CN109740679B

CN109740679B - Target identification method based on convolutional neural network and naive Bayes

Info

Publication number: CN109740679B
Application number: CN201910013650.2A
Authority: CN
Inventors: 胡燕祝; 王松
Original assignee: Individual
Current assignee: Individual
Priority date: 2019-01-13
Filing date: 2019-01-13
Publication date: 2020-10-30
Anticipated expiration: 2039-01-13
Also published as: CN109740679A

Abstract

The invention relates to a target recognition method based on a convolutional neural network and naive Bayes, belonging to the field of image processing and pattern recognition and characterized by comprising the following steps: (1) determining coordinates (x ', y ') of translated and normalized picture pixel points in a training set sample A to obtain a new training set A '; (2) training the convolutional neural network and updating the connection weight omega_i(ii) a (3) Extracting a feature vector X; (4) taking the feature vector X as the input of a naive Bayes model, and establishing the naive Bayes model for training; (5) the test set samples are used as input and sent to the combined network for classification to obtain classification results. The target identification method based on the convolutional neural network and the naive Bayes replaces a Softmax classifier in the traditional convolutional neural network with the naive Bayes to realize classification prediction, and fully utilizes the characteristic information of the characteristic vector output by the full-connection layer of the convolutional neural network. Through a plurality of groups of data experiments, the method is accurate and reliable in calculation and relatively stable in data result, and a stable identification method on the basis of ensuring accurate classification is provided for image target identification.

Description

Target identification method based on convolutional neural network and naive Bayes

Technical Field

The invention relates to the field of picture processing and pattern recognition, in particular to a method for recognizing a picture target.

Background

At present, most of technologies cannot achieve high accuracy, are unstable and have poor robustness aiming at the problem of image target identification. Some techniques, while capable of achieving high accuracy, require some complex pre-processing steps. Taking handwritten digit recognition as an example, handwritten digit pictures are often subjected to operations such as expanding a training set by distortion deformation and preprocessing images by simulating various dithering operations, so that high accuracy can be achieved, but the practicability is reduced to a certain extent, and the stability requirement cannot be met. In the classical pattern recognition, features are generally extracted in advance, after a plurality of features are extracted, correlation analysis is carried out on the features to find out the feature which can represent the characters most, and the feature which is irrelevant to classification is removed. However, the extraction of these features is too dependent on human experience and subjective awareness, the different classification performance of the extracted features is greatly affected, and even the order of the extracted features may affect the final classification performance. Meanwhile, the quality of image preprocessing also affects the extracted features.

At present, the digital identification technology in the picture target is often associated with economy and business, and the technology put into use must ensure higher accuracy, because if the identification is wrong, even a very small error can cause a series of business disputes, even bring about huge loss, and cause irrecoverable results. Therefore, in the process of research and development, in order to avoid a series of problems, an accurate and reliable handwritten digit recognition model must be established, so that when the model recognizes handwritten digits, the model simultaneously meets higher accuracy and stronger robustness, meets the requirements of the fields of economy, commerce and the like on handwritten digit recognition, thereby saving time and money, improving efficiency and saving cost.

Disclosure of Invention

In view of the problems in the prior art, the technical problem to be solved by the present invention is to provide an accurate and stable image target identification method, and the specific flow of the method is shown in fig. 1.

The technical scheme comprises the following implementation steps:

(1) determining coordinates (x ', y ') of the translated and normalized picture pixel points in the training set sample A to obtain a new training set A ':

according to the original coordinates (x) of the picture pixel points in the sample set₀，y₀) Maximum x of horizontal coordinate of pixel point after translation operation_maxMinimum value x_minMaximum value y of ordinate_maxMinimum value y_minAnd dx and dy respectively represent the moving size of the pixel point on the x axis and the y axis, and the coordinate of the pixel point of the training sample set is determined as follows:

(2) training the convolutional neural network and updating the connection weight omega_i：

Constructing a convolutional neural network, setting convolutional layers of the convolutional neural networkM, number of pooling layers n, convolutional layer convolution kernel size(s)₁，s₂，...，s_m) Size of the pooling layer (t)₁，t₂，...，t_n) Convolution layer step length r₁Step length r of the pooling layer₂The number N of neurons in the full connection layer, an activation function f and iteration times l. Method for applying back propagation algorithm and BP algorithm to connection weight omega in convolutional neural network_iUpdating:

wherein the content of the first and second substances,

is the actual output value, y_iEta is the step length per update, omega'_iJ is the cost function of the neural network for the updated connection weight.

(3) Extracting a feature vector X:

and reserving the alternately connected parts of the convolutional layer and the pooling layer in front of the fully-connected layer of the convolutional neural network, removing the Softmax layer behind the fully-connected layer, and recording the fully-connected layer of the convolutional neural network as C. And after the training of the convolutional neural network model is finished, the output of the full connection layer C is the extracted feature vector X.

(4) And taking the feature vector X obtained in the last step as the input of a naive Bayes model, establishing the naive Bayes model for training, and realizing the classification of different training samples. Naive bayes rule is as follows:

wherein C is_iIs class, P (C)_i|x_i) Is shown as having x_iIs characterized by being divided into C_iThe probability of (c).

(5) And (4) taking the test set sample as input, sending the test set sample to a combined network for classification to obtain a classification result, and completing the image target identification based on the convolutional neural network and naive Bayes.

Compared with the prior art, the invention has the advantages that:

(1) the invention adopts a mode of combining the convolutional neural network and the naive Bayes, replaces a Softmax classifier in the traditional convolutional neural network with the naive Bayes to realize classification prediction, fully utilizes the characteristic information of the characteristic vector output by the full connection layer of the convolutional neural network, and improves the accuracy.

(2) The invention does not need to carry out complex preprocessing operation on the image, only needs to carry out normalization operation on the image when carrying out feature extraction by using the convolutional neural network, and simplifies the fussy preprocessing flow.

(3) The invention performs experiments on a plurality of groups of data, and compared with the prior art, the experimental results have obvious advantages and the data results are relatively stable. The method improves the stability of the model on the basis of ensuring the accurate classification and can better complete the task of image target identification.

Drawings

For a better understanding of the present invention, reference is made to the following further description taken in conjunction with the accompanying drawings.

FIG. 1 is a flowchart of the steps for building a convolutional neural network and naive Bayes based target recognition model;

FIG. 2 is a flow chart of an algorithm for building a target recognition model based on a convolutional neural network and naive Bayes;

FIG. 3 is a sample of a MNIST dataset of handwritten digital pictures;

FIG. 4 is a simplified neural network architecture diagram;

FIG. 5 is a block diagram of a convolutional neural network;

FIG. 6 is a schematic diagram of a combined network architecture according to the present invention;

FIG. 7 is a comparison of experimental results for various models;

FIG. 8 is a comparison of results of multiple sets of simulation experiments;

detailed description of the preferred embodiments

The present invention will be described in further detail below with reference to examples.

Taking handwritten digit recognition as an example, the dataset chosen is the MNIST public dataset, the samples in the MNIST dataset being shown in FIG. 3, which is a database of handwritten digits established by Corinna cortex of the Google laboratory and by YannleCun, the Koron research institute of New York university. The handwritten digital picture shown in fig. 2 is taken as an original handwritten digital picture data set, a part of the handwritten digital picture data set is marked as a training set, a part of the handwritten digital picture data set is marked as a testing set, and 60000 training sample sets and 10000 testing sample sets are provided. The picture sizes were all 28X 28.

The whole flow of the handwritten number recognition method provided by the invention is shown in figure 1, and the specific steps are as follows:

according to the original coordinates (x) of the picture pixel points in the sample set₀，y₀) Maximum value x of pixel point after translation operation_maxMinimum value x_minDetermining the coordinates of the pixel points of the training sample set:

Constructing a convolutional neural network, setting the number m of convolutional layers of the convolutional neural network as 2, the number n of pooling layers as 2,The number of convolutional layer convolutional cores is 32 and 64, the number of pooling layer cores is 32 and 32, and the convolutional layer step length r₁Is 1, step length r of the pooling layer ₂2, the number N of neurons in the full connecting layer is 200, the activation function f (x) adopts a ReLu function, and the iteration number l is 30. Method for applying back propagation algorithm and BP algorithm to connection weight omega in convolutional neural network_iUpdating:

f(x)＝max(0，x)

wherein the content of the first and second substances,

is the actual output value, y_iFor an ideal output value, η is the step size of each update.

(3) Extracting a feature vector X:

and reserving the alternately connected parts of the convolutional layer and the pooling layer in front of the fully-connected layer of the convolutional neural network, removing the Softmax layer behind the fully-connected layer, and recording the fully-connected layer of the convolutional neural network as C. And after the model training is finished, the output of the full connection layer C is the extracted feature vector X.

wherein C is_iAre of the

classes

0, 1, 2, 3, 4, 5, 6, 7, 8, 9, P (C)_i|x_i) Is shown as having x_iProbability under the feature of being classified into these 10 categories.

(5) And (4) taking the test set sample as input, sending the test set sample to a combined network for classification to obtain a classification result, and completing handwritten digit recognition based on the convolutional neural network and naive Bayes.

In order to verify the accuracy of the invention in identifying the picture target, a plurality of groups of handwritten number identification simulation experiments are carried out on the invention, and the results are compared with some model algorithms for identifying handwritten numbers, wherein the simulation results are shown in tables 1 and 2. The simulation result shows that the handwritten number recognition method established by the invention can achieve higher accuracy rate without complex preprocessing, and has better stability under the condition of ensuring that the accuracy rate is not reduced.

TABLE 1 comparison of experimental results for various models

Experimental methods	Identification accuracy (%)
		Handwritten digit recognition based on convolutional neural network	99.20
Maxout network	99.55
		Handwritten digit recognition based on convolutional neural network and support vector machine	99.60
The invention	99.80

As can be seen from the simulation result table 1, in the case of using the same data set, the recognition accuracy can reach 99.8% after the simple picture pixel normalization preprocessing. Compared with the other three methods, the method has higher accuracy. The target identification method established by the invention is accurate, provides an effective method for establishing an accurate image target identification model, and is more suitable for practical use.

TABLE 2 comparison of multiple sets of simulation experiments

Serial number	Identification accuracy (%)
		1	99.70
2	99.77
		3	99.75
4	99.82
		5	99.85

As can be seen from the simulation result table 2, after a plurality of groups of experiments are carried out by using the same data set, the recognition accuracy is 99.7-99.9%, and the fluctuation range is only 0.2%, which shows that the target recognition method established by the invention has higher stability on the basis of keeping higher accuracy, and can meet the handwritten number recognition in most scenes. The method adopted by the invention is accurate and reliable, and a reliable method is provided for establishing an accurate image target identification model.

Claims

1. A target identification method based on a convolutional neural network and naive Bayes comprises the following specific identification steps:

(1) determining coordinates (x ', y ') of translated and normalized picture pixel points in a training set sample A to obtain a new training set A ';

according to the original coordinates (x) of the picture pixel points in the sample set₀，y₀) Maximum x of horizontal coordinate of pixel point after translation operation_maxMinimum value x_minMaximum value y of ordinate_maxMinimum value y_minThe dx and dy respectively represent the moving size of the pixel points on the x axis and the y axis, and the coordinates of the pixel points of the training sample set are determined;

(2) training the convolutional neural network and updating the connection weight omega_i；

Constructing a convolutional neural network, setting the number m of convolutional layers, the number n of pooling layers and the size(s) of convolutional layer convolutional core of the convolutional neural network₁，s₂，...，s_m) Size of the pooling layer (t)₁，t₂，...，t_n) Convolution layer step length r₁Step length r of the pooling layer₂The number N of neurons in the full connection layer, the activation function f and the iteration number l, and a connection weight omega in the convolutional neural network by adopting a back propagation algorithm and a BP algorithm_iUpdating is carried out;

wherein the content of the first and second substances,

is the actual output value, y_iFor ideal output value, η is the step length of each update, ω_i' is the updated connection weight, J is the cost function of the neural network;

(3) extracting a feature vector X;

reserving the alternately connected parts of the convolutional layer and the pooling layer in front of the fully-connected layer of the convolutional neural network, removing a Softmax layer behind the fully-connected layer, recording the fully-connected layer of the convolutional neural network as C, and outputting the fully-connected layer C as the extracted feature vector X after the training of the convolutional neural network model is finished;

(4) taking the feature vector X obtained in the last step as the input of a naive Bayes model, establishing the naive Bayes model for training, and realizing the classification of different training samples, wherein the naive Bayes rule is as follows;

wherein C is_iIs class, P (C)_i|x_i) Is shown as having x_iIs characterized by being divided into C_iThe probability of (d);