CN111242063A - Construction method of small sample classification model based on transfer learning and application of iris classification - Google Patents

Abstract

The invention discloses a small sample classification model construction method based on migration learning and the application of the small sample classification model constructed by the method in iris image classification. Based on the migration learning of VGG16 model, an ICP-VGG model is constructed; The task is to configure the activation function of the fully connected layer in the custom network and the dropout ratio of the Dropout layer; fine-tune the network, set the relevant parameters of model training; obtain a small sample iris dataset, perform data preprocessing and data enhancement on the dataset; train and The model is verified, and the recognition result image is output; the method proposed by the present invention can better apply the deep learning model in the field of small sample iris, reduce overfitting and improve the recognition accuracy.

Description

Construction method of small sample classification model based on transfer learning and application of iris classification

technical field

The invention belongs to the technical field of computer images, in particular to a small sample classification model construction method based on migration learning and an iris image classification application.

Background technique

As human beings enter the era of big data and the rapid development of computing equipment, deep learning has entered a period of rapid development. In recent years, with the powerful expressive ability of deep learning models and large-scale training data sets, deep learning has achieved remarkable results in computer vision, speech recognition and other fields, especially in the field of image classification, showing an explosive growth , the accuracy of image classification on large-scale datasets continues to improve. Large-scale data sets are the cornerstone for deep learning to achieve remarkable results in various fields, but in practical applications, the acquisition of large-scale data sets requires expensive manpower and material resources or due to the limitations of certain fields. Therefore, compared with large-scale datasets, small-sample datasets are more common in practical applications. For example, in the field of iris, due to the particularity of its physiological structure, iris image acquisition is difficult, so the annotated iris image dataset is relatively small.

In the small-sample iris data set, due to the lack of training data, it is difficult to obtain ideal results by applying traditional deep learning models, and overfitting often occurs, that is, the model will perform very well on the training data set, and often errors will tend to zero, but the performance on the test data set is very poor and the accuracy rate is low. This is because when the deep model is very complex, it is easy to regard the noise of the training dataset of the small-sample iris dataset as a feature of the whole sample, and the learned model performs poorly on the test dataset. In order to better apply the deep learning model in the field of small sample iris, reduce overfitting, reduce the error of iris image classification detection results, and improve the recognition accuracy, the present invention proposes a transfer learning based on convolutional neural network. iris image recognition method.

SUMMARY OF THE INVENTION

In view of the deficiencies in the prior art, the present invention proposes a small sample classification model construction method and iris image classification application based on migration learning, which can better apply the deep learning model in the small sample iris field and reduce overfitting. , to improve the recognition accuracy.

The technical scheme adopted in the present invention is as follows:

A method for constructing a small sample classification model based on transfer learning, comprising the following steps:

Step 1, remove the three fully connected layers in the pre-trained P-VGG16 model, and retain 5 convolution blocks; and add a custom network after the convolution block to construct an ICP-VGG model; the custom networks are in turn: are the Flatten layer, the first fully connected layer, the Dropout layer and the second fully connected layer;

Step 2, configure the number of neurons in the fully connected layer in the custom network, the activation function of the fully connected layer and the dropout ratio of the Dropout layer; the activation functions in the first fully connected layer and the second fully connected layer are from Sigmoid, Choose from Tanh, ReLU, PReLu, ELU or SoftMax; the number of neurons in the first fully connected layer is 512 or 1024, and the number of neurons in the second fully connected layer is the same as the number of categories in the sample set to be tested; set Dropout the dropout ratio in the layer;

Step 3, train the ICP-VGG model network; determine the loss function in the IC-VGG model according to the type of the problem to be solved; set the relevant parameters for model training; determine the optimizer of the ICP-VGG model; Convolutional Neural Network Model.

其中，

为初始学习率，Epoch为模型在数据集上运行的次数，rate为衰减值；Further, the optimizer selects RMSProp, and sets the learning rate of RMSProp

in,

is the initial learning rate, Epoch is the number of times the model runs on the data set, and rate is the decay value;

Further, the method for training the ICP-VGG model is: keeping the weights of the network in the first four convolution blocks of the model unchanged, and only training the last convolution block and the custom network of the model;

Further, the loss function is selected as: for binary classification problems, use binary cross-entropy loss function; for multi-classification problems, use categorical cross-entropy loss function; for regression problems, use mean square error loss function; for sequence learning problems, use Use the connectionist temporal classification function;

Further, the model training-related parameters include the running times of the model and the batch size (batchSize) of the model;

Based on a convolutional neural network model suitable for image recognition constructed by the above method, the present invention also proposes an iris image recognition method, which obtains a small sample iris image data set, and divides the data set into a training set and a test set. The training set and the test set are respectively subjected to data preprocessing and data enhancement; the processed training set and test set are trained and tested on the convolutional neural network model suitable for image recognition constructed by the above method, and the iris image is obtained. classification results.

Further, the data preprocessing method is: modifying the size of the iris image in the data set to 224*224; converting the single-channel image into a three-channel image at the same time;

Further, the data enhancement method is: using the ImageDataGenerator method in Keras to perform data enhancement on the data in the training set and the test set; setting the parameter rescale=1./255; setting the parameter shear_range=0.2; setting the parameter zoom_range=0.2; setting Parameter horizontal-flip=True.

Beneficial effects of the present invention:

(1) The small sample classification model constructed based on the transfer learning of the present invention can reduce overfitting, improve the recognition accuracy, and obtain ideal results.

(2) The pre-trained VGG16 model is used for migration, and the ICP-VGG model is constructed and applied in the small-sample iris dataset, which can not only make full use of the network with a good classification foundation, but also learn the model through the sharing of the source domain and the target domain. parameters to transfer knowledge, is a very efficient way to apply deep learning to small image datasets.

(3) According to the iris image task, configure the activation function of the fully connected layer in the custom network and the dropout ratio of the dropout layer. The activation function of the first fully connected layer selects the ReLU activation function to enhance the fitting ability of the ICP-VGG model. The activation function of the second fully connected layer is classified with the SoftMax function. The dropout ratio in the dropout layer is set to 0.5, and half of the features of the first fully connected layer are randomly discarded during the training process to further reduce the overfitting of the ICP-VGG model.

(4) Fine-tune the network and set the relevant parameters for model training. Fine-tuning the ICP-VGG16 model and freezing some network weights can avoid the difficulty of training from scratch and effectively prevent network overfitting. Compared with training from scratch, fine-tuning will save a lot of computing resources and computing time, and improve computing efficiency and accuracy. . Choosing categorical cross-entropy as the loss function can better measure the probability distribution of network output and the true distribution of labels. Choosing RMSProp as the optimizer of the model and assigning a dynamic learning rate can better update the weights of the network based on the training data and loss function.

(5) Obtain a small sample iris dataset, and perform data preprocessing and data enhancement on the dataset. The data is preprocessed to make the original data more suitable for processing by the neural network and improve the algorithm effect. Use data augmentation techniques to generate more training data from existing training samples, allowing the model to observe more of the data and thus have better generalization capabilities.

Description of drawings

Fig. 1 is the method realization principle flow chart of the present invention;

Fig. 2 is the model structure diagram of the migration model VGG16 of the present invention;

Fig. 3 is the simplified network structure diagram of CP-VGG16 model of the present invention;

Fig. 4 is the simplified network structure diagram of ICP-VGG model of the present invention;

Fig. 5 is the fine-tuning simplified network structure diagram of the ICP-VGG model of the present invention;

FIG. 6 is an image of the recognition result output by the ICP-VGG model of the present invention on a small sample iris data set.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

In order to better apply the deep learning model in the field of small sample iris, reduce overfitting and improve the recognition accuracy, as shown in Figure 1, the present invention proposes a method for constructing a small sample classification model based on transfer learning, including: The following steps are implemented:

Step 1: Construct ICP-VGG model based on VGG16 model transfer learning

(1) Use the VGG16 model that has been pre-trained on the large dataset ImageNet (including 1.4 million images) and record it as P-VGG16. The initial input size of the model is 224*224, and the network structure is shown in Figure 2. The accumulation block is composed of three fully connected layers. The convolution block is composed of 13 layers of convolution layers (convolution) and 3 layers of pooling layers (maxpooling) and is divided into five groups. The first and second groups are composed of two layers. It consists of three convolutional layers and one pooling layer, and the third, fourth, and fifth groups are all constructed by three convolutional layers and pooling layers.

(2) Remove the last three fully connected layers in P-VGG16, and keep the convolutional basis in P-VGG16 as CP-VGG16, as shown in Figure 3.

(3) Migrate CP-VGG16 as a convolution base to the small-sample iris (Iris) dataset, and add four layers of custom network behind CP-VGG16 to construct an ICP-VGG model with a total of 22 layers, as shown in Figure 4 Show. The custom network is the Flatten layer, the first fully connected layer, the Dropout layer, and the second fully connected layer. Among them, the role of the Flatten layer: to "flatten" the input, that is, to make the multi-dimensional input one-dimensional. The role of the fully connected layer: map the learned features to sample labels for classification. The role of the Dropout layer: In order to reduce the overfitting phenomenon of the model in the small sample data set, some output features of this layer are randomly discarded during the training process.

Step 2: According to the iris image classification task, configure the number of neurons and activation function of the fully connected layer in the custom network, and the dropout ratio of the dropout layer.

2.1, Determine the activation function in the fully connected layer. In the model, the activation function defines the mapping relationship from the input neuron to the output neuron. Its function is mainly to add some nonlinear factors to the neural network, so that the neural network has a better fitting ability and can solve the problem of more complex problems. complicated question. Common activation functions are Sigmoid, Tanh, ReLU, PReLu, ELU, and SoftMax. Generally speaking, the ReLU activation function or the ELU activation function is used in the classification problem. The present invention studies the iris image classification problem. Therefore, the activation function in the first fully connected layer uses ReLU, and the activation function in the second fully connected layer uses ReLU. Softmax for classification.

2.2. The number of neurons in the first fully connected layer is set to 1024, and the number of neurons in the second fully connected layer is the same as the number of categories in the iris data set, which is 407 in this embodiment.

2.3, set the dropout ratio in the Dropout layer to 0.5, and randomly discard half of the features of the first fully connected layer during the training process to further reduce the overfitting of the ICP-VGG model.

Step 3: Fine-tune the network and set the relevant parameters for model training

3.1. Freeze the first 4 convolution blocks in ICP-VGG16, that is, keep the weights of these 4 convolution blocks unchanged during the model training process, and only train the 5th convolution block of the model and the custom network (three Convolutional layer, a pooling layer, a Flatten layer, a first fully connected layer, a dropout layer, a second fully connected layer), as shown in Figure 5.

3.2, Determine the loss function in the IC-VGG model according to the type of problem to be solved. For example, for binary classification problems, you can use the binary crossentropy loss function; for multiclass problems, you can use the categorical crossentropy loss function; for regression problems, you can use the mean-squared error -error) loss function; for sequence learning problems, the connection temporal classification (CTC, connection temporal classification) function can be used. The problem studied in the present invention is the iris image classification problem, so the selected loss function is categorical crossentropy.

3.3, set the number of runs (Epoch) of the model to 50, that is, the model runs 50 times on the entire dataset. Set the batch size (batchSize) of the model to 10, that is, the number of samples for the model to be trained at a time is 10.

3.4, Determine the optimizer of the ICP-VGG model. In the model, the optimizer can be used to update and influence the network parameters of the model training and model output to approximate or reach the optimal value, thereby minimizing the value of the loss function. The problem studied in the present invention is the iris image classification problem, and RMSProp is selected as the optimizer of the ICP-VGG model. The learning rate lr in the RMSProp optimizer is set to the dynamic learning rate according to equation (1).

为初始学习率，本发明设置为0.0001，rate为衰减值，本实用发明取1.8×10^-8。本发明选择RMSProp作为模型的优化器，并且赋值动态的学习率，可以更好地基于训练数据和损失函数来更新网络的权重。where Epoch is the number of times the model runs on the dataset,

is the initial learning rate, which is set to 0.0001 in the present invention, and the rate is the attenuation value, which is 1.8×10 ^-8 in the present invention. The present invention selects RMSProp as the optimizer of the model, and assigns a dynamic learning rate, which can better update the weight of the network based on the training data and the loss function.

Based on the small sample classification model based on migration learning constructed by the above method, the present invention also proposes an iris image recognition method,

1.1. Download the iris dataset CASIA-Iris-Lamp on the website of the Institute of Automation, Chinese Academy of Sciences. The dataset contains left and right eyes, and the right eye is taken as the research object. There are 407 categories and a total of 8050 pictures. Divide 3/5 of the pictures in each class into the training set and 2/5 of the pictures in each class into the test set.

2.2, preprocess the images in the training set and test set. The size of the original image in the dataset is 640*480 pixels, and it is a single-channel image. In order to facilitate network training, the size of the image in the dataset is modified to 224*224, and the single-channel image is converted into a three-channel image.

3.3. Since the training data in the iris sample set is relatively small, in order to reduce over-fitting, the ImageDataGenerator method in Keras is used to perform data augmentation on the data in the training set and test set. The specific parameter settings in ImageDataGenerator are as follows:

a, set the parameter rescale=1./255, that is, scale the pixels 0-255 in the picture to between 0-1;

b, set the parameter shear_range=0.2, that is, the angle of the random staggered transformation of the image is 0.2;

c, set the parameter zoom_range=0.2, that is, the range of random zooming of the image is 0.2;

d. Set the parameter horizontal-flip=True, that is, randomly flip half of the images horizontally.

In order to verify the effect of applying the convolutional neural network model for image recognition of the present invention on iris image recognition, the constructed ICP-VGG model is trained and tested on a small sample iris data set as shown in Figure 6, and the output is accurate according to the classification result. rate curve, output loss value graph. The accuracy rate on the test set is maintained at about 97.63%, and the loss function loss value is about 0.1898. The verification results prove that the model can effectively avoid overfitting on the small-sample iris data set and reduce the error of iris image classification and detection results. , which greatly improves the accuracy.

The above embodiments are only used to illustrate the design ideas and features of the present invention, and the purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly, and the protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made according to the principles and design ideas disclosed in the present invention fall within the protection scope of the present invention.

Claims (10)

1. a small sample classification model construction method based on migration learning, is characterized in that, comprises the following steps: Step 1, remove the three fully connected layers in the pre-trained P-VGG16 model and retain 5 convolution blocks; and add a custom network after the convolution blocks to construct the ICP-VGG model; Step 2, configure the number of neurons in the fully connected layer in the custom network, the activation function of the fully connected layer, and the dropout ratio of the Dropout layer; Step 3, train the ICP-VGG model network; determine the loss function in the IC-VGG model according to the type of the problem to be solved; set the relevant parameters for model training; determine the optimizer of the ICP-VGG model; Convolutional Neural Network Model. 2. The method for constructing a small sample classification model based on migration learning according to claim 1, wherein the self-defined network is followed by a Flatten layer, a first fully connected layer, a Dropout layer and a second fully connected layer. Floor. 3. The method for constructing a small sample classification model based on migration learning according to claim 2, wherein the activation functions in the first fully connected layer and the second fully connected layer are from Sigmoid, Tanh, ReLU, Choose from PReLu, ELU or SoftMax; the number of neurons in the first fully connected layer is 512 or 1024, and the number of neurons in the second fully connected layer is the same as the number of categories in the sample set to be tested. 4. a kind of small sample classification model construction method based on transfer learning according to claim 1, is characterized in that, the method for described training ICP-VGG model is: keep the weight of the network in the first 4 convolution blocks of the model different. Change, only train the last convolutional block and custom network of the model. 5. a kind of small sample classification model construction method based on transfer learning according to claim 1, is characterized in that, described optimizer selects RMSProp, and sets the learning rate of RMSProp

in,

is the initial learning rate, Epoch is the number of times the model runs on the dataset, and rate is the decay value. 6 . The method for constructing a small sample classification model based on transfer learning according to claim 1 , wherein the parameters related to the model training include the number of runs of the model and the batch size of the model. 7 . 7. The method for constructing a small sample classification model based on transfer learning according to any one of claims 1 to 6, wherein the loss function is selected as: for a binary classification problem, use binary crossover The entropy loss function; for multi-classification problems, use the categorical cross-entropy loss function; for regression problems, use the mean squared error loss function; for sequence learning problems, use the connectionist time series classification function. 8. a kind of iris image recognition method based on the small sample classification model constructed by the described method of claim 7, is characterized in that, obtains small sample iris image data set, and this data set is divided into training set and test set, respectively for The training set and the test set are subjected to data preprocessing and data enhancement; the processed training set and test set are trained and tested on the convolutional neural network model suitable for image recognition constructed by the above method, and the classification result of the iris image is obtained. 9 . The method for iris image recognition according to claim 8 , wherein the method for data preprocessing is: modifying the size of the iris image in the data set to 224*224; simultaneously converting the single-channel image into three channel image. 10. a kind of iris image recognition method according to claim 8, is characterized in that, the method for described data enhancement is: use ImageDataGenerator method in Keras to carry out data enhancement to the data in training set and test set; Setting parameter rescale= 1./255; set parameter shear_range=0.2; set parameter zoom_range=0.2; set parameter horizontal-flip=True.

Images