CN104102919B - Image classification method capable of effectively preventing convolutional neural network from being overfit - Google Patents

Abstract

The invention relates to an image classification method for effectively preventing over-fitting of a convolutional neural network, comprising: obtaining an image training set and an image test set; training a convolutional neural network model; using the trained convolutional neural network model to test images set for image classification. Among them, the steps of convolutional neural network model training are: preprocessing and sample amplification of the image data in the image training set to form training samples; performing forward propagation on the training samples to extract image features; calculating each sample in the Softmax classifier The classification probability of ; the training error is calculated according to the probability y _i ; the training error is used to backpropagate from the last layer of the convolutional neural network in turn, and the network weight matrix W is modified by the stochastic gradient descent method SGD. Compared with the prior art, the present invention has the advantages of high classification accuracy, fast convergence speed, high calculation efficiency and the like.

Description

An Image Classification Method Effectively Preventing Overfitting of Convolutional Neural Networks

technical field

The invention relates to the field of image processing, in particular to an image classification method for effectively preventing over-fitting of a convolutional neural network.

Background technique

With the wide application of multimedia technology and computer network, a large amount of image data appears on the network. In order to effectively manage these image files and provide users with better experience services, it is becoming more and more important to automatically identify the contents of these images.

With the continuous improvement and development of random machine learning methods, more and more attention has been paid to deep learning algorithms. Among them, convolutional neural network is an important algorithm in deep learning, and it has become a research hotspot in the fields of speech analysis and image recognition. The convolutional neural network breaks the way of full connection of neurons between layers in the traditional neural network. Its weight sharing network structure makes it more similar to the biological neural network, which reduces the complexity of the network model and reduces the weight value. quantity. This advantage is more obvious when the input of the network is the image, so that the image can be directly used as the input of the network, avoiding the complicated feature extraction and data reconstruction process in the traditional recognition algorithm. The convolutional network is a multi-layer perceptron specially designed to recognize two-dimensional shapes. This network structure is highly invariant to translation, scaling, tilting, or other forms of deformation.

Image classification technology based on convolutional neural network can effectively and automatically extract feature information from images, and the extracted features have very good image expression ability, so this technology has achieved satisfactory experimental results in some image classification problems. However, the technology currently has the following drawbacks:

First, because the labeled data in the image database is limited, as the scale of the convolutional neural network increases, the weights that need to be trained will also increase, which will inevitably lead to over-fitting of the neural network, that is, training The classification accuracy at test time is much better than that at test time.

Second, in order to obtain better feature expression ability and better classification accuracy, some researchers adopt the method of increasing network depth and expanding network size. However, this method will greatly increase the computational complexity, and the traditional CPU operation speed can no longer meet such computational complexity.

Contents of the invention

The object of the present invention is to provide an image classification method that can effectively prevent over-fitting of convolutional neural networks with high classification accuracy, fast convergence speed and high calculation efficiency in order to overcome the above-mentioned defects in the prior art.

The purpose of the present invention can be achieved through the following technical solutions:

An image classification method that effectively prevents overfitting of convolutional neural networks, the method runs on the GPU, including:

Step 1, obtain image training set and image test set;

Step 2, the training of the convolutional neural network model, specifically includes the following steps:

a) set the structure of the convolutional neural network and the upper limit of the number of training times N, initialize the neural network weight matrix W, the structure includes the number of layers of the convolutional neural network and the number of feature maps in each layer;

b) obtaining image data from the image training set for preprocessing, and performing sample amplification to form training samples;

c) performing forward propagation on the training sample to extract image features, the forward propagation including calculation of convolutional layer, nonlinear normalization layer and mixed pooling layer;

d) Calculate the classification probability of each sample in the Softmax classifier:

In the formula, s _i represents the output value of the i-th neuron of the Softmax classifier, s _i =F η, F is the image feature vector of a certain training sample, η is the corresponding weight, n is the number of categories to be classified ;

e) Calculate the training error according to the probability y _i

When i=k, θ _ik =1, i represents the i-th category, when the original input belongs to category i,

f) Utilizing the training error from the last layer of the convolutional neural network to backpropagate forward sequentially, and simultaneously utilizing the stochastic gradient descent method SGD to modify the network weight matrix W;

g) judge whether the model training is completed, if so, execute step 3 after saving the convolutional neural network model and the Softmax classifier, if not, then return to step b);

Step 3, using the trained convolutional neural network model to perform image classification on the image test set.

In a) of the step 2, the value range of the elements of the initial weight matrix W is [-0.01, 0.01].

b) of said step 2 is specifically:

b1) For images with equal length and width, use the cvResize function in OPENCV to scale, and the scaled picture size is N×N;

b2) For images with unequal length and width, the short side S is fixed, and S continuous pixels in the middle of the long side are intercepted to form an image of S×S size, and then step b1) is repeated to finally form an image of N×N size;

b3) Calculate the sum of the pixel values of all images, and divide by the number of images to obtain an average image, and subtract the average image from each image to obtain an input sample;

b4) performing data amplification on the input samples to form final training samples.

In c) of step 2, the calculation of the convolutional layer is specifically:

y _k = max{w _k *x,0}

Among them, x represents the output of the previous layer, that is, the input of the current layer, y _k represents the output of the kth feature map, w _k represents the kth weight matrix connected to the output of the previous layer, and "*" represents two Dimension inner product operation;

The calculation of the nonlinear normalization layer is specifically:

Among them, x _kij is the output of the k-th feature map of the previous layer during the calculation of the nonlinear normalization layer, and the accumulation operation is completed on the same position (i, j) of the N feature maps adjacent to the k-th feature map , α and β are the preset normalization parameters, and y _kij is the newly generated feature map;

The calculation of the mixed pooling layer is specifically:

Among them, λ is a random parameter with a value of 0 or 1, x _kpq is the output of the kth feature map of the previous layer when the hybrid pooling layer is calculated, and R _ij is the area to be down-sampled.

In the step g), the criterion for judging whether the model training is completed is: reaching the upper limit of training times.

Compared with the prior art, the present invention has the following advantages:

First, the present invention proposes for the first time the method of using mixed down-sampling Mixed Pooling when down-sampling the convolutional neural network, which can effectively prevent the over-fitting phenomenon of the neural network, and finally achieve the effect of improving the classification accuracy, and has good robustness specialty.

Second, the present invention uses an image processing unit GPU (Graphic Processing Units) to accelerate processing and calculation, so that the convolutional neural network can quickly calculate and converge when processing a large amount of data.

Third, the recognition accuracy of the present invention is superior to mainstream algorithms on CIFAR10, CIFAR100, and SVHN data sets, and has higher computational efficiency.

Description of drawings

Fig. 1 is the schematic diagram of the model training process of the present invention;

Fig. 2 is a schematic diagram of the image classification process of the present invention.

detailed description

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

As shown in Figures 1-2, an image classification method that effectively prevents convolutional neural networks from overfitting, the method first obtains an image set M from an image database and divides it into a training set M _t and a test set M _y , Then a convolutional neural network model is established according to the training set M _t , and finally the image classification is performed on the test set M _y with the trained convolutional neural network model.

As shown in Figure 1, the training of the convolutional neural network model specifically includes the following steps:

In step S101: set the structure of the convolutional neural network and the upper limit of training times N, and initialize the weight matrix W of the neural network, specifically:

1a) According to the scale of the problem, the level of the convolutional neural network and the number of feature maps of each layer are preset. The number n indicates the number of feature maps in this layer, Conv indicates the convolutional layer, LRN indicates the nonlinear normalization layer, Pooling indicates the downsampling layer, and Softmax indicates the structure of the classifier layer of the last layer);

1b) Connect the layers and initialize the weight matrix W of the neural network. The initialization method is to randomly generate a floating-point number of [-0.01, 0.01] for each element in W and assign it.

In step S102: obtain image data from training set M _t , carry out preprocessing and data amplification to the obtained image number:

2a) For images equal in length and width, directly utilize the cvResize function in OPENCV to zoom, so that the zoomed picture is N×N pixel size, in the present embodiment, N=32;

2b) For images with unequal length and width, fix the short side S unchanged, and intercept S pixels in the middle of the long side to form an image of S×S size, and then repeat the steps of 2a) to finally form a 32×32 size image image;

2c) Use all images to calculate the sum of the pixel values at each position, and divide by the number of images to obtain an average image, and finally subtract the input sample obtained by the average image from each image;

2d) Perform data amplification on each 32×32 input sample, regard the sample as a 32×32 matrix, and intercept 24×24 elements in the upper left, 24×24 elements in the lower left, 24×24 elements in the upper right, and 24×24 elements in the right The lower 24×24 elements and the central 24×24 are used as new input samples, and these 5 new input samples are continuously flipped horizontally to form 5 new input samples, so that the original 32×32 samples are processed by data expansion. After increasing the technology, 10 new samples of 24×24 are obtained.

In step S103: use the new sample obtained in step S102 as the input layer Input1 of the convolutional neural network to perform forward propagation to extract image features, and the forward propagation needs to go through the process of Conv64-LRN64-Pooling64-Conv64-LRN64-Pooling64, That is, feature extraction is performed after two stages of convolution. The specific steps are as follows:

3a) Convolution layer (Conv) calculation:

y _k = max{w _k *x,0}

Among them, x represents the output of the previous layer (that is, the input of this layer), the output value of the Input1 layer in the first Conv64 layer, and the output of the first Pooling64 in the second Conv64 layer, and y _k represents The k-th feature map output of the Conv64 layer (that is, the k-th output component of the previous layer), w _k represents the k-th weight matrix connected to the output of the previous layer, and "*" represents a two-dimensional inner product operation .

3b) Nonlinear Normalization (LRN) calculation:

Among them, x _kij is the output of the kth feature map of the Conv64 layer (that is, the kth output component of the previous layer), and the accumulation operation is the same position of the 5 feature maps adjacent to the kth feature map (i ,j) completed above, α=0.001 and β=0.75 are the preset normalization parameters, so far the newly generated feature map of the LRN layer can be expressed as y _kij .

3c) Hybrid pooling layer calculation:

Among them, λ is a random parameter with a value of 0 or 1, x _kpq is the output of the kth feature map of the previous LRN64 layer (that is, the kth output component of the previous layer), and R _ij is the area to be downsampled , the selected downsampling area is 3×3 in size.

The above three calculations are performed sequentially until all convolution stages are completed.

In step S104: use the image feature vector F obtained in step S103 to calculate the probability _yi of the sample being classified into each class in the Softmax classifier:

Among them, s _i represents the output value of the i-th neuron of the Softmax classifier, which is calculated from the corresponding weight of the image feature vector F dot product, that is, s _i =F·η, η is the corresponding weight, and n is the required The number of categories to classify. Assuming that there are n categories of pictures to be classified, an actual output vector of Y=y ₁ ,y ₁ ,...,y _n } is finally formed. Assuming that a certain picture belongs to the i-th category, its expected output vector is That is, the i-th element of the vector is 1, and the rest are 0, according to The error vector δ of the sample can be calculated, and the calculation formula of the i-th training error δ _i is:

When i=k, θ _ik =1, i represents the i-th category, when the original input belongs to category i,

In step S105: use the training error to propagate forward and backward sequentially from the last layer of the convolutional neural network, and propagate the error forward to the pooling layer, the nonlinear normalization layer and the convolutional layer sequentially from the Softmax layer, and at the same time Use the stochastic gradient descent method SGD to modify the network weight matrix W.

In step S106: After the backpropagation is completed, it is judged whether the number of training times has reached the upper limit N of training times set in step 1, and if it is reached, the training and storage model is stopped; if it is not reached, then continue to return to step S102 to continue training.

In step S107: save the trained model and Softmax classifier.

As shown in Figure 2, the specific steps for image classification of the test set M _y with the trained convolutional neural network model are as follows:

In step S201: extract test samples from the test set M _y for preprocessing and amplification, and obtain the data as network input;

In step S202: according to step S103, the calculation of the convolutional layer, the nonlinear normalization layer and the mixed pooling layer is performed sequentially until all convolutional stages are completed;

In step S203: obtain the feature vector of the test sample;

In step S204: use the feature vector and use the Softmax classifier to calculate the probability y _i of the sample being classified into each category, and calculate the largest element of {y ₁ , y ₁ ,...,y _n }, assuming that the element is y _j , Then the final judged category of the sample is the jth category.

In order to verify the performance of the present invention, this embodiment carried out experiments on three public data sets (CIFAR-10, CIFAR-100, SVHN), and used the convolutional neural network with hybrid downsampling and only ordinary downsampling The convolutional neural network methods are analyzed and compared. The experiments are carried out in accordance with the experimental regulations of the corresponding data sets for training and testing. From the comparison in Table 1, Table 2, and Table 3, it can be seen that the hybrid downsampling method can still obtain a better test error than the traditional downsampling method when the training error is relatively large, which fully proves that the hybrid downsampling method is beneficial to It plays an important role in preventing overfitting of convolutional neural networks. On the above three data sets, the test errors of the present invention are 10.80%, 38.07%, and 3.01%, respectively. The experimental results have a high recognition rate due to the currently published mainstream algorithms.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. The present invention also includes technical solutions composed of any combination of the above technical features.

Table 1 Experimental data of CIFAR-10 dataset

downsampling method training error test error Traditional Max Downsampling 3.01% 11.36% Traditional Mean Downsampling 4.52% 13.75% Mixed Downsampling Mix Pooling 6.25% 10.80%

Table 2 Experimental data of CIFAR-100 dataset

downsampling method training error test error Traditional Max Downsampling 5.42% 40.09% Traditional Mean Downsampling 14.61% 44.01% Mixed Downsampling Mix Pooling 25.71% 38.07%

Table 3 Experimental data of SVHN dataset

Claims (5)

1. it is a kind of effectively to prevent the image classification method of convolutional neural networks over-fitting, it is characterised in that the method operates in GPU In, including：

Step one, obtains training set of images and image measurement collection；

Step 2, the training of convolutional neural networks model, specifically includes following steps：

A) structure and frequency of training upper limit N of setting convolutional neural networks, initialize neural network weight matrix W, the structure Quantity including characteristic pattern in the number of plies of convolutional neural networks and every layer；

B) view data is obtained from described image training set to be pre-processed, and carry out sample amplification, form training sample；

C) propagated forward is carried out to the training sample and extracts characteristics of image, the propagated forward includes convolutional layer, non-linear returns One calculating for changing pooling layers of layer and mixing；

D) class probability of various kinds sheet is calculated in Softmax graders：

$y_i=\frac{\exp \left(s_i\right)}{{\mathrm{\Σ}}_{k=1}^n\exp \left(s_k\right)}$

In formula, s_iRepresent the output valve of Softmax i-th neuron of grader, s_i=F η, F are the image of certain training sample Characteristic vector, η is corresponding weights, and n is the categorical measure for needing classification；

E) according to probability y_iIt is calculated training error

${\mathrm{\δ}}_i=\underset{k=1}{\overset{n}{\Σ}}\frac{y_k^{*}}{y_k}{y}_k({\mathrm{\θ}}_{ik}-y_i)$

As i=k, θ_ik=1, i represent i-th classification, when be originally inputted belong to classification i when,

F) last layer successively forward backpropagation of the training error from convolutional neural networks is utilized, while using boarding steps Degree descent method SGD modification network weight matrix Ws；

Whether g) judgment models training completes, if so, performing step after then preserving convolutional neural networks model and Softmax graders Rapid three, if it is not, then return to step b)；

Step 3, image classification is carried out using the convolutional neural networks model after training to image measurement collection.

2. it is according to claim 1 a kind of effectively to prevent the image classification method of convolutional neural networks over-fitting, its feature Be, the step 2 a) in, the span of the element of initial weight matrix W is [- 0.01,0.01].

3. it is according to claim 1 a kind of effectively to prevent the image classification method of convolutional neural networks over-fitting, its feature It is, the b of the step 2) it is specially：

B1) for as broad as long image, zoomed in and out using the cvResize functions in OPENCV, the picture size after scaling It is N × N；

B2) to the unequal image of length and width, fixed short side S is constant, intercepts the continuous S pixel in the middle of side long, forms S × S big Small image, repeats step b1) ultimately form the image of N × N sizes；

B3 the pixel value sum of all images) is calculated, and quantity divided by image obtains an average image, in every piece image In subtract the average image and obtain input sample；

B4 data amplification) is carried out to the input sample, final training sample is formed.

4. it is according to claim 1 a kind of effectively to prevent the image classification method of convolutional neural networks over-fitting, its feature It is, the c of step 2) in, the calculating of the convolutional layer is specially：

y_k=max { w_k*x,0}

Wherein, x represents the input of the output of preceding layer, i.e. current layer, y_kRepresent k-th output of characteristic pattern, w_kRepresent with it is previous K-th weight matrix that the output of layer is connected, " * " represents the inner product operation of two dimension；

The calculating of the non-linear normalizing layer is specially：

$y_{kij}=x_{kij}/{(1+\frac{\mathrm{\α}}{N}\·\underset{l=k-\frac{N}{2}}{\overset{k+\frac{N}{2}}{\Σ}}{\left(x_{lij}\right)}^2)}^{\mathrm{\β}}$

Wherein, x_kijThe output of k-th characteristic pattern of preceding layer when being calculated for non-linear normalizing layer, accumulating operation is special at k-th Levy what is completed in the same position (i, j) of the adjacent N number of characteristic pattern of figure, α and β is default normalized parameter, y_kijFor newly-generated Characteristic pattern；

The calculating of pooling layers of the mixing is specially：

$y_{kij}=\mathrm{\λ}\·{\max }_{(p,q)\∈R_{ij}}\left(x_{kpq}\right)+(1-\mathrm{\λ})\·\frac{1}{\left|R_{ij}\right|}\·\underset{(p,q)\∈R_{ij}}{\Σ}{x}_{kpq}$

Wherein, λ is the random parameter that value is 0 or 1, x_kpqK-th characteristic pattern of preceding layer during for pooling layers of calculating of mixing Output, R_ijTo treat down-sampled region.

5. it is according to claim 1 a kind of effectively to prevent the image classification method of convolutional neural networks over-fitting, its feature It is that in the step g), the criterion whether judgment models training completes is：Reach the frequency of training upper limit.