CN106991440A - A kind of image classification algorithms of the convolutional neural networks based on spatial pyramid - Google Patents

Abstract

The invention discloses a kind of image classification algorithms of the convolutional neural networks based on spatial pyramid, spatial pyramid is used for reference and has first extracted global characteristics, then each pyramid levels are drawn and the global feature that local feature constitutes spatial pyramid is obtained in the way of grid.A new convolutional neural networks model is constructed, this model first half is traditional convolutional network, there are 3 convolutional layers, 2 pond layers；This 3 convolutional layers are subjected to uniform pondization in the way of gridding afterwards and obtain respective characteristic pattern.Every layer of characteristic pattern is linked to be a characteristic vector by row, and then this 3 characteristic vectors are linked in sequence as a total characteristic vector.This total characteristic vector is the feature for covering classical convolutional layer, while also with the addition of the feature of above convolutional layer, it is to avoid the loss of key character, while sizing grid have adjusted the weight of each convolutional layer characteristic pattern, is favorably improved the recognition efficiency of network.

Description

A kind of image classification algorithms of the convolutional neural networks based on spatial pyramid

Technical field

The invention belongs to image processing and pattern recognition field, and in particular to a kind of depth based on spatial pyramid The image recognition algorithm of convolutional neural networks

Background technology

Spatial pyramid is the global characteristics for extracting original image first, and then image is divided into each pyramid levels Refined net sequence, extracts feature from each grid of each pyramid level, and it is connected into a big characteristic vector.

The original achievement of what convolutional neural networks took in terms of image procossing in recent years, has obtained extensive utilization.With Afterwards, more researchers are modified to classic network.In order to obtain more preferably image recognition result, this patent is used for reference A kind of thinking of space pyramid, it is proposed that new depth convolutional neural networks, is obtained more preferable compared to conventional method Recognition effect.

The content of the invention

The purpose of the present invention is to propose to a kind of image classification of the depth convolutional neural networks based on spatial pyramid mode Mode, improves the ability of image steganalysis.

The technical solution adopted in the present invention is：A kind of image classification of convolutional neural networks based on spatial pyramid is calculated Method, it is characterised in that comprise the following steps：

Step 1：Propagated forward, is implemented including following sub-step：

Step 1.1：Set up the first half convolutional neural networks with M convolutional layer, M-1 pond layer；

Step 1.2：M convolutional layer is subjected to pond respectively, M category features are obtained, then connected into respectively one big Characteristic vector, finally reconnect into a total characteristic vector as the final feature of image；

Step 1.3：Full connection and softmax classification are carried out once to final characteristic vector, convolutional neural networks are obtained；

Step 1.4：All weights of whole convolutional neural networks are initialized by empirical equation, then will training Convolutional neural networks after picture x input initializations, are propagated according to propagated forward formula；

Step 2：Reversely regulation.

The beneficial effects of the invention are as follows：A kind of new convolutional neural networks algorithm structure is proposed, and improves identification effect Rate.

Brief description of the drawings

Fig. 1：The method schematic of the embodiment of the present invention.

Embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with the accompanying drawings and embodiment is to this hair It is bright to be described in further detail, it will be appreciated that implementation example described herein is merely to illustrate and explain the present invention, not For limiting the present invention.

See the image classification algorithms of Fig. 1, the present invention a kind of convolutional neural networks based on spatial pyramid provided, bag Include following steps：

Step 1：Propagated forward, is implemented including following sub-step：

Step 1.1：Set up the first half convolutional neural networks of 3 convolutional layers of band, 2 pond layers；

Step 1.2：3 convolutional layers are subjected to pond respectively, 3 category features are obtained, then connected into respectively one big Characteristic vector, finally reconnect into a total characteristic vector as the final feature of image；

Input after picture and the characteristic pattern of first convolutional layer is obtained by convolution kernel and hidden layer biasing, the convolution of first layer is special Levy figure x¹Obtained by equation below；

Wherein：The jth characteristic pattern of the 1st convolutional layer is represented,Represent input picture x after pretreatment⁰I-th Pictures, n₀Represent x⁰The number of picture；The 1st layer of j-th of two-dimensional convolution core is represented,Represent j-th of spy of the 1st hidden layer Levy the biasing of figure；δ represents sigmiod functions, and mp represents that what is obtained is characterized figure；

J=1.2......n1；N1 is the convolution kernel number of first layer, is also of the 1st convolution characteristic pattern Number；

The convolutional layer characteristic pattern of acquisition is subjected to down-sampling by 2*2 uniform pondization, it is original half to obtain ranks Characteristic pattern v¹；

v¹=mean-pooling { x¹}；

Wherein, mean-pooling represents uniform pond

Then the characteristic pattern of each convolutional layer can be transferred through below equation and obtain；

The characteristic pattern of each pond layer can be transferred through below equation and obtain；

v^l=mean-pooling { x^l}；

The characteristic pattern of 3 convolutional layers, i.e. x are obtained altogether¹,x²,x³, the mode of space pyramid is then used for reference, by 3 volumes The mode that the characteristic pattern of lamination carries out gridding carries out feature extraction, and what the present embodiment was taken is to draw the 1st convolutional layer to be divided into 4*4 grid, then each grid is by uniform pondization one feature of extraction, and final 1st convolutional layer is after feature extraction Characteristic pattern p as a 4*4¹；

1st convolutional layer is drawn to the grid for being divided into 4*4, then each grid extracts a feature by uniform pondization, most The 1st convolutional layer turns into 4*4 characteristic pattern p after feature extraction eventually¹；

p¹=mean-pooling (v¹)；

3 category feature figure p are similarly obtained according to below equation¹,p²,p³；

p^l=mean-pooling (v^l)；

Wherein, pond window size and step-length change with the change of input picture size；p¹,p²,p³Size difference For default 4*4,2*2,1*1；Then willThe column vector that size is 16, p are aggregated into by row¹Aggregate into 16*6=96 row to Amount, similarly can be by p²The column vector that size is 2*2*16=64 is polymerized to, by p³Be polymerized to size be 1*1*120=120 row to Amount, finally aggregate into a total size be 280 column vector p as input picture feature.

Step 1.3：Full connection and softmax classification are carried out once to final characteristic vector, convolutional neural networks are obtained；

Step 1.4：All weights of whole convolutional neural networks are initialized by empirical equation, then will training Convolutional neural networks after picture x input initializations, are propagated according to propagated forward formula；

All weights of whole convolutional neural networks are initialized by empirical equation, are that rule of thumb formula is initial Change the weight w between random generation input block and Hidden unit_kjWith the biasing b of Hidden unit_j,

It is 0 to set b initial values；

Wherein, w represents weights, and l represents l layers of convolutional network, and j represents j-th of nerve of convolutional neural networks l convolutional layers Member, k represents full articulamentum kth layer, and layerinput represents this layer of input neuron number, and layeroutput represents this The output neuron number of layer；k^lIt is the size of l-th of convolutional layer convolution kernel, this formula can make the weights of initialization be arrived -1 Between 1.

The picture of each input is expressed as x, and the graphical representation of input convolutional neural networks is x⁰；When the picture of input is During gray scale picture, x⁰=x；When it is colour picture to input picture, pass through below equation gray processing, x⁰=rgb2gray (x).

Training image x and its label are inputted, every layer of output valve is calculated using following forward conduction formula；h_w,b(x)=f (w^Tx+b)

Wherein, h_(w,b)(x) neuron output value, w are represented^TThe transposition of weights is represented, b represents biasing, and f represents to activate letter Number.

Step 2：Reversely regulation；Implement including following sub-step：

Step 2.1：By equation below according to label value and using forward conduction formula calculate obtain last layer it is defeated Go out value and calculate last layer of deviation；

Wherein, J_lFor l layers of loss function,For output layer neuron output value, h_w,b(x⁽ⁱ⁾) it is the i-th pictures Output valve, y⁽ⁱ⁾Represent the label of i-th input picture.

Step 2.2：Each layer deviation is calculated according to last layer of deviation, so as to try to achieve gradient direction, according to equation below more New weights：

The present embodiment regulation convolutional layer x during reverse regulation¹,x²When, the gradient for thering is both direction to transmit, this algorithm By the way that the gradient phase Calais of both direction is adjusted.

The present embodiment inputs a number of picture into the convolutional neural networks trained, is divided by propagated forward Class result is identical then correct with being compared from tape label.Thus the accuracy of network algorithm is obtained.

It should be appreciated that the part that this specification is not elaborated belongs to prior art.

It should be appreciated that the above-mentioned description for preferred embodiment is more detailed, therefore it can not be considered to this The limitation of invention patent protection scope, one of ordinary skill in the art is not departing from power of the present invention under the enlightenment of the present invention Profit is required under protected ambit, can also be made replacement or be deformed, each fall within protection scope of the present invention, this hair It is bright scope is claimed to be determined by the appended claims.

Claims (7)

1. a kind of image classification algorithms of the convolutional neural networks based on spatial pyramid, it is characterised in that comprise the following steps：

Step 1：Propagated forward, is implemented including following sub-step：

Step 1.1：Set up the first half convolutional neural networks with M convolutional layer, M-1 pond layer；

Step 1.2：M convolutional layer is subjected to pond respectively, M category features are obtained, a big spy is then connected into respectively Vector is levied, a total characteristic vector is finally reconnected into as the final feature of image；

Step 1.3：Full connection and softmax classification are carried out once to final characteristic vector, convolutional neural networks are obtained；

Step 1.4：All weights of whole convolutional neural networks are initialized by empirical equation, picture then will be trained Convolutional neural networks after x input initializations, are propagated according to propagated forward formula；

Step 2：Reversely regulation.

2. the image classification algorithms of the convolutional neural networks according to claim 1 based on spatial pyramid, its feature exists In：In step 1.2, if what is set up in step 1.1 is the first half convolutional Neural net with 3 convolutional layers, 2 pond layers Network, then carry out pond by 3 convolutional layers, obtain 3 category features respectively；Step 1.2 to implement process as follows：

Input after picture and the characteristic pattern of first convolutional layer, the convolution characteristic pattern of first layer are obtained by convolution kernel and hidden layer biasing x¹Obtained by equation below；

$x_j^1=mp\left(\mathrm{\δ}\left(\underset{i=1}{\overset{n_0}{\mathrm{\Σ}}}{x}_i^0*k_j^1+b_j^1\right)\right);$

Wherein：The jth characteristic pattern of the 1st convolutional layer is represented,Represent input picture x after pretreatment⁰I-th figure Piece, n₀Represent x⁰The number of picture；The 1st layer of j-th of two-dimensional convolution core is represented,Represent j-th of characteristic pattern of the 1st hidden layer Biasing；δ represents sigmiod functions, and mp represents that what is obtained is characterized figure；

N1 is the convolution kernel number of first layer, is also the number of the 1st convolution characteristic pattern；

The convolutional layer characteristic pattern of acquisition is subjected to down-sampling by 2*2 uniform pondization, the spy that ranks are original half is obtained Levy figure v¹；

v¹=mean-pooling { x¹}；

Wherein, mean-pooling represents uniform pond

Then the characteristic pattern of each convolutional layer can be transferred through below equation and obtain；

$x_j^l=mp\left(\mathrm{\δ}\left(\underset{i=1}{\overset{n_l}{\mathrm{\Σ}}}{v}_i^{l-1}*k_j^l+b_j^l\right)\right);$

The characteristic pattern of each pond layer can be transferred through below equation and obtain；

v^l=mean-pooling { x^l}；

The characteristic pattern of 3 convolutional layers, i.e. x are obtained altogether¹,x²,x³, then the characteristic pattern of 3 convolutional layers is carried out to the side of gridding Formula carries out feature extraction；

1st convolutional layer is drawn to the grid for being divided into 4*4, then each grid extracts a feature, the final 1st by uniform pondization Individual convolutional layer turns into 4*4 characteristic pattern p after feature extraction¹；

p¹=mean-pooling (v¹)；

3 category feature figure p are similarly obtained according to below equation¹,p²,p³；

p^l=mean-pooling (v^l)；

Wherein, pond window size and step-length change with the change of input picture size；p¹,p²,p³Size be respectively pre- If 4*4,2*2,1*1；Then willThe column vector that size is 16, p are aggregated into by row¹16*6=96 column vector is aggregated into, Similarly can be by p²The column vector that size is 2*2*16=64 is polymerized to, by p³The column vector that size is 1*1*120=120 is polymerized to, It is last sequentially to aggregate into column vector p that a total size is 280 as the feature of input picture.

3. the image classification algorithms of the convolutional neural networks according to claim 1 based on spatial pyramid, its feature exists In all weights of whole convolutional neural networks are initialized by empirical equation described in step 1.4, are rule of thumb Weight w between the random generation input block of formula initialization and Hidden unit_kjWith the biasing b of Hidden unit_j,

It is 0 to set b initial values；

$w_{lj}=\left(rand\right(k^l)-0.5)*2*\frac{\sqrt{6}}{\sqrt{layerinput+layeroutput}};$

$w_{kj}=\left(rand\left(layerinput,layeroutput\right)-0.5\right)*2*\frac{\sqrt{6}}{\sqrt{layerinput+layeroutput}}$

Wherein, w represents weights, and l represents l layers of convolutional network, and j represents convolutional neural networks l j-th of neuron of convolutional layer, k Full articulamentum kth layer is represented, layerinput represents this layer of input neuron number, and layeroutput represents this layer Output neuron number；k^lThe size of l-th of convolutional layer convolution kernel, this formula can make the weights of initialization -1 to 1 it Between.

4. the image classification algorithms of the convolutional neural networks according to claim 2 based on spatial pyramid, its feature exists In：In step 1.4, input training image x and its label calculate every layer of output valve using following forward conduction formula；

h_w,b(x)=f (w^Tx+b)

Wherein, h_(w,b)(x) neuron output value, w are represented^TThe transposition of weights is represented, b represents biasing, and f represents activation primitive.

5. the image classification of the convolutional neural networks based on spatial pyramid according to claim 1-4 any one is calculated Method, it is characterised in that：In step 1.4, the picture of each input is expressed as x, and the graphical representation of input convolutional neural networks is x⁰；When the picture of input is gray scale picture, x⁰=x；When it is colour picture to input picture, pass through below equation gray processing, x⁰ =rgb2gray (x).

6. the image classification algorithms of the convolutional neural networks according to claim 4 based on spatial pyramid, its feature exists In implementing including following sub-step for, step 2：

Step 2.1：Last layer of output valve obtained by equation below according to label value and using the calculating of forward conduction formula Calculate last layer of deviation；

${{\mathrm{\δ}}_l}^{\left(nl\right)}=\frac{\∂J_l}{\∂{Z_l}^{\left(nl\right)}}=\frac{\∂}{\∂{Z_l}^{\left(nl\right)}}\frac{1}{2}\left|\right|h_{w,b}\left(x^{\left(i\right)}\right)-y^{\left(i\right)}|{|}^2$

Wherein, J_lFor l layers of loss function,For output layer neuron output value, h_w,b(x⁽ⁱ⁾) be the i-th pictures output Value, y⁽ⁱ⁾Represent the label of i-th input picture.

Step 2.2：Each layer deviation is calculated according to last layer of deviation, so as to try to achieve gradient direction, is updated and weighed according to equation below Value：

$w_{ij}^{\left(l\right)}=w_{ij}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂w_{ij}^l}J(w,b)$

${b_i}^{\left(j\right)}={b_i}^{\left(j\right)}-\mathrm{\α}\frac{\∂}{\∂{b_i}^{\left(j\right)}}J(W,b).$

7. the image classification algorithms of the convolutional neural networks according to claim 6 based on spatial pyramid, its feature exists In：The regulation convolutional layer x during reverse regulation¹,x²When, the gradient for thering is both direction to transmit, this algorithm is by by both direction Gradient phase Calais be adjusted.