrinsic relationship between the characteristics. As the relationships of different signer represent differently, we can distinguish different signature with these relationships. By far, this approach has changed the former way, which takes the characteristic as importation of classifier directly.In details, the method which describes the characteristics of off-line handwritten signatures is under the premise of having no priori samples, and as a result, we should first abstract all of the possible characteristics in the signatures to get several numbers of features. With these features, we create a set of features, which is referred to as the template characteristics in the paper. It's worthy to note that the template theory of characteristics can represent all the signatures effectively. We'd better set up a space of feature weight with weight coefficients, which are acquired through describing the characterization of each signer, to shows the contribution of each character. Because the macroeconomic performances between individual signatures are different, we can get different combinations characteristics when using the characteristics in the template to illustrate signatures.In other words,the combination of personal signatures and the characteristic weight coefficients is a one-to-one mapping relation ,which means that each signature can establish a mapping of a point of the characteristics weight space, and then, we can distinguish the identifies of different writers according to the'distance'of different points in the space .In other words the characteristics of individuals with Signature portfolio weights are 11 mapping relation, that is, each signature can be mapped into a space on the characteristics of the weight of a point, then according to the different points of space between the "distance" to the different Writing the identity for the difference.It is worthy of note that this part is the core of the article. In this section, we try to adjust the characteristic differences in degrees according to the stabilization and the characterization ability in the parties'signatures. That is reducing the difference degree of the instability characteristics (the identification interference characteristics) and retain the one of the stability characteristics.2. Simulation resultsAccording to the put forward feature selection algorithm and static signature feature description algorithm, the relative position characteristics of the largest pixel and the distribution direction characteristics of the main energy will be picked up which belong to the parties and the replica's signatures. Then according to the algorithm the extracted characteristics will be processed and analysed. From the data obtained from the experiment we can conclude that without considering the stabilization and the characterization contribution specific feature of a certain specific feature in the parties'signatures ,the difference degrees on a certain characteristic of the two separately are 0.0600(the relative position characteristics of the largest pixel)and 0.2476(the distribution direction characteristics of the main energy).Here the latter difference degree is bigger. While we consider the stabilization and the characterization contribution specific feature in the parties'signatures, the difference degrees on a certain characteristic of the two separately are 0.0599(the relative position characteristics of the largest pixel)and 0.0139(the distribution direction characteristics of the main energy).Here the latter difference degree get small and is even smaller than the one of the relative position characteristics of the largest pixel)and 0.2476(the distribution direction characteristics of the main energy. It fully illustrates the characterization algorithm put forward in this paper. The characteristic differences in degrees can be adjusted according to the stabilization and the characterization ability in the parties'signatures. Therefore, when we measure the difference degree, some difference degree interferences of non-essential characteristics can be eliminated. |
PDF Full Text Request