CN111242214A - Small animal identification method based on image - Google Patents

Abstract

An image-based small animal identification method, comprising the steps of: step 1: reducing the image to 32 × 32; step 2: inputting 32 × 32 images to an RBG three-color channel; and step 3: inputting the 32 x 3 images into a convolutional neural network model, and identifying and classifying; and 4, step 4: outputting the identified result, adding an L2 regularization term into a convolutional layer of the convolutional neural network, wherein the introduction of the regularization term balances deviation and variance, fitting capability and generalization capability, empirical risk and structural risk; the excitation function used by the convolutional layer is a Relu function, which makes the output of a part of neurons 0; a dropout method is added in the full link layer, and the dropout method can randomly delete the neurons and the affiliated relationship among the neurons in training. The method improves generalization ability and enhances recognition degree.

Description

An Image-Based Small Animal Recognition Method

technical field

The invention relates to the field of image recognition, in particular to an image-based small animal recognition method.

Background technique

With the acceleration of economic globalization, more and more people have started to keep pets, especially small animals, such as cats and dogs in China, and lizards and snakes in foreign countries. However, many people who do not keep small animals do not understand and are not familiar with the living habits of the corresponding small animals, so they are accidentally injured by them when they play and play, so it is very important to identify and understand the habits of the small animals that they get along with. Convolutional neural networks are commonly used in the recognition of small animal images, but training a general convolutional neural network model requires a large amount of sample data, which is time-consuming and labor-intensive. Therefore, a convolutional neural network model with few samples can bring great convenience and save time. Effortless.

SUMMARY OF THE INVENTION

The training of the common convolutional neural network model requires a large amount of sample data. If the number of sample data is insufficient, the problem of over-fitting will occur, which will lead to the reduction of the animal recognition rate. The invention provides an image-based small animal recognition method, which can optimize the small number of samples. The problem of reducing the accuracy of the model caused by over-fitting is generated, and the accuracy of recognition is improved.

In order to solve the above-mentioned technical problems, the present invention provides the following technical solutions:

An image-based small animal identification method, comprising the following steps:

Step 1: Reduce the image to a 32×32 image;

Step 2: The input layer inputs the length, width and depth of the image. The length and width represent the size of the image. The input of the present invention is the 32×32 image reduced in step 1, and its length is 32 and the width is 32; the depth represents the color channel of the image, and the input is the RGB three-color channel, so The color image depth is 3, that is, the input image is 32×32×3;

Step 3: Build a convolutional neural network structure, the process is as follows:

Step 3-1: The convolution layer C1 performs a convolution operation on the image matrix input by the input layer, extracts features, and obtains a feature map through the action of the excitation function. The formula of the feature map of the convolution layer is shown in formula (1).

代表第l层第j个神经元，f为激励函数，M_j代表前一层输出特征图集合，

代表第l-1层上第i个神经元，*代表卷积运算，

代表l-1层第i个神经元到l层第j个神经元的卷积核矩阵，

代表l层第j个神经元的偏置；In formula (1),

represents the _jth neuron of the lth layer, f is the excitation function, Mj represents the output feature map set of the previous layer,

represents the ith neuron on the l-1th layer, * represents the convolution operation,

Represents the convolution kernel matrix from the ith neuron in layer l-1 to the jth neuron in layer l,

represents the bias of the jth neuron in layer l;

The excitation function is the Relu function, and the Relu function is shown in formula (2).

Relu(x)=max(0,x) (2)

When the number of samples is less than the number of parameters, the sample matrix is likely to be irreversible, and the introduction of the regularization term will solve this problem; the introduction of the regularization term balances bias and variance, fitting ability and generalization ability, experience Risk and structural risk, using the L2 regularization method, is shown in formula (3).

C ₀ represents the original cost parameter, w is the shared weight, λ is the given regular term coefficient, n represents the number of neurons, then the coefficient of the regularization term is 1×10 ^-4 ;

The convolutional layer C1 performs L2 regularization on the weights at the same time through 32 5×5 convolution kernels, and obtains 32 28×28×3 feature maps;

Step 3-2: According to the local correlation principle of the image, the sub-sampling layer S1 aggregates and samples the feature map output by the convolution layer in a small adjacent area. While reducing the features and parameters, it retains the useful information of the image and uses Stochastic -pooling method, assign probability to pixels according to the size of the value, and then perform sub-sampling according to the probability;

The subsampling layer S1 subsamples the feature map generated by the convolutional layer C1 in step 3-1 in a 2×2 area, with a step size of 1, to obtain 32 feature maps of 14×14×3;

Step 3-3: The convolutional layer C2 extracts the feature map in the same way as the convolutional layer C1 in step 3-1. The only difference is that the convolution kernel of the convolutional layer C2 is 3×3, and the convolutional layer feature map formula As shown in formula (1), the excitation function used is shown in formula (2), and the L2 regularization method used is shown in formula (3);

The convolutional layer C2 performs L2 regularization on the weights through 32 3×3 convolution kernels at the same time, and obtains 1024 12×12×3 feature maps;

Step 3-4: The sub-sampling layer S2 has the same method as the sub-sampling layer S1 in step 3-2. The sub-sampling layer S2 sub-samples the feature map generated by the convolutional layer C2 in the 2×2 area, and the step size is 1. Obtain 1024 6×6×3 feature maps;

Step 3-5: Flatten the feature map in step 3-4 to generate a one-dimensional feature vector and input it to the full link layer. The full link layer has three layers:

The size of the full link layer F1 is 256;

The size of the full link layer F2 is 128, and then the dropout method is used to activate the neurons in a certain layer with a certain probability. The probability parameter of dropout is 0.5, and the activation function of the full link layer F2 uses the Relu function. ) shown;

The size of the full link layer F3 is 64, and then the dropout method is used to activate the neurons in a certain layer with a certain probability. The probability parameter of dropout is 0.5, and the activation function of the full link layer F3 uses the Relu function. ) shown;

Step 4: The output layer is a classifier, which is used to output the probability of each category, and the sum of the probability values of all categories is 1. The classifier function used is the soft-max function. The soft-max function formula is shown in formula (4) :

In formula (4), h(x ⁽ⁱ⁾ ) represents the probability that the sample i belongs to the kth class, and the total number of classes is K.

The technical idea of the present invention is as follows: considering the time-consuming and labor-intensive problems of training the model and the large number of samples required, introducing an L2 regularization term in the convolution layer, using the Relu function and adding the dropout method in the full link layer can effectively optimize the The problem of model accuracy reduction caused by fitting.

The beneficial effects of the present invention are as follows: in this method, when the number of samples is less than the number of parameters, the sample matrix is likely to be irreversible, and this problem can be solved by introducing a regularization term in the convolution layer. The introduction of regularization term balances bias and variance, fitting ability and generalization ability, empirical risk and structural risk, so it can improve the accuracy of identification. Using the Relu function as the excitation function will cause the output of some neurons to be 0, which will cause the sparsity of the network, reduce the interdependence of parameters, and alleviate the overfitting problem and improve the recognition accuracy.

Description of drawings

FIG. 1 is an overall flow chart of the method.

Figure 2 is a flow chart of the CNN network.

Figure 3 shows the structure of the CNN network.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

Referring to Figures 1 to 3, an image-based small animal recognition method, the whole process is shown in Figure 1, divided into four steps, the more critical step is the third step.

As shown in Figure 2, after inputting the image into the convolutional neural network, the input layer reduces the image to a size of 32 × 32, and then inputs the RGB three-color channels, that is, the input image is 32 × 32 × 3.

Then the 32×32×3 image is input to the convolutional layer C1 with L2 regularization. The convolution kernel of the convolutional layer C1 is 32 5×5 convolution kernels. The matrix is convolved to extract features, and after the action of the excitation function, 32 feature maps of 28×28×3 are obtained. The excitation function of the convolutional layer C1 is the Relu function, and the formula is as follows:

Relu(x)=max(0,x)

The formula of the feature map of the convolutional layer is as follows:

When the number of samples is less than the number of parameters, the sample matrix is likely to be irreversible, and the introduction of the regularization term will solve this problem. The introduction of the regularization term balances bias and variance, fitting ability and generalization ability, experience Risk and structural risk, the L2 regularization method formula used is as follows:

Then 32 28×28×3 feature maps are input into the sub-sampling layer S1, and the sub-sampling layer S1 uses the Stochastic-pooling method to sub-sample the feature maps generated by the convolutional layer C1 in the 2×2 area. is 1, and 32 feature maps of 14×14×3 are obtained;

Then 32 14×14×3 feature maps are input to the convolutional layer C2 added with L2 regularization. The convolution kernel of the convolutional layer C1 is 32 3×3 convolution kernels. The input image matrix is subjected to convolution operation to extract features, and through the action of the excitation function, 1024 feature maps of 12×12×3 are obtained. The excitation function of the convolutional layer C1 is the Relu function, and the formula is as follows:

Relu(x)=max(0,x)

The formula of the feature map of the convolutional layer is as follows:

The L2 regularization formula used is as follows:

Then 1024 12×12×3 feature maps are input into the sub-sampling layer S2, and the sub-sampling layer S2 uses the Stochastic-pooling method to sub-sample the feature maps generated by the convolutional layer C2 in the 2×2 area. is 1, and 1024 6×6×3 feature maps are obtained.

Then the 1024 6×6×3 feature maps are flattened to generate a one-dimensional feature vector.

Then the generated one-dimensional vector is input into the fully linked layer F1, and the size of the fully linked layer F1 is 256.

Then, the result generated by the full link layer F1 is input into the full link layer F2, the size of the full link layer F2 is 128, and the dropout method is used to activate the neurons of a certain layer according to a certain probability. The probability parameter of the dropout method is 0.5.

Then, the result generated by the full link layer F2 is input into the full link layer F3, and the size of the full link layer F3 is 64, and then the dropout method is used to activate the neurons of a certain layer according to a certain probability. The probability parameter of the dropout method is 0.5.

Finally, the results are classified and output.

The above is an introduction to an embodiment of an image-based small animal recognition method of the present invention. The convolutional neural network introduced by the present invention includes an input layer, two convolution layers with L2 regularization added, and two sub-sampling layers. Three full-link layers and one output layer are added to the dropout method. Through this convolutional neural network, the requirements of enhancing the recognition rate of small animals can be achieved.

It should be pointed out here that the specific implementation in the present invention merely enumerates individual examples of the present invention, and the solutions implemented using the design idea of the present invention and its equivalent changes shall all belong to the protection scope of the present invention.

Claims (1)

1. An image-based small animal identification method, characterized in that it comprises the following steps:

step 1: reducing the image to a 32 × 32 image;

step 2: inputting the length, width and depth of the picture input by the input layer, wherein the length and width represent the size of the image, and the input in the invention is the image with the size of 32 multiplied by 32 reduced in the step 1, the length is 32, and the width is 32; the depth represents the color channel of the image, and the input is RGB three-color channel, so the color image depth is 3, namely the input image is 32 × 32 × 3;

and step 3: constructing a convolutional neural network structure, wherein the process is as follows:

step 3-1: the convolutional layer C1 performs convolution operation on the image matrix input by the input layer, extracts features, and obtains a feature map through the action of an excitation function, wherein the formula of the convolutional layer feature map is shown as formula (1):

in the formula (1), the reaction mixture is,

represents the jth neuron of layer 1, f is the excitation function, M_jRepresents the set of output feature maps of the previous layer,

represents the ith neuron on the l-1 th layer, represents convolution operation,

a convolution kernel matrix representing the i-th neuron of layer l-1 to the j-th neuron of layer l,

represents the bias of the jth neuron at layer 1;

the excitation function is a Relu function, which is expressed by equation (2):

Relu(x)＝max(0，x) (2)

in the case that the number of samples is less than the number of parameters, the sample matrix is very likely to be irreversible, and the introduction of the regularization term solves the problem; the introduction of the regularization term balances deviation and variance, fitting ability and generalization ability, empirical risk and structural risk, and the regularization method using L2 is shown in formula (3):

C₀representing the original cost parameter, w being the shared weight, λ being a given regular top coefficient, n representing the number of neurons, the coefficient of the term being 1 × 10^-4；

The convolutional layer C1 performs L2 regularization on the weights through 32 5 × 5 convolution kernels to obtain 32 28 × 28 × 3 feature maps;

step 3-2: the sub-sampling layer S1 performs aggregate sampling on the feature map output by the convolutional layer in adjacent small areas according to the local correlation principle of the image, retains useful information of the image while reducing features and parameters, gives probability to pixel points according to the value by using a Stochastic-posing method, and performs sub-sampling according to the probability;

the sub-sampling layer S1 sub-samples the feature map generated by the convolutional layer C1 in step 3-1 in a 2 × 2 region with a step size of 1, to obtain 32 14 × 14 × 3 feature maps;

step 3-3: the convolutional layer C2 has the same manner as the convolutional layer C1 in step 3-1 for extracting the feature map, and the only difference is that the convolutional layer C2 has a convolution kernel of 3 × 3, the formula of the convolutional layer feature map is shown in formula (1), the excitation function used is shown in formula (2), and the L2 regularization method used is shown in formula (3);

the convolutional layer C2 performs L2 regularization on the weights through 32 convolution kernels of 3 × 3 at the same time to obtain 1024 feature maps of 12 × 12 × 3;

step 3-4: the sub-sampling layer S2 is similar to the sub-sampling layer S1 in step 3-2 in that the sub-sampling layer S2 sub-samples the feature map generated by the convolutional layer C2 in a 2 × 2 region with a step size of 1, and 1024 feature maps of 6 × 6 × 3 are obtained;

step 3-5: adopting a flat operation on the characteristic diagram in the step 3-4 to generate a one-dimensional characteristic vector to be input into a full link layer, wherein the full link layer comprises three layers:

the size of the full link layer F1 is 256;

the size of the full link layer F2 is 128, then a dropout method is adopted to activate neurons of a certain layer according to a certain probability, the probability parameter of the dropout is 0.5, the excitation function of the full link layer F2 uses a Relu function, and the specific expression is shown as the formula (2);

the size of the full link layer F3 is 64, then a dropout method is adopted to activate neurons of a certain layer according to a certain probability, the probability parameter of the dropout is 0.5, the excitation function of the full link layer F3 uses a Relu function, and the specific expression is shown as the formula (2);

and 4, step 4: the output layer is a classifier and is used for outputting the probability belonging to each class, the sum of the probability values of all the classes is 1, and the used classifier function is a soft-max function which is expressed by the formula (4):

in the formula (4), h (x)⁽ⁱ⁾) The probability that the sample i belongs to the kth class is shown, and the total class number is K.

Images