The Improvement Of PNN Pattern Recognition Method Combined MrFCM And DE Algorithm

The disadvantage of parameter model method traditional in algorithm of pattern recognition is that the assumed probabilistic distribution function didn't match the data of classification. Non- parameter model method, take the Probabilistic Neural Networks (PNN) for example, can overcome the disadvantage of parameter model, but when encountering the training sets with large amounts and more dimentions, the lower classification speed and huge storage make it impractical in data classification. So, the article does some useful works for the two major disadvantages.Because the hidden arrangement function of PNN which use the same parameter which can lead to reduce the precision of the last recognition, so we provide an improved method which is DEPNN method, we also give the process of the algorithm, the whole process is divided into two parts, first we distiss the character of the recognition problem, then we use the Probabilistic Neural Networks method to recognize. The experiment result shows that the DEPNN method is more effective than PNN method.Because of the other disadvantage of PNN method which has too much consumption of memory when confronted a mass of and more dimentional training samples, the article uses the Fuzzy C Means method to do some preliminary work to the training samples which would train the PNN neural net work. Though the FCM method have the excellent clustering ability, it depends on the initialization of the pattern, the result of clustering would easily trap into the part extremum which is not the extremum of the whole model. To keep the last precision of the PNN recognition method article adopt the MrFCM cluster algorithm to save the memory, the MrFCM built on the FCM cluster method, it divided the process of the clustering into two parts, as we call Multisage random sampling FCM(MrFCM) algorithm, the experiment show the MrFCM method is more effective than FCM method.At last, article give the MrFCM-DEPNN recognition method after we improve the disadvantages of the PNN recognition method. Based on DEPNN and MrFCM method the passage gives MrFCM-DEPNN pattern recognition method, and gives the serial structure and steps of the MrFCM-DEPNN algorithm, it based on the improvement of FCM algorithm and PNN algorithm, using PNN to do probabilistic classfication based on Bayes confidence measure, this combined use of FCM and PNN speeds up the PNN evaluation significantly and increases its accuracy as well. The experiment result demonstrate that MrFCM-DEPNN recognition algorithm is a faster and accurater method than traditional method for more dimentional data.