CN111368900A - Image target object identification method - Google Patents

Abstract

The invention relates to the technical field of image recognition, and specifically discloses an image target recognition method. ; Establish a support vector machine, and use the bat harmony algorithm to optimize the network parameters of the support vector machine; divide the marked image samples to be identified into a training set and a test set, and train the support vector machine; A convolutional neural network is built, and the convolutional neural network is trained; the real-time recognition image is collected by the camera in real time, and the real-time recognition image is input to the trained support vector machine for classification, and then the classified real-time recognition image is input to the corresponding volume The product neural network recognizes the target. This method takes advantage of the small-sample learning strength of support vector machines and the advantages of convolutional neural networks that are good at predicting pictures, and improves the accuracy and efficiency of prediction.

Description

An image target recognition method

technical field

The invention belongs to the technical field of image recognition, and particularly relates to an image target recognition method.

Background technique

Robot learning algorithms include support vector machine SVM, convolutional neural network, etc. SVM is a novel small-sample learning method with a solid theoretical foundation. It basically does not involve probability measures and the law of large numbers, etc., so it is different from the existing statistical methods. Essentially, it avoids the traditional process from induction to deduction, realizes efficient "transduction reasoning" from training samples to forecast samples, and greatly simplifies the usual problems of classification and regression. However, for the prediction of picture sets with large data, obviously, SVM is not enough. Therefore, because multi-layer networks such as convolution pooling of convolutional neural networks automatically extract image features and classify them, convolutional neural networks are generally used for Image recognition.

However, for the recognition of image targets, if the recognition of multiple targets is used, due to the huge amount of image data for each target, if only one convolutional neural network is used for recognition, the readiness rate and efficiency of image recognition will be reduced.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an image target recognition method, which can overcome the recognition of multiple targets. Because the image data of each target is huge, if only one convolutional neural network is used for recognition, it will reduce the image recognition time. Defects in readiness and efficiency.

In order to achieve the above purpose, the present invention provides an image target recognition method, including:

S1, the camera collects images of different types of objects to be recognized, and marks the characteristics of the objects of different types of objects to be recognized to form a set of images to be recognized;

S2, establish a support vector machine, and optimize the network parameters of the support vector machine by using the bat harmony hybrid algorithm to form a support vector machine with optimal network parameters;

S3: Divide the marked image samples to be identified into a training set and a test set, input the training set into the support vector machine training target feature data with optimal network parameters, and use the test set to test the trained support vector machine to obtain the A support vector machine for classifying object features;

S4 , establishing a plurality of convolutional neural networks corresponding to different targets respectively, and using the images to be identified of different target types to input the corresponding convolutional neural network training respectively, so as to obtain a plurality of convolutional neural networks capable of predicting the same type respectively. ;

S5, the real-time recognition image is collected by the camera in real time, the real-time recognition image is input into the trained support vector machine for classification, and then the classified real-time recognition image is input into the corresponding convolutional neural network to recognize the target object.

Preferably, in the above technical solution, step S2 specifically includes:

S201, setting the parameters of the support vector machine to be optimized: penalty parameter C, RBF kernel parameter δ, loss function ε parameter;

S202, initialize the harmony population size HMS, learn the probability HMCR of the harmony library, the pitch adjustment rate PAR, the distance bandwidth bw, the maximum number of iterations J and the harmony memory library HM;

S203, bat population initialization: initialize the population number N, the maximum pulse volume A ₀ , the maximum pulse rate R ₀ , the search lower limit x _min , the search upper limit x _max , the attenuation coefficient α of the volume, the enhancement coefficient γ of the search frequency, and the maximum number of iterations _Imax ;

S204, update the bat population, and update each bat in the bat population according to the following steps:

Generate new bats according to formula (1):

BatX(i)=gbest*(N(0,1)+1),as r(i)<rand(1)

where N(0,1) is a Gaussian distribution function with mean 0 and variance 1; r(i) is the pulse sent by the i-th individual bat; rand is a uniformly distributed random number in the range of [0,1]; gbest is the optimal value;

Out-of-bounds processing for newly spawned bats;

Calculate the fitness value f(BatX(i)) of the bat;

If the fitness of the newly generated bat is less than gbest, update gbest with the current bat;

Update the impulse loudness A and impulse rate r according to formula (2) (3)

r _i ^t+1 =R ₀ [1-exp(-γt)] (3)

S205, update the harmony population and gbest, randomly select a harmony in the harmony library to mutate according to formula (4) and calculate its fitness

If the newly generated harmony fitness is less than gbest, update gbest with the current bat;

S206, repeating S204 to S205 until the preset search precision is met or the maximum number of searches is reached, then go to step S207, otherwise go to S204 to continue the calculation;

S207, output gbest, and obtain the parameters of the optimized support vector machine to establish the optimal support vector machine.

Preferably, in the above technical solution, the convolutional neural network is a multi-layer perceptual convolutional neural network model based on the NIN network structure.

Preferably, in the above-mentioned technical solution, it also includes a normalizing step, which specifically includes:

Establish a coordinate system, input the acquired image to be recognized into the coordinate system, and obtain whether the edge of the image to be recognized is parallel to the coordinate axis of the coordinate system. If not, obtain the angle between the edge of the recognized image and the coordinate axis. The angle at which the to-be-recognized image is rotated is obtained by the included angle, and the to-be-recognized image is rotated according to the angle, so that the to-be-recognized image is turned upright.

Compared with the prior art, the present invention has the following beneficial effects:

1. The image target recognition method in the present invention uses a support vector machine to classify the target in the picture, and then input it into the corresponding target recognition network model for recognition, classify the picture and then recognize it, making full use of the The advantages of the support vector machine's small sample learning and the convolutional neural's good at predicting pictures are improved, and the prediction accuracy and efficiency are improved.

2. The present invention utilizes the support vector machine optimized by the bat harmony hybrid algorithm, and utilizes the local optimization capability of the bat algorithm and the global optimization capability of the harmony search algorithm for collaborative optimization, thereby improving the performance of the algorithm.

Description of drawings

FIG. 1 is a flow chart of a method for recognizing an image object according to the present invention.

Fig. 2 is the flow chart of the optimization of the bat harmony mixing algorithm of the present invention.

FIG. 3 is a schematic diagram of the NIN network structure of the present invention.

FIG. 4 is a recognition flow chart of the image target recognition method of the present invention.

Detailed ways

The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

As shown in Figure 1, an image target recognition method in this embodiment includes:

In step S1, the camera collects images to be recognized of different target types, and marks the target features of the different images to be recognized, that is, feature extraction, to form a set of images to be recognized. For example, select pictures of different items: images of trees, dogs, and cars, and mark the features of the items in the images.

In step S2, a support vector machine is established, and the network parameters of the support vector machine are optimized by using the bat harmony mixing algorithm, so as to form a support vector machine with optimal network parameters. Wherein, suppose the input of the image to be recognized is x, the output is y, y=f(x) is a nonlinear relationship, and the inverse function of y=f(x) can be obtained by using the support vector machine. In this embodiment, in order to simplify the operation, the SVM support vector machine is used to take the image to be recognized as the input, and the marked target feature as the output.

Step S3: Divide the marked image samples to be identified into a training set and a test set, input the training set into the support vector machine training target feature data with optimal network parameters, and use the test set to test the trained support vector machine, and obtain A support vector machine capable of classifying object features.

In step S4, multiple convolutional neural networks corresponding to different targets are established, and images to be identified of different target types are used to input the corresponding convolutional neural network training respectively, so as to obtain multiple convolutional neural networks capable of predicting the same type respectively. network.

In step S5, the real-time recognition image is collected by the camera in real time, and the real-time recognition image is input into the trained support vector machine for classification, and then the classified real-time recognition image is input into the corresponding convolutional neural network to recognize the target object, and the recognition effect is shown in the figure. 4 shown.

Step S2 adopts the bat harmony mixing algorithm. In this embodiment, the harmony and bat algorithms are introduced respectively:

The harmony search algorithm simulates the process of music creation, adjusting its pitch within the range of values each time to achieve the optimum. Its main parameters are harmony population size (HMS), probability of learning harmony library (HMCR), pitch adjustment ratio (PAR), distance bandwidth (bw), and maximum number of iterations (J). The steps of the HS algorithm are as follows:

Step1: Initialize the above parameters;

Step2: Initialize the harmony memory bank;

Step3: Improvise a new harmony. The update process of each dimension of the harmony is as follows:

Step3.1: Generate a random number rand1 between [0,1], if rand1 is less than HMCR, select a random tone in the harmony library, otherwise, generate a random number between [x _min , x _max ] as a new The pitch of , x _max and x _min are the upper and lower limits of the harmony, respectively.

是新产生的和声的第i个维度的值，微调公式如下：Step3.2: If it comes from the harmony library, generate a random number rand2 between [0,1], if rand2 is less than PAR, fine-tune it,

is the value of the i-th dimension of the newly generated harmony, and the fine-tuning formula is as follows:

Step4: Update the harmony memory bank. If the fitness of the new harmony is better than the fitness of the worst harmony in the harmony library, then replace the worst harmony in the harmony library with the searched harmony.

Step5: Repeat step 3 to step 4 until the maximum number of iterations is reached.

The bat algorithm is an intelligent optimization algorithm that simulates bats looking for prey. The bat algorithm judges the quality of the current position by the fitness of the current position. At each iteration, the bats are randomly shifted to the position of the best bat individual. The bat algorithm looks like this:

Step1: Population initialization: initialize the population number N, the maximum pulse volume A ₀ , the maximum pulse rate R ₀ , the search lower limit x _min , the search upper limit x _max , the attenuation coefficient α of the volume, the enhancement coefficient γ of the search frequency, and the maximum number of iterations I _max , the bat's position _xi .

Step2: Calculate the fitness f(x) of the bat, x=(x ₁ ,...x _d ) ^T , and evaluate the fitness value of each bat according to the fitness evaluation function to find the current optimal solution x ^* .

Step3: Update the search pulse frequency, speed and position of the bat.

p _i =p _min +(p _max -p _min )β (2)

where β is a random number and x ^* is the position of the optimal bat.

Step4: Generate a random number rand, if rand>r _i , perform random mutation on the optimal bat position in the current population to generate new bats.

Step5: Generate a random number rand, if rand>A _i and the fitness of the newly generated bat is better, update the position of the current bat, and update as follows

r _i ^t+1 =R ₀ [1-exp(-γt)] (6)

Step6: Find the fitness value of all bats and find the optimal solution.

Step7: Repeat Step2 to Step5 until the set optimal solution condition threshold is met or the maximum number of iterations is reached.

The invention introduces the bat algorithm into the harmony search algorithm for collaborative optimization. Its main operation is to initialize a bat population BatX to store the local solutions generated near the global optimal harmony gbest, and to generate a local solution BatX(i) after Gaussian perturbation of gbest in each optimization. Position information BatX(i) for evaluation: if the fitness value of the bat individual is better than gbest, and the loudness A(i) it generates is greater than the randomly generated loudness, then replace the worst harmony in the harmony library with bat and update gbest. The bat generation formula is shown in 7.

BatX(i)=gbest*(N(0,1)+1),as r(i)<rand(7)

where N(0,1) is a Gaussian distribution function with mean 0 and variance 1; r(i) is the pulse sent by the i-th individual bat; rand is a uniformly distributed random number in the range of [0,1]; gbest is the optimal value.

Therefore, as shown in Figure 2, step S2 specifically includes:

S201, setting the parameters of the support vector machine to be optimized: the parameter range of the penalty parameter C is [1, 100], the range of the RBF kernel parameter δ is [0.1, 100], the parameter range of the loss function ε is [0.001, 1].

S202 , initialize the harmony population size HMS, learn the probability HMCR of the harmony library, the pitch adjustment rate PAR, the distance bandwidth bw, and the maximum number of iterations J and the harmony memory library HM.

S203, bat population initialization: initialize the population number N, the maximum pulse volume A ₀ , the maximum pulse rate R ₀ , the search lower limit x _min , the search upper limit x _max , the attenuation coefficient α of the volume, the enhancement coefficient γ of the search frequency, and the maximum number of iterations _Imax ;

S204, update the bat population, and update each bat in the bat population according to the following steps:

Generate a new bat according to formula (7): BatX(i)=gbest*(N(0,1)+1), as r(i)<rand

Out-of-bounds processing for newly spawned bats;

Calculate the fitness value f(BatX(i)) of the bat;

If the fitness of the newly generated bat is less than gbest, update gbest with the current bat;

Update the impulse loudness A and impulse rate r according to formula (5) (6)

r _i ^t+1 =R ₀ [1-exp(-γt)]

S205, update the harmony population and gbest, randomly select a harmony in the harmony library to mutate according to formula (1) and calculate its fitness

If the newly generated harmony fitness is less than gbest, update gbest with the current bat;

S206, repeating S204 to S205 until the preset search precision is met or the maximum number of searches is reached, then go to step S207, otherwise go to S204 to continue the calculation;

S207, output gbest, select the optimal support vector machine model and its parameters, including: training parameters (including penalty factor C, radial basis kernel function parameters, etc.), model type, kernel function type, loss function and its parameters, to establish the optimal support vector machine.

In this way, the support vector machine optimized by the bat harmony hybrid algorithm utilizes the local optimization ability of the bat algorithm and the global optimization ability of the harmony search algorithm to perform collaborative optimization, which improves the performance of the algorithm.

Further, in this embodiment, the convolutional neural network uses a multi-layer perceptual convolutional neural network model based on the NIN network structure, and a deep learning model is constructed based on the NIN network structure. The overall structure of the NIN is composed of multiple multi-layer perceptual convolution layers. (Mlpconv) and a fully connected layer. Among them, each Mlpconv is the convolution operation of each local receptive field by the multi-layer perceptron micro-network structure, which consists of one convolution and two perceptual layers; a fully connected layer is added after the last layer of Mlpconv, and the The recognition output of the image target is realized by the method of linear regression. As shown in FIG. 3 , the network architecture of a deep learning model for high-precision detection of image target recognition provided by an embodiment of the present invention includes:

Input layer: The input object image is a three-channel true color image, and the image size is a dataset of 32x32x3.

The first Mlpconv layer: the convolution kernel is set to 16 in size of 3x3x3, the perceptron is implemented by 1x1 convolution, which are 1x1x16 and 1x1x32 respectively, and finally the dimensionality reduction is processed by the pooling layer;

The second Mlpconv layer: the convolution kernel is set to 32 in size of 3x3x16, and the perceptron is implemented by 1x1 convolution, which are 1x1x32 and 1x1x32 respectively, and finally through the pooling layer dimensionality reduction processing;

The third Mlpconv layer: the convolution kernel is set to 64 in size of 3x3x32, and the perceptron is implemented by 1x1 convolution, which are 1x1x64 and 1x1x64 respectively, and finally through the pooling layer dimensionality reduction processing;

The fourth Mlpconv layer: the convolution kernel is set to 128 in size of 3x3x64, and the perceptron is implemented by 1x1 convolution, which are 1x1x128 and 1x1x128 respectively, and finally through the pooling layer dimensionality reduction processing;

Fully connected layer: The images output through the convolutional layer and the pooling layer have relatively high-level features. The two-dimensional feature map is mapped to a one-dimensional space through the fully connected layer, and the 20x1 fully connected parameters are set. method, output it as the image target detection value, use the Leaky-Relu function for nonlinear activation, and use L1 regularization.

Further, when obtaining a picture and inputting it into a support vector machine or a convolutional neural network, it also includes a positive transformation step, which specifically includes:

Establish a coordinate system, input the acquired image to be recognized into the coordinate system, and obtain whether the edge of the image to be recognized is parallel to the coordinate axis of the coordinate system. If not, obtain the angle between the edge of the recognized image and the coordinate axis. The angle at which the to-be-recognized image is rotated is obtained by the included angle, and the to-be-recognized image is rotated according to the angle, so that the to-be-recognized image is turned upright.

To sum up, the image target recognition method in the present invention uses the support vector machine to classify the target in the picture, and then input it into the corresponding target recognition network model for recognition, and then classify the picture and then perform the recognition, which is sufficient. Taking advantage of the small-sample learning strength of the support vector machine and the advantages of the convolutional neural network being good at predicting pictures, the accuracy and efficiency of the prediction are improved.

The foregoing descriptions of specific exemplary embodiments of the present invention have been presented for purposes of illustration and description. These descriptions are not intended to limit the invention to the precise form disclosed, and obviously many changes and modifications are possible in light of the above teachings. The exemplary embodiments were chosen and described for the purpose of explaining certain principles of the invention and their practical applications, to thereby enable one skilled in the art to make and utilize various exemplary embodiments and various different aspects of the invention. Choose and change. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims (4)

1. an image target object recognition method, is characterized in that, comprises: S1, the camera collects images of different types of objects to be recognized, and marks the characteristics of the objects of different types of objects to be recognized to form a set of images to be recognized; S2, establish a support vector machine, and optimize the network parameters of the support vector machine by using the bat harmony hybrid algorithm to form a support vector machine with optimal network parameters; S3: Divide the marked image samples to be identified into a training set and a test set, input the training set into the support vector machine training target feature data with optimal network parameters, and use the test set to test the trained support vector machine to obtain the A support vector machine for classifying object features; S4 , establishing a plurality of convolutional neural networks corresponding to different targets respectively, and using the images to be identified of different target types to input the corresponding convolutional neural network training respectively, so as to obtain a plurality of convolutional neural networks capable of predicting the same type respectively. ; S5, the real-time recognition image is collected by the camera in real time, the real-time recognition image is input into the trained support vector machine for classification, and then the classified real-time recognition image is input into the corresponding convolutional neural network to recognize the target object. 2. The image target recognition method according to claim 1, wherein step S2 specifically comprises: S201, setting the parameters of the support vector machine to be optimized: penalty parameter C, RBF kernel parameter δ, loss function ε parameter; S202, initialize the harmony population size HMS, learn the probability HMCR of the harmony library, the pitch adjustment rate PAR, the distance bandwidth bw, the maximum number of iterations J and the harmony memory library HM; S203, bat population initialization: initialize the population number N, the maximum pulse volume A ₀ , the maximum pulse rate R ₀ , the search lower limit x _min , the search upper limit x _max , the attenuation coefficient α of the volume, the enhancement coefficient γ of the search frequency, and the maximum number of iterations _Imax ; S204, update the bat population, and update each bat in the bat population according to the following steps: Generate new bats according to formula (1): BatX(i)=gbest*(N(0,1)+1),as r(i)<rand(1) where N(0,1) is a Gaussian distribution function with mean 0 and variance 1; r(i) is the pulse sent by the i-th individual bat; rand is a uniformly distributed random number in the range of [0,1]; gbest is the optimal value; Out-of-bounds processing for newly spawned bats; Calculate the fitness value f(BatX(i)) of the bat; If the fitness of the newly generated bat is less than gbest, update gbest with the current bat; Update the impulse loudness A and impulse rate r according to formula (2) (3)

S205, update the harmony population and gbest, randomly select a harmony in the harmony library to mutate according to formula (4) and calculate its fitness

If the newly generated harmony fitness is less than gbest, update gbest with the current bat; S206, repeating S204 to S205 until the preset search precision is met or the maximum number of searches is reached, then go to step S207, otherwise go to S204 to continue the calculation; S207, output gbest, and obtain the parameters of the optimized support vector machine to establish the optimal support vector machine. 3 . The image target recognition method according to claim 1 , wherein the convolutional neural network is a multi-layer perceptual convolutional neural network model based on the NIN network structure. 4 . 4. The image target object recognition method according to claim 1, characterized in that, further comprising a step of correcting, specifically comprising: Establish a coordinate system, input the acquired image to be recognized into the coordinate system, and obtain whether the edge of the image to be recognized is parallel to the coordinate axis of the coordinate system. If not, obtain the angle between the edge of the recognized image and the coordinate axis. The angle at which the to-be-recognized image is rotated is obtained by the included angle, and the to-be-recognized image is rotated according to the angle, so that the to-be-recognized image is turned upright.

Images