A distance mapping pattern classification method

Abstract

The invention relates to a distance mapping pattern classification method. The established distance mapping classifier comprises four parts, namely an input layer, a hidden layer 1, a hidden layer 2 and an output layer. The input layer comprises d neuron nodes, and each node represents a feature vector of a training or test sample. The feature vector set of the extracted samples is used as the input layer of the distance mapping classifier. The hidden layer 1 contains N neuron nodes, N representing the total number of training samples; and calculates the Euclidean distance between the input eigenvector and the eigenvectors of N training samples as the hidden layer 1. The hidden layer 2 consists of L neuron nodes. The hidden layer 1 and the hidden layer 2 can be connected by linear activation function. The two parameter matrices of the linear activation function are the connection weights and offsets of the hidden layer 1 and 2 respectively. When the classifier model is used to classifythe test samples, the output matrix can be calculated directly by using the transformation matrix obtained from the training.

Description

technical field [0001] The invention relates to a method for (artificial intelligence) pattern recognition and classification. Background technique [0002] At present, pattern classification is widely used in various industries such as medicine, intelligent transportation, weather forecast, and liquor recognition. To this end, many pattern classification methods have been presented. Generally, pattern classification methods can be summarized into two types, namely simple classifiers and complex classifiers. Simple classifiers, such as k-nearest neighbors, Bayesian, Fisher, and extreme learning machines, are suitable for fast classification requirements; while complex classifiers, such as support vector machines, neural networks, and deep learning, have poor real-time performance. [0003] Different classifiers have different characteristics. Take two simple classifiers as examples, namely k-nearest neighbor {B.K.Samanthula, Y.Elmehdwi, and J.Wei, "k-nearest neighbor clas...

AI Technical Summary

Problems solved by technology

[0004] The traditional k-nearest neighbor is to select k samples that are close to the test sample in all training samples. This method has two shortcomings: one is that it only considers the role of some training samples that are close to the test sample, while ignoring the effects of other training samples. Second, different k values ​​have a greater impact on classification accuracy, and continuous training is required to determine an ideal k value

Images