CN107578054A

CN107578054A - Image processing method and device

Info

Publication number: CN107578054A
Application number: CN201710888763.8A
Authority: CN
Inventors: 万韶华
Original assignee: Beijing Xiaomi Mobile Software Co Ltd
Current assignee: Beijing Xiaomi Mobile Software Co Ltd
Priority date: 2017-09-27
Filing date: 2017-09-27
Publication date: 2018-01-12

Abstract

The disclosure is directed to image processing method and device.This method includes：Processing image is treated using the first convolution module of default network model and carries out feature extraction, obtains the fisrt feature figure of pending image, wherein, the size of fisrt feature figure is less than pending image, and the first convolution module includes at least one first convolutional layer；And then fisrt feature figure is up-sampled using the second convolution module of default network model, obtain second feature figure, the size of second feature figure is more than fisrt feature figure, and the second convolution module includes the second convolutional layer, and the step-length of the convolution kernel movement of the second convolutional layer is proper fraction；According to second feature figure, characteristic pattern after the processing equal with pending picture size is obtained.On the premise of disclosed method can reach the size of amplification fisrt feature figure, image processing efficiency is effectively lifted.

Description

Image processing method and device

Technical field

This disclosure relates to technical field of image processing, more particularly to image processing method and device.

Background technology

Convolutional neural networks (Convolutional Neural Network, referred to as：CNN) since 2012, Image classification and image detection etc. achieve huge achievement and are widely applied.

CNN powerful part is its automatic learning characteristic of sandwich construction energy, and may learn many levels Feature：Shallower convolutional layer perception domain is smaller, the feature of study to some regional areas；Deeper convolutional layer has larger sense Know domain, can learn to the more abstract feature of some.These abstract characteristics are to sensitiveness such as the size of object, position and direction It is lower, so as to contribute to the raising of recognition performance.

These abstract features are helpful to classifying, and can judge what classification is included in piece image well Object, but because lost the details of some objects, it is impossible to the specific profile of object is provided well, points out each pixel tool Which object body belongs to, therefore accomplishes that accurate segmentation is just very difficult.

The way of traditional dividing method based on CNN is typically：In order to a pixel classifications, around the pixel An image block as CNN input be used for train and predict.This method has several drawbacks in that：First, storage overhead is very big. Such as the size of the image block used each pixel is 15 × 15, then required memory space is 225 times of original image.Two It is that computational efficiency is low.Adjacent block of pixels is substantially what is repeated, and convolution, this calculating are calculated one by one for each block of pixels Also there is repetition largely.Third, the size that size limit sensing region of block of pixels.The size ratio of usual block of pixels The size of entire image is much smaller, can only extract some local features, so as to cause the performance of classification to be restricted.

For this problem, UC Berkeley Jonathan Long et al. propose full convolutional neural networks (Fully Convolutional Networks, referred to as：FCN) it is used for the segmentation of image.The network attempts to recover from abstract feature Go out the classification belonging to each pixel.The classification of pixel scale is further extended into from the classification of image level.But pass through FCN In convolutional layer for several times and pond layer processing after, the size of image can constantly diminish, that is, the resolution ratio of image is increasingly It is low.Therefore need to recover the original size of image using measure.

The content of the invention

To overcome problem present in correlation technique, the embodiment of the present disclosure provides image processing method and device.The skill Art scheme is as follows：

According to the first aspect of the embodiment of the present disclosure, there is provided a kind of image processing method, including：

Processing image is treated using the first convolution module of default network model and carries out feature extraction, is obtained described pending The fisrt feature figure of image；The size of the fisrt feature figure is less than the pending image, and first convolution module includes At least one first convolutional layer；

The fisrt feature figure is up-sampled using the second convolution module of the default network model, obtains second Characteristic pattern, the size of the second feature figure are more than the fisrt feature figure；Second convolution module includes the second convolutional layer, The step-length of the convolution kernel movement of the second convolutional layer is proper fraction in second convolution module；

According to the second feature figure, characteristic pattern after the processing equal with the pending picture size is obtained.

The technical scheme provided by this disclosed embodiment can include the following benefits：Using the of default network model One convolution module treats processing image and carries out feature extraction, obtains the fisrt feature figure of pending image, wherein, fisrt feature figure Size be less than pending image, the first convolution module includes at least one first convolutional layer；And then using default network model The second convolution module fisrt feature figure is up-sampled, obtain second feature figure, the size of second feature figure is more than first Characteristic pattern, the second convolution module includes the second convolutional layer, and the step-length of the convolution kernel movement of the second convolutional layer is proper fraction；According to Second feature figure, obtain characteristic pattern after the processing equal with pending picture size.Processing is being treated by the first convolution module After image carries out feature extraction, the size of obtained fisrt feature figure can be less than pending image, in order that obtaining output end output Image recover the original size of image so that user can more be clearly seen that the image after processing, lead in the disclosure Cross and set the second convolution module for including the second convolutional layer to up-sample fisrt feature figure, now, due to volume Two product module The step-length of the convolution kernel movement of the second convolutional layer is proper fraction in block, so as to reach the mesh of the size of amplification fisrt feature figure , and due to being operated using deconvolution, required data are more during processing, treatment effeciency can be reduced, and in the disclosure It on the premise of being operated without using deconvolution, can equally reach the purpose of the size of amplification fisrt feature figure, and effectively improve The efficiency of image procossing.

According to the second aspect of the embodiment of the present disclosure, there is provided a kind of image processing apparatus, including：

First acquisition module, carried for treating processing image progress feature using the first convolution module of default network model Take, obtain the fisrt feature figure of the pending image；The size of the fisrt feature figure is less than the pending image, described First convolution module includes at least one first convolutional layer；

Second acquisition module, for the second convolution module using the default network model to first acquisition module The fisrt feature figure obtained is up-sampled, and obtains second feature figure, and the size of the second feature figure is more than described the One characteristic pattern；Second convolution module includes the second convolutional layer, the convolution kernel of the second convolutional layer in second convolution module Mobile step-length is proper fraction；

3rd acquisition module, for the second feature figure obtained according to second acquisition module, obtain with it is described Characteristic pattern after the equal processing of pending picture size.

According to the third aspect of the embodiment of the present disclosure, there is provided a kind of image processing apparatus, including：

Processor；

For storing the memory of processor-executable instruction；

Wherein, the processor is configured as：

According to the fourth aspect of the embodiment of the present disclosure, there is provided a kind of computer-readable recording medium, be stored thereon with calculating Machine instructs, and the instruction realizes following steps when being executed by processor：

It should be appreciated that the general description and following detailed description of the above are only exemplary and explanatory, not The disclosure can be limited.

Brief description of the drawings

Accompanying drawing herein is merged in specification and forms the part of this specification, shows the implementation for meeting the disclosure Example, and be used to together with specification to explain the principle of the disclosure.

Fig. 1 is the flow chart of the image processing method according to an exemplary embodiment.

Fig. 2 is the schematic diagram of the default network model according to an exemplary embodiment one.

Fig. 3 is the schematic diagram of the default network model according to an exemplary embodiment two.

Fig. 4 is the schematic diagram of the default network model according to an exemplary embodiment three.

Fig. 5 is the element group schematic diagram of the fisrt feature figure according to an exemplary embodiment.

Fig. 6 is the convolution kernel schematic diagram of the second convolutional layer according to an exemplary embodiment.

Fig. 7 is the convolution operation schematic diagram according to an exemplary embodiment one.

Fig. 8 is the convolution operation schematic diagram according to an exemplary embodiment two.

Fig. 9 is the convolution operation schematic diagram according to an exemplary embodiment three.

Figure 10 is the convolution operation schematic diagram according to an exemplary embodiment four.

Figure 11 is the application scenario diagram of the embodiment of the present disclosure method according to an exemplary embodiment one.

Figure 12 is the application scenario diagram of the embodiment of the present disclosure method according to an exemplary embodiment two.

Figure 13 is a kind of block diagram of image processing apparatus according to an exemplary embodiment one.

Figure 14 is a kind of block diagram of image processing apparatus according to an exemplary embodiment two.

Figure 15 is a kind of block diagram of image processing apparatus according to an exemplary embodiment three.

Figure 16 is a kind of block diagram of image processing apparatus according to an exemplary embodiment four.

Figure 17 is a kind of block diagram of image processing apparatus according to an exemplary embodiment five.

Figure 18 is a kind of block diagram of image processing apparatus according to an exemplary embodiment six.

Figure 19 is a kind of block diagram of image processing apparatus according to an exemplary embodiment seven.

Figure 20 is a kind of block diagram for image processing apparatus 80 according to an exemplary embodiment.

Figure 21 is a kind of block diagram of device 90 for image procossing according to an exemplary embodiment.

Embodiment

Here exemplary embodiment will be illustrated in detail, its example is illustrated in the accompanying drawings.Following description is related to During accompanying drawing, unless otherwise indicated, the same numbers in different accompanying drawings represent same or analogous key element.Following exemplary embodiment Described in embodiment do not represent all embodiments consistent with the disclosure.On the contrary, they be only with it is such as appended The example of the consistent apparatus and method of some aspects be described in detail in claims, the disclosure.

In the embodiment of the present disclosure, processing image progress feature is treated using the first convolution module of default network model and carried Take, obtain the fisrt feature figure of pending image, wherein, the size of fisrt feature figure is less than pending image, first volume product module Block includes at least one first convolutional layer；And then fisrt feature figure is carried out using the second convolution module of default network model Sampling, obtains second feature figure, the size of second feature figure is more than fisrt feature figure, and the second convolution module includes the second convolution Layer, and the step-length of the convolution kernel movement of the second convolutional layer is proper fraction；According to second feature figure, obtain and pending picture size Characteristic pattern after equal processing.After processing image progress feature extraction is treated by the first convolution module, first obtained is special The size of sign figure can be less than pending image, in order that the image for obtaining output end output recovers the original size of image, to cause User can more be clearly seen that the image after processing, include the volume Two product module of the second convolutional layer in the disclosure by setting Block up-samples to fisrt feature figure, now, due to the step-length of the convolution kernel movement of the second convolutional layer in the second convolution module For proper fraction, so as to reach the purpose of the size of amplification fisrt feature figure, such as：Second convolutional layer in second convolution module Convolution kernel movement step-length be 1/2, after being up-sampled using the second convolution module to fisrt feature figure, obtain second The size of characteristic pattern will be 2 times of the size of fisrt feature figure.And due to being operated using deconvolution, required data during processing It is more, treatment effeciency can be reduced, and in the disclosure on the premise of being operated without using deconvolution, it can equally reach amplification The purpose of the size of fisrt feature figure, and effectively improve the efficiency of image procossing.

Fig. 1 is the flow chart of the image processing method according to an exemplary embodiment one, as shown in figure 1, this method Comprise the following steps S101-S103：

In step S101, processing image progress feature is treated using the first convolution module in default network model and carried Take, obtain the fisrt feature figure of pending image；The size of fisrt feature figure is less than pending image, and the first convolution module includes At least one first convolutional layer.

Example, the size of fisrt feature figure can be represented with the lateral length of fisrt feature figure with longitudinally wide, and Lateral length with it is longitudinally wide can in units of pixel, can also by centimetre in units of.Such as：In units of pixel, the The size of one characteristic pattern is 480 × 800, wherein, the transverse direction of 480 expression fisrt feature figures has 480 pixels, and 800 represent first There are 800 pixels the longitudinal direction of characteristic pattern.

When treating processing image progress feature extraction using the first convolution module, pending image is inputted to the first volume Lamination 1, process of convolution is carried out to pending image in the first convolutional layer 1, and by the pending image after process of convolution Output is to pond layer 1, and to the image progress pond processing of input in pond layer 1, the image after pondization is handled is exported to the One convolutional layer 2, the like, last first convolutional layer N carries out the obtained as fisrt feature figure after process of convolution.

Because the size of fisrt feature figure is less than pending image, if directly exporting the fisrt feature figure to user, use Family clearly can not see desired processing result image from the fisrt feature figure, so as to make it that user satisfaction is relatively low.

In step s 102, fisrt feature figure is up-sampled using the second convolution module in default network model, Second feature figure is obtained, the size of second feature figure is more than fisrt feature figure；Second convolution module includes the second convolutional layer, and second The step-length of the convolution kernel movement of the second convolutional layer is proper fraction in convolution module.

In order to lift the satisfaction of user, that is, in order to amplify the size of fisrt feature figure, exported in the first convolution module After fisrt feature figure, up-sampling processing can be carried out to fisrt feature image using the second convolution module.As shown in Fig. 2 when second , will when being up-sampled using the second convolution module to fisrt feature figure when convolution module includes multiple second convolutional layers Fisrt feature figure is inputted to the second convolutional layer 1, and up-sampling processing is carried out in the second convolutional layer 1, and will be passed through up-sampling and be handled Fisrt feature figure afterwards is exported to the second convolutional layer 2, the like, after last second convolutional layer M carries out up-sampling processing What is obtained is second feature figure.

In a kind of achievable mode, the second convolution module that network model is preset in above-mentioned use is entered to fisrt feature figure Row up-sampling can be implemented in the following manner：The convolution kernel of second convolutional layer and fisrt feature figure are subjected to convolution.

It is proper fraction step-length Move Volumes during performing convolution operation in each second convolutional layer of the second convolution module Product core, such as：Convolution kernel can be moved with 1/2 step-length, after now carrying out convolution to fisrt feature figure, the second feature of output The size of figure would is that 2 times of fisrt feature figure size, thus reach the purpose of up-sampling, namely reach amplification first The purpose of the size of characteristic pattern.

Because amount of calculation is small when convolution is handled than deconvolution, therefore operate, but use without using deconvolution in the disclosure Convolution operation, so as to effectively lift treatment effeciency.

In step s 103, according to second feature figure, characteristic pattern after the processing equal with pending picture size is obtained.

It is worth noting that, above-mentioned default network model includes but is not limited to, full convolutional neural networks (Fully Convolutional Networks, referred to as：FCN) model.

FCN is the general depth convolutional network framework for carrying out image segmentation, and the network attempts to recover from abstract feature Go out the classification belonging to each pixel, i.e., further extend into the classification of pixel scale from the classification of image level.FCN is by tradition Full articulamentum in CNN changes into convolutional layer one by one.In traditional CNN structures, first 5 layers are convolutional layers, the 6th layer and 7 layers be respectively a length be 4096 one-dimensional vector, the 8th layer be length be 1000 one-dimensional vector, respectively correspond to 1000 The probability of classification.FCN is expressed as convolutional layer, the size (port number, wide, height) of corresponding convolution kernel by the 6th layer, the 7th layer and the 8th layer Respectively (4096,1,1), (4096,1,1) and (1000,1,1).Because all layers are all convolutional layers, therefore referred to as full convolution net Network.When carrying out image segmentation using FCN, after the convolutional layer for several times in FCN and the processing of pond layer, the size meeting of image Constantly diminish, in order that the image that must be exported recovers the original size of image, at present, can be realized and adopted by deconvolution operation Sample recovers the purpose of the original size of image to reach.Deconvolution is similar with convolution, is all the computing being added that is multiplied.It is anti-realizing When the forward and backward of convolution operation is propagated, the forward, backward of reverse convolution can be used to propagate.But using deconvolution During operation, treatment effeciency can be caused than relatively low.

In one embodiment, feature extraction can be realized by way of convolution operation, now, presets network model In the first convolution module convolution operation only can be carried out to pending image, as shown in figure 3, first in default network model Convolution module just only includes at least one first convolutional layer, and those first convolutional layers are sequentially connected.So, the first convolution is being used When module carries out feature extraction to pending image, pending image is inputted to the first convolutional layer 1, in the first convolutional layer 1 Process of convolution is carried out, and the pending image after process of convolution is exported to the first convolutional layer 2, the like, last Individual first convolutional layer N carries out the obtained as fisrt feature figure after process of convolution.

Because the picture quality of the fisrt feature figure only got by convolution operation is relatively low, therefore, in another implementation In example, feature extraction can also be realized by way of convolution and pondization operation, now, preset first in network model Convolution module can carry out convolution simultaneously to pending image and pondization operates, then, the first volume product module of default network model Not only include at least one first convolutional layer, in addition at least one pond layer in block.Wherein, the quantity of pond layer is no more than the The quantity of one convolutional layer, as shown in figure 4, each pond layer is arranged between two the first convolutional layers, and in the first convolution module The step-length of convolution kernel movement be integer more than or equal to 1.

Example, the step-length of the convolution kernel movement in the first convolution module can refer to the first convolution in the first convolution module The step-length of the convolution kernel movement of layer.

The technical scheme provided by this disclosed embodiment can include the following benefits：By being carried out to pending image Convolution and pondization processing so that the picture quality of obtained fisrt feature figure is higher, effectively improves the quality of image procossing.

Because each convolution kernel in the second convolutional layer is moved with fraction step-length, then, using the second convolutional layer Convolution kernel up-sampled during, it may appear that each element in the convolution kernel of the second convolutional layer and element to be sampled Group situation about misplacing, the further comprising the steps of A1 of the above method, now above-mentioned steps S102 may be embodied as step A2：

In step A1, computing is carried out to the element of fisrt feature figure using bilinear interpolation, after obtaining interpolation processing Fisrt feature figure.

In step A2, the fisrt feature figure after difference processing is up-sampled using the second convolution module.

Example, Fig. 5 is the element group schematic diagram of the fisrt feature figure according to an exemplary embodiment, first spy Pel element group is levied as 5 × 5 matrix, Fig. 6 is the convolution kernel schematic diagram of the second convolutional layer according to an exemplary embodiment, The convolution kernel of second convolutional layer is 3 × 3 matrix.

Illustrated so that the processing of upper and lower sample is process of convolution as an example, when the convolution kernel that the second convolutional layer is moved with step-length 1, As shown in fig. 7, when it carries out convolution operation with first group of fisrt feature pel element group, in fact, above-mentioned convolution operation is just It is a dot product operations, as shown in figure 8, the value after convolution is：1×1+0×1+1×1+0×0+1×1+0×1+1×0+0×0 + 1 × 1=4, the convolution kernel of the second convolutional layer is moved by above-mentioned step-length, namely the convolution kernel of the second convolutional layer moves every time One pixel (rectangular block in figure), often moves and once performs once dot product operations as shown in Figure 8.

When with step-length 1 to move the convolution kernel of the second convolutional layer, the convolution kernel of the second convolutional layer all can be with first every time The element group alignment of convolution is treated in the element group of characteristic pattern, so as to directly carry out very much convolution operation, and when with Fractional-step It is long to move the convolution kernel of the second convolutional layer when, the convolution kernel of the second convolutional layer be possible to can with the element group of fisrt feature figure Treat that the element group of convolution misplaces, so as to cannot directly carry out convolution operation.

Assuming that on the basis of Fig. 7, after the convolution kernel that the second convolutional layer is moved with step-length 1, what is obtained as shown in Figure 9 shows It is intended to, and after the convolution kernel of the second convolutional layer is moved with step-length 1/2, obtained schematic diagram as shown in Figure 10.

As shown in Figure 10, the convolution kernel of the second convolutional layer and treat convolution element group be dislocation, now, in order to perform volume Product operation, can enter row interpolation using bilinear interpolation in the element group for treating convolution, obtain the first spy after interpolation processing The element group of figure is levied, is then rolled up using the element group of the fisrt feature figure after the convolution kernel and interpolation processing of the second convolutional layer Product.

It is worth noting that, above-mentioned bilinear interpolation may be replaced by linear interpolation method etc., the disclosure is not to it It is any limitation as.

The technical scheme provided by this disclosed embodiment can include the following benefits：It is special to first using double interpolation methods The element for levying figure carries out computing, obtains the fisrt feature figure after interpolation processing, in the convolution kernel that can avoid the second convolutional layer The problem of element group in each element and fisrt feature figure during dislocation there occurs that can not up-sample, adopts so as to effectively improve The accuracy of the numerical value obtained after sample.

The step-length of convolution kernel movement in the quantity of above-mentioned second convolutional layer and each second convolutional layer can be according to reality Using adjustment, in one embodiment, in quantity and each second convolutional layer that the second convolutional layer can be determined according to following formula Convolution kernel movement step-length：

Wherein, the M be first convolutional layer quantity, the N be the second convolutional layer quantity, the F_jFor jth The step-length of convolution kernel movement in individual second convolutional layer, the E_iFor the step-length of the convolution kernel movement in i-th of first convolutional layers.

Due to the step-length E of the convolution kernel movement in each first convolutional layer_iIt is to determine with the quantity M of the first convolutional layer, There are two unknown numbers in above-mentioned formula, it is determined that the convolution kernel in the quantity and each second convolutional layer of the second convolutional layer moves Step-length when, be premised on meeting following constraints：

Example, it is assumed that M=4, then：

When the quantity of the second convolutional layer is N=2, then：

log E₁+log E₂+log E₃+log E₄+log F_i+log F₂=0

When the step-length of the convolution kernel of each first convolutional layer is 2, then：

log2+log2+log2+log2+log F₁+log F₂=0

That is, F₁And F₂Value to meet：2×2×2×2×F₁×F₂=1

Due to F₁And F₂It is proper fraction, then, F₁And F₂Value can beWithCan also beWithCan also beWith

As the step-length E of the convolution kernel movement in the 1st the first convolutional layer₁Convolution kernel in=2, the 2nd the first convolutional layer Mobile step-length E₂The step-length E of convolution kernel movement in=3, the 3rd the first convolutional layer₃Volume in=2, the 4th the first convolutional layer The step-length E of product core movement₄When=4, then：

log2+log3+log2+log4+logF₁+log F₂=0

That is, F₁And F₂Value to meet：2×3×2×4×F₁×F₂=1

It can release

Due to F₁And F₂It is proper fraction, then, F₁And F₂Value for example can beWithAlso may be used ThinkWithCan also beWithCan also be F₁=1 HeDeng.

The technical scheme provided by this disclosed embodiment can include the following benefits：Can by above-mentioned constraints In the hope of the step that when the quantity of the second convolutional layer is some numerical value, now the convolution kernel in each second convolutional layer moves It is long, so as to which corresponding perform of the step-length moved by the convolution kernel in obtained each second convolutional layer is accumulated with fraction step-length Move Volumes The convolution operation of core so that the size of the second feature figure finally given is with the premise of pending image identical, effectively being lifted The speed of image procossing.

After the step-length of the movement of the convolution kernel in each second convolutional layer has been obtained, during convolution is performed, also need Know the convolution kernel size in each second convolutional layer.

When the second convolution module includes second convolutional layer, in one embodiment, the convolution of the second convolutional layer Core can redefine, or the convolution kernel size in any one first convolutional layer.

Example, there are 3 the first convolutional layers in the first convolution module, be respectively：First convolutional layer 1, the and of the first convolutional layer 2 First convolutional layer 3；Wherein, the convolution kernel size of the first convolutional layer 1 is 3 × 3, and the convolution kernel size of the first convolutional layer 2 is 3 × 3, The convolution kernel size of first convolutional layer 3 is 5 × 5；There are 1 the second convolutional layer, now, the second convolutional layer in second convolution module Convolution kernel size can be：3 × 3 or 5 × 5.

Convolution kernel so is selected for the second convolutional layer at random, may be such that the second feature image obtained after final convolution Effect it is poor.

In another embodiment, the convolution kernel in the second convolutional layer is the convolution kernel in the first convolutional layer of predetermined number The average value of size.

Such as：There are 3 the first convolutional layers in first convolution module, be respectively：First convolutional layer 1, the first convolutional layer 2 and One convolutional layer 3；There is 1 the second convolutional layer in second convolution module, be：Second convolutional layer 1；So, the convolution of the second convolutional layer 1 Core size can be：The volume of the convolution kernel size of first convolutional layer 1, the convolution kernel size of the first convolutional layer 2 and the first convolutional layer 3 The average value of product core size；The convolution kernel size of second convolutional layer 1 can also be：The convolution kernel size of first convolutional layer 1 and The average value of the convolution kernel size of one convolutional layer 3.

In another embodiment, convolution kernel corresponding to each first convolutional layer in the first convolution module can first be obtained； Then the size for determining the convolution kernel that size is maximum in the first convolutional layer is the convolution kernel size of the second convolutional layer.

It is determined that the second convolutional layer in the second convolution module convolution kernel size when, can first obtain each first convolution The convolution kernel of layer, can be by those the first convolutional layers because the convolution kernel of each the first convolutional layer has the size of oneself The size of convolution kernel is contrasted respectively, selects the size of the maximum convolution kernel of size as the convolution kernel size of the second convolutional layer.

Example, can by the width of the convolution kernel of each first convolutional layer and it is high be compared respectively, selection it is wherein wide and Height is convolution kernel of the convolution kernel of maximum as the second convolutional layer.

Continue according to above-mentioned example, in 3 the first convolutional layers of the first convolution module, the convolution of the first convolutional layer 3 The size of core is maximum, then, the convolution kernel size of the second convolutional layer is：5×5.

The technical scheme provided by this disclosed embodiment can include the following benefits：By determining each first convolution In the convolution kernel of layer the size of the maximum convolution kernel of size for the convolution kernel of the second convolutional layer size, so as to pass through obtain the The convolution kernel of two convolutional layers comes after carrying out convolution operation to fisrt feature figure, can effectively lift the effect of second feature figure.

When the second convolution module includes at least two second convolutional layers, at this point it is possible to according in the first convolution module The convolution kernel and preset rules of each first convolutional layer determine the convolution kernel of each second convolutional layer in the second convolution module.

In a kind of achievable mode, preset rules can be：According to preset order in the first convolution module each Convolution kernel in one convolutional layer is ranked up, and the convolution kernel in the first convolutional layer after sequence is used as in the second convolution module successively The convolution kernel of each second convolutional layer.

Example, the mode arranged in sequence is carried out to the convolution kernel in each first convolutional layer in the first convolution module Sequence, the convolution kernel in the first convolutional layer after sorting in sequence are used as each second convolutional layer in the second convolution module successively Convolution kernel, that is, the convolution kernel of each second convolutional layer is each first volume in the first convolution module in the second convolution module The order arrangement of convolution kernel in lamination.

Assuming that have 3 the first convolutional layers in the first convolution module, respectively the first convolutional layer 1, the first convolutional layer 2 and first Convolutional layer 3；Also there are 3 the second convolutional layers, respectively the second convolutional layer 1, the second convolutional layer 2 and volume Two in second convolution module Lamination 3.Then the convolution kernel of the second convolutional layer 1 is the convolution kernel of the first convolutional layer 1, and the convolution kernel of the second convolutional layer 2 is the first volume The convolution kernel of lamination 2, the convolution kernel of the second convolutional layer 3 are the convolution kernel of the first convolutional layer 3.

Example, the convolution kernel in each first convolutional layer in the first convolution module is carried out in the way of reversing Sequence, the convolution kernel in the first convolutional layer after being sorted according to backward are used as each second convolutional layer in the second convolution module successively Convolution kernel, that is, the convolution kernel of each second convolutional layer is each first volume in the first convolution module in the second convolution module Convolution kernel in lamination reverses.

Assuming that have 3 the first convolutional layers in the first convolution module, respectively the first convolutional layer 1, the first convolutional layer 2 and first Convolutional layer 3；Also there are 3 the second convolutional layers, respectively the second convolutional layer 1, the second convolutional layer 2 and volume Two in second convolution module Lamination 3.Then the convolution kernel of the second convolutional layer 1 is the convolution kernel of the first convolutional layer 3, and the convolution kernel of the second convolutional layer 2 is the first volume The convolution kernel of lamination 2, the convolution kernel of the second convolutional layer 3 are the convolution kernel of the first convolutional layer 1.

In a kind of achievable mode, preset rules can be：The convolution of each second convolutional layer in second convolution module Core size is the average value of the convolution kernel size in the first convolutional layer of predetermined number in the first convolution module.

Such as：There are 4 the first convolutional layers in first convolution module, be respectively：First convolutional layer 1, the first convolutional layer 2, One convolutional layer 3 and the first convolutional layer 4；There are 2 the second convolutional layers in second convolution module, be respectively：Second convolutional layer 1 and second Convolutional layer 2；So, the convolution kernel size of the second convolutional layer 1 can be the convolution kernel size and the first convolutional layer of first volume lamination 1 The average value of 2 convolution kernel size, the convolution kernel size of the second convolutional layer 2 can be first volume lamination 3 convolution kernel size and The average value of the convolution kernel size of first convolutional layer 4；Or second the convolution kernel size of convolutional layer 1 can be first volume lamination 1 Convolution kernel size and the first convolutional layer 3 convolution kernel size average value, the convolution kernel size of the second convolutional layer 2 can be the The average value of the convolution kernel size of one convolutional layer 2 and the convolution kernel size of the first convolutional layer 4；Or second convolutional layer 1 convolution Core size can be the convolution kernel size of first volume lamination 1, and the convolution kernel size of the second convolutional layer 2 can be first volume lamination 2 Convolution kernel size, the average value of the convolution kernel size of the first convolutional layer 3 and the convolution kernel size of the first convolutional layer 4 be averaged Value；Or second convolutional layer 1 convolution kernel size can be first volume lamination 2 convolution kernel size, the volume of the second convolutional layer 2 Product core size can be first volume lamination 1 convolution kernel size and the first convolutional layer 4 convolution kernel size average value.

In a kind of achievable mode, preset rules can be the convolution kernel of each second convolutional layer of preset in advance.

Such as：There are 2 the second convolutional layers, respectively the second convolutional layer 1 and the second convolutional layer 2, can be with preset in advance second The convolution kernel of convolutional layer 1 is：The convolution kernel of the second convolutional layer of preset in advance 2 is：Each second When convolutional layer carries out convolution algorithm, only convolution kernel corresponding to second convolutional layer need to be used.

In a kind of achievable mode, preset rules can be：The convolution kernel of each second convolutional layer can be any one The convolution kernel of first convolutional layer.

Such as：There are 3 the first convolutional layers in first convolution module, be respectively：First convolutional layer 1, the first convolutional layer 2 and One convolutional layer 3；Wherein, the convolution kernel of the first convolutional layer 1 isThe convolution kernel of first convolutional layer 2 isThe convolution kernel of first convolutional layer 3 isThere are 2 the second convolutional layers in second convolution module, be respectively Second convolutional layer 1 and the second convolutional layer 2, then the convolution kernel of the second convolutional layer 1 can be in the convolution kernel of 3 the first convolutional layers Any one, that is, can be：OrSimilarly, the convolution kernel of the second convolutional layer 2 can also be 3 Any one in the convolution kernel of individual first convolutional layer, that is, can be：Or

The technical scheme provided by this disclosed embodiment can include the following benefits：Determined by preset rules each The convolution kernel of second convolutional layer, it can effectively lift the display effect of second feature figure.

During the first convolutional layer or the second convolutional layer carry out convolution, the convolution kernel movement of each first convolutional layer Step-length can differ, likewise, the step-length of the convolution kernel movement of each second convolutional layer can also differ, but using During this kind of step-length carries out convolution, it can make it that operand is larger, so that image processing efficiency is relatively low.

In one embodiment, in order to reduce operand, and image processing efficiency is lifted, the convolution of each first convolutional layer The step-length all same of core movement；The step-length all same of the corresponding convolution kernel movement of each second convolutional layer.

The technical scheme provided by this disclosed embodiment can include the following benefits：By setting each first convolution The step-length all same of the convolution kernel movement of layer；The step-length all same of the corresponding convolution kernel movement of each second convolutional layer, from can be with Operand is effectively reduced, and lifts image processing efficiency.

Because full convolutional network model is important models realizing function of image segmentation, therefore, in one embodiment In, when the default network model in above-mentioned each embodiment is full convolutional network model, the above method also includes：To step Characteristic pattern is classified pixel-by-pixel after the processing that S103 is obtained, and obtains carrying out the image after image segmentation.

It is worth noting that, the embodiment of the present disclosure include but is not limited to by way of softmax graders to processing after Characteristic pattern is classified pixel-by-pixel.

The technical scheme provided by this disclosed embodiment can include the following benefits：Use with fraction step-length Move Volumes The mode of product core realizes the purpose of up-sampling, and the characteristic pattern after handle, and then by the characteristic pattern progress after processing Classification obtains carrying out the image after image segmentation pixel-by-pixel, due to realizing up-sampling in a manner of fraction step-length Move Volumes product core Purpose compared in a manner of deconvolution efficiency it is higher, so as to effectively lifted image segmentation efficiency.

Portrait segmentation is an important subdomains of image segmentation, and mobile phone camera can be split by portrait, be background blurring Two steps simulate the background blurring effect of slr camera.Full convolutional network (FCN) is realize portrait dividing function one Important algorithm, in order to improve the low problem of efficiency that FCN networks are brought using deconvolution, the disclosure proposes that one kind is based on fraction The realization of the convolution operation of step-length, the operation still fall within the category of convolution operation, but can reach as deconvolution operation Function, while can save deconvolution operation low efficiency problem.Efficient portrait partitioning algorithm, significantly improves hand The experience of taking pictures of machine camera.

Figure 11 is the application scenario diagram of the embodiment of the present disclosure method according to an exemplary embodiment one, such as Shown in Figure 11, it is assumed that the quantity M of the first convolutional layer in the first convolution module is 5, is respectively：First convolutional layer 1, One convolutional layer 2, the first convolutional layer 3, the first convolutional layer 4 and the first convolutional layer 5；.Convolution kernel in each first convolutional layer Mobile step-length E_iIt is 2；The quantity N of second convolutional layer is 1, that is, in the present embodiment by fraction step-length The operation of convolution make it that the resolution ratio of second feature figure is identical with the resolution ratio of pending image.Pass through constraintsUnderstand, the step-length of the convolution kernel of the second convolutional layer in the second convolution module is

In the convolution kernel of each first convolutional layer of the first convolution module, it is volume Two to select the maximum convolution kernel of size The convolution kernel size of the second convolutional layer in volume module.

In the first convolution module after 5 process of convolution, the size of fisrt feature figure is relative to pending image 32 times of size reduction, second convolutional layer in the second convolution module is used to fisrt feature figure in the second convolution module When convolution kernel performs convolution operation, withStep-length movement convolution kernel, the size of the second feature figure now obtained is relative to the The size of one characteristic pattern is exaggerated 32 times, that is, the size of the second feature figure now obtained and the size phase of pending image Together.

Due to having carried out multiple convolution operation in the first convolution module, if accumulating core by a fraction step-length Move Volumes Convolution operation just obtain the second feature image that there is identical size with pending image, the second feature image now obtained Although having identical size with pending image, second feature image may may also be relatively rough, and therefore, Figure 12 is basis The application scenario diagram of embodiment of the present disclosure method shown in one exemplary embodiment two, as shown in figure 12, on the basis of Figure 11, The quantity N=5 of the second convolutional layer in the present embodiment in the second convolution module, it is respectively：Second convolutional layer 1, the second convolutional layer 2nd, the second convolutional layer 3, the second convolutional layer 4 and the second convolutional layer 5.

Because the quantity M of the first convolutional layer is 5, the step-length E of the convolution kernel of each first convolutional layer_iIt is 2, it is assumed that second The step-length of the convolution kernel movement of each second convolutional layer in convolution module is identical, passes through constraintsUnderstand, the convolution kernel movement of each second convolutional layer in the second convolution module Step-length can be

And determine that the convolution kernel of the second convolutional layer 1 is identical with the convolution kernel of the first convolutional layer 5, the convolution of the second convolutional layer 2 Core is identical with the convolution kernel of the first convolutional layer 4, and the convolution kernel of the second convolutional layer 3 is identical with the convolution kernel of the first convolutional layer 3, and second The convolution kernel of convolutional layer 4 is identical with the convolution kernel of the first convolutional layer 2, the volume of the convolution kernel of the second convolutional layer 5 and the first convolutional layer 1 Product nuclear phase is same.

In the first convolution module after 5 convolution operations, the size of fisrt feature image is relative to pending figure 32 times of the size reduction of picture, fisrt feature image is performed using the convolution kernel of the second convolutional layer in each second convolutional layer During convolution operation, respectively withStep-length movement convolution kernel, when using the second convolutional layer convolution kernel carry out convolution process In, computing is carried out to the element of fisrt feature figure using bilinear interpolation algorithm, the fisrt feature figure after interpolation processing is obtained, adopts Convolution operation is carried out to the fisrt feature figure after difference processing with the second convolution module.Then, the second feature figure finally given The size of picture is exaggerated 32 times relative to the size of fisrt feature image, that is, the size of the second feature image now obtained It is identical with the size of pending image.

Following is embodiment of the present disclosure, can be used for performing embodiments of the present disclosure.

Figure 13 is a kind of block diagram of image processing apparatus according to an exemplary embodiment one.As shown in figure 13, should Image processing apparatus includes：

First acquisition module 11, feature is carried out for treating processing image using the first convolution module of default network model Extraction, obtain the fisrt feature figure of the pending image；The size of the fisrt feature figure is less than the pending image, institute Stating the first convolution module includes at least one first convolutional layer；

Second acquisition module 12, mould is obtained to described first for the second convolution module using the default network model The fisrt feature figure that block 11 obtains is up-sampled, and obtains second feature figure, the size of the second feature figure is more than institute State fisrt feature figure；Second convolution module includes the second convolutional layer, the volume of the second convolutional layer in second convolution module The step-length of product core movement is proper fraction；

3rd acquisition module 13, for the second feature figure obtained according to second acquisition module 12, obtain with Characteristic pattern after the equal processing of the pending picture size.

In one embodiment, first convolution module also includes at least one pond layer, the quantity of the pond layer No more than the quantity of first convolutional layer, each pond layer is arranged between two first convolutional layers, and described the The step-length of convolution kernel movement in one convolution module is the integer more than or equal to 1.

In one embodiment, as shown in figure 14, described device also includes：Interpolating module 14, second acquisition module 12 include：Up-sample submodule 121；

The interpolating module 14, for first acquisition module 11 is obtained using bilinear interpolation algorithm described the The element of one characteristic pattern carries out computing, obtains the fisrt feature figure after interpolation processing；

The up-sampling submodule 121, it is described for being carried out using second convolution module to the interpolating module 14 Fisrt feature figure after difference processing is up-sampled.

In one embodiment, as shown in figure 15, second acquisition module 12 includes：Convolution submodule 122；

The convolution submodule 122, for the convolution kernel of second convolutional layer and first acquisition module to be obtained The fisrt feature figure carry out convolution.

In one embodiment, as shown in figure 16, described device also includes：First determining module 15；

First determining module 15, for determined according to following formula second convolutional layer quantity and it is each described in The step-length that the second convolution kernel moves in second convolutional layer：

In one embodiment, as shown in figure 17, when second convolution module includes second convolutional layer, institute Stating device also includes：4th acquisition module 16 and the second determining module 17；

4th acquisition module 16, for obtaining in first convolution module corresponding to each first convolutional layer Convolution kernel；

Second determining module 17, chi in first convolutional layer obtained for determining the 4th acquisition module 16 The size of very little maximum convolution kernel is the convolution kernel size of second convolutional layer.

In one embodiment, as shown in figure 18, when second convolution module includes at least two second convolutional layers When, described device also includes：3rd determining module 18；

3rd determining module 18, for the volume in each first convolutional layer in first convolution module Product core and preset rules determine the convolution kernel of each second convolutional layer in second convolution module；

Wherein, the preset rules comprise at least at least one of following rule：

The convolution kernel in each first convolutional layer in first convolution module is ranked up according to preset order, The convolution kernel in first convolutional layer after sequence is used as each second convolutional layer in second convolution module successively Convolution kernel；

Or the convolution kernel size of each second convolutional layer is the first volume product module in second convolution module The average value of convolution kernel size in block in first convolutional layer of predetermined number.

In one embodiment, as shown in figure 19, the default network model is full convolutional network, and described device is also wrapped Include：Sort module 19；

The sort module 19, carried out pixel-by-pixel for characteristic pattern after the processing that is obtained to the 3rd acquisition module Classification, obtain carrying out the image after image segmentation.

Processor；

For storing the memory of processor-executable instruction；

Wherein, processor is configured as：

Above-mentioned processor is also configured to：

First convolution module also includes at least one pond layer, and the quantity of the pond layer is not more than the first volume The quantity of lamination, each pond layer are arranged between two first convolutional layers, the volume in first convolution module The step-length of product core movement is the integer more than or equal to 1.

Before being up-sampled to the fisrt feature figure, methods described also includes：

Computing is carried out to the element of the fisrt feature figure using bilinear interpolation algorithm, obtains first after interpolation processing Characteristic pattern；

Second convolution module using the default network model carries out up-sampling to the fisrt feature figure to be included：

The fisrt feature figure after difference processing is up-sampled using second convolution module.

The convolution kernel of second convolutional layer and the fisrt feature figure are subjected to convolution.

Methods described also includes：The quantity of second convolutional layer and each second convolution are determined according to following formula The step-length that the second convolution kernel moves in layer：

When second convolution module includes second convolutional layer, methods described also includes：

Obtain each convolution kernel corresponding to first convolutional layer in first convolution module；

The size for determining the convolution kernel that size is maximum in first convolutional layer is the convolution kernel chi of second convolutional layer It is very little.

When second convolution module includes at least two second convolutional layers, methods described also includes：

Described in convolution kernel and preset rules in each first convolutional layer in first convolution module determine The convolution kernel of each second convolutional layer in second convolution module；

Wherein, the preset rules comprise at least at least one of following rule：

The default network model is full convolutional network, and methods described also includes：

Characteristic pattern after the processing is classified pixel-by-pixel, obtains carrying out the image after image segmentation.On above-mentioned reality The device in example is applied, wherein modules perform the concrete mode operated and carried out in the embodiment about this method in detail Thin description, will be not set forth in detail explanation herein.

Figure 20 is a kind of block diagram for image processing apparatus 80 according to an exemplary embodiment, and the device is applicable In terminal device.For example, device 80 can be mobile phone, and computer, digital broadcast terminal, messaging devices, game control Platform processed, tablet device, Medical Devices, body-building equipment, personal digital assistant etc..

Device 80 can include following one or more assemblies：Processing component 802, memory 804, power supply module 806 are more Media component 808, audio-frequency assembly 810, the interface 812 of input/output (I/O), sensor cluster 814, and communication component 816。

The integrated operation of the usual control device 80 of processing component 802, such as communicated with display, call, data, camera The operation that operation and record operation are associated.Processing component 802 can carry out execute instruction including one or more processors 820, To complete all or part of step of above-mentioned method.In addition, processing component 802 can include one or more modules, it is easy to Interaction between processing component 802 and other assemblies.For example, processing component 802 can include multi-media module, to facilitate more matchmakers Interaction between body component 808 and processing component 802.

Memory 804 is configured as storing various types of data to support the operation in device 80.These data are shown Example includes the instruction of any application program or method for being operated on device 80, contact data, telephone book data, disappears Breath, picture, video etc..Memory 804 can be by any kind of volatibility or non-volatile memory device or their group Close and realize, as static RAM (SRAM), Electrically Erasable Read Only Memory (EEPROM) are erasable to compile Journey read-only storage (EPROM), programmable read only memory (PROM), read-only storage (ROM), magnetic memory, flash Device, disk or CD.

Power supply module 806 provides electric power for the various assemblies of device 80.Power supply module 806 can include power management system System, one or more power supplys, and other components associated with generating, managing and distributing electric power for device 80.

Multimedia groupware 808 is included in the screen of one output interface of offer between described device 80 and user.One In a little embodiments, screen can include liquid crystal display (LCD) and touch panel (TP).If screen includes touch panel, screen Curtain may be implemented as touch-screen, to receive the input signal from user.Touch panel includes one or more touch sensings Device is with the gesture on sensing touch, slip and touch panel.The touch sensor can not only sensing touch or sliding action Border, but also detect and touched or the related duration and pressure of slide with described.In certain embodiments, more matchmakers Body component 808 includes a front camera and/or rear camera.When device 80 is in operator scheme, such as screening-mode or During video mode, front camera and/or rear camera can receive outside multi-medium data.Each front camera and Rear camera can be a fixed optical lens system or have focusing and optical zoom capabilities.

Audio-frequency assembly 810 is configured as output and/or input audio signal.For example, audio-frequency assembly 810 includes a Mike Wind (MIC), when device 80 is in operator scheme, during such as call model, logging mode and speech recognition mode, microphone is configured To receive external audio signal.The audio signal received can be further stored in memory 804 or via communication component 816 send.In certain embodiments, audio-frequency assembly 810 also includes a loudspeaker, for exports audio signal.

I/O interfaces 812 provide interface between processing component 802 and peripheral interface module, and above-mentioned peripheral interface module can To be keyboard, click wheel, button etc..These buttons may include but be not limited to：Home button, volume button, start button and lock Determine button.

Sensor cluster 814 includes one or more sensors, for providing the state estimation of various aspects for device 80. For example, sensor cluster 814 can detect opening/closed mode of device 80, the relative positioning of component, such as the component For the display and keypad of device 80, sensor cluster 814 can be with the position of 80 1 components of detection means 80 or device Change, the existence or non-existence that user contacts with device 80, the orientation of device 80 or acceleration/deceleration and the temperature change of device 80. Sensor cluster 814 can include proximity transducer, be configured to detect object nearby in no any physical contact Presence.Sensor cluster 814 can also include optical sensor, such as CMOS or ccd image sensor, in imaging applications Use.In certain embodiments, the sensor cluster 814 can also include acceleration transducer, gyro sensor, magnetic sensing Device, pressure sensor or temperature sensor.

Communication component 816 is configured to facilitate the communication of wired or wireless way between device 80 and other equipment.Device 80 can access the wireless network based on communication standard, such as WiFi, 2G or 3G, or combinations thereof.In an exemplary implementation In example, communication component 816 receives broadcast singal or broadcast related information from external broadcasting management system via broadcast channel. In one exemplary embodiment, the communication component 816 also includes near-field communication (NFC) module, to promote junction service.Example Such as, in NFC module radio frequency identification (RFID) technology can be based on, Infrared Data Association (IrDA) technology, ultra wide band (UWB) technology, Bluetooth (BT) technology and other technologies are realized.

In the exemplary embodiment, device 80 can be believed by one or more application specific integrated circuits (ASIC), numeral Number processor (DSP), digital signal processing appts (DSPD), PLD (PLD), field programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic building bricks are realized, for performing the above method.

In the exemplary embodiment, a kind of non-transitorycomputer readable storage medium including instructing, example are additionally provided Such as include the memory 804 of instruction, above-mentioned instruction can be performed to complete the above method by the processor 820 of device 80.For example, institute State non-transitorycomputer readable storage medium can be ROM, random access memory (RAM), CD-ROM, tape, floppy disk and Optical data storage devices etc..

A kind of non-transitorycomputer readable storage medium, when the instruction in the storage medium is by the processor of device 80 During execution so that device 80 is able to carry out above-mentioned image processing method, and methods described includes：

Wherein, the preset rules comprise at least at least one of following rule：

Characteristic pattern after the processing is classified pixel-by-pixel, obtains carrying out the image after image segmentation.

Figure 21 is a kind of block diagram of device 90 for image procossing according to an exemplary embodiment.For example, dress Put 90 and may be provided in a server.Device 90 includes processing component 902, and it further comprises one or more processors, And as the memory resource representated by memory 903, can be by the instruction of the execution of processing component 902 for storing, such as should Use program.The application program stored in memory 903 can include it is one or more each correspond to one group of instruction Module.In addition, processing component 902 is configured as execute instruction, to perform the above method.

Device 90 can also include the power management that a power supply module 906 is configured as performs device 90, and one wired Or radio network interface 905 is configured as device 90 being connected to network, and input and output (I/O) interface 908.Device 90 It can operate based on the operating system for being stored in memory 903, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or similar.

A kind of non-transitorycomputer readable storage medium, when the instruction in the storage medium is by the processor of device 90 During execution so that device 90 is able to carry out above-mentioned image processing method, and methods described includes：

Wherein, the preset rules comprise at least at least one of following rule：

Those skilled in the art will readily occur to the disclosure its after considering specification and putting into practice disclosure disclosed herein Its embodiment.The application is intended to any modification, purposes or the adaptations of the disclosure, these modifications, purposes or Person's adaptations follow the general principle of the disclosure and including the undocumented common knowledges in the art of the disclosure Or conventional techniques.Description and embodiments are considered only as exemplary, and the true scope of the disclosure and spirit are by following Claim is pointed out.

It should be appreciated that the precision architecture that the disclosure is not limited to be described above and is shown in the drawings, and And various modifications and changes can be being carried out without departing from the scope.The scope of the present disclosure is only limited by appended claim.

Claims

A kind of 1. image processing method, it is characterised in that including：

Processing image is treated using the first convolution module of default network model and carries out feature extraction, obtains the pending image Fisrt feature figure；The size of the fisrt feature figure is less than the pending image, and first convolution module is included at least One the first convolutional layer；

The fisrt feature figure is up-sampled using the second convolution module of the default network model, obtains second feature Figure, the size of the second feature figure are more than the fisrt feature figure；Second convolution module includes the second convolutional layer, described The step-length of the convolution kernel movement of the second convolutional layer is proper fraction in second convolution module；

According to the second feature figure, characteristic pattern after the processing equal with the pending picture size is obtained.
2. according to the method for claim 1, it is characterised in that first convolution module also includes at least one pond Layer, the quantity of the pond layer are not more than the quantity of first convolutional layer, and each the pond layer is arranged at two described the Between one convolutional layer, the step-length of the convolution kernel movement in first convolution module is the integer more than or equal to 1.
3. according to the method for claim 1, it is characterised in that before being up-sampled to the fisrt feature figure, institute Stating method also includes：

Computing is carried out to the element of the fisrt feature figure using bilinear interpolation algorithm, obtains the fisrt feature after interpolation processing Figure；

Second convolution module using the default network model carries out up-sampling to the fisrt feature figure to be included：

The fisrt feature figure after difference processing is up-sampled using second convolution module.
4. according to the method described in any one of claims 1 to 3, it is characterised in that described using the default network model Second convolution module carries out up-sampling to the fisrt feature figure to be included：

The convolution kernel of second convolutional layer and the fisrt feature figure are subjected to convolution.
5. according to the method described in any one of claims 1 to 3, it is characterised in that methods described also includes：According to following formula Determine the quantity of second convolutional layer and the step-length that each the second convolution kernel moves in second convolutional layer：

<mrow> <munderover> <mo>&Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <mi>log</mi> <mi> </mi> <msub> <mi>E</mi> <mi>i</mi> </msub> <mo>+</mo> <munderover> <mo>&Sigma;</mo> <mrow> <mi>j</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mi>log</mi> <mi> </mi> <msub> <mi>F</mi> <mi>j</mi> </msub> <mo>=</mo> <mn>0</mn> </mrow>

Wherein, the M be first convolutional layer quantity, the N be the second convolutional layer quantity, the F_jFor j-th second The step-length of convolution kernel movement in convolutional layer, the E_iFor the step-length of the convolution kernel movement in i-th of first convolutional layers.
6. according to the method described in any one of claims 1 to 3, it is characterised in that when second convolution module includes one During individual second convolutional layer, methods described also includes：

Obtain each convolution kernel corresponding to first convolutional layer in first convolution module；

The size for determining the convolution kernel that size is maximum in first convolutional layer is the convolution kernel size of second convolutional layer.
7. according to the method described in any one of claims 1 to 3, it is characterised in that when second convolution module include to During few two the second convolutional layers, methods described also includes：

Convolution kernel and preset rules in each first convolutional layer in first convolution module determine described second The convolution kernel of each second convolutional layer in convolution module；

Wherein, the preset rules comprise at least any of following rule：

The convolution kernel in each first convolutional layer in first convolution module is ranked up according to preset order, sorted The volume as each second convolutional layer in second convolution module successively of the convolution kernel in first convolutional layer afterwards Product core；

Or the convolution kernel size of each second convolutional layer is in first convolution module in second convolution module The average value of convolution kernel size in first convolutional layer of predetermined number.
8. according to the method described in any one of claims 1 to 3, it is characterised in that the default network model is full convolution net Network, methods described also include：

Characteristic pattern after the processing is classified pixel-by-pixel, obtains carrying out the image after image segmentation.
A kind of 9. image processing apparatus, it is characterised in that including：

First acquisition module, feature extraction is carried out for treating processing image using the first convolution module of default network model, Obtain the fisrt feature figure of the pending image；The size of the fisrt feature figure is less than the pending image, and described One convolution module includes at least one first convolutional layer；

Second acquisition module, for being obtained using the second convolution module of the default network model to first acquisition module The fisrt feature figure up-sampled, obtain second feature figure, it is special that the size of the second feature figure is more than described first Sign figure；Second convolution module includes the second convolutional layer, the convolution kernel movement of the second convolutional layer in second convolution module Step-length be proper fraction；

3rd acquisition module, for the second feature figure obtained according to second acquisition module, obtain waiting to locate with described Characteristic pattern after the equal processing of reason picture size.
10. device according to claim 9, it is characterised in that first convolution module also includes at least one pond Layer, the quantity of the pond layer are not more than the quantity of first convolutional layer, and each the pond layer is arranged at two described the Between one convolutional layer, the step-length of the convolution kernel movement in first convolution module is the integer more than or equal to 1.
11. device according to claim 9, it is characterised in that described device also includes：Interpolating module, described second obtains Modulus block includes：Up-sample submodule；

The interpolating module, for the fisrt feature figure obtained using bilinear interpolation algorithm to first acquisition module Element carry out computing, obtain the fisrt feature figure after interpolation processing；

The up-sampling submodule, after carrying out the difference processing to the interpolating module using second convolution module Fisrt feature figure up-sampled.
12. according to the device described in any one of claim 9 to 11, it is characterised in that second acquisition module includes：Convolution Submodule；

The convolution submodule, for the convolution kernel of second convolutional layer and first acquisition module are obtained described the One characteristic pattern carries out convolution.
13. according to the device described in any one of claim 9 to 11, it is characterised in that described device also includes：First determines mould Block；

First determining module, for the quantity that second convolutional layer is determined according to following formula and each volume Two The step-length that the second convolution kernel moves in lamination：

<mrow> <munderover> <mo>&Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <mi>log</mi> <mi> </mi> <msub> <mi>E</mi> <mi>i</mi> </msub> <mo>+</mo> <munderover> <mo>&Sigma;</mo> <mrow> <mi>j</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mi>log</mi> <mi> </mi> <msub> <mi>F</mi> <mi>j</mi> </msub> <mo>=</mo> <mn>0</mn> </mrow>

Wherein, the M be first convolutional layer quantity, the N be the second convolutional layer quantity, the F_jFor j-th second The step-length of convolution kernel movement in convolutional layer, the E_iFor the step-length of the convolution kernel movement in i-th of first convolutional layers.
14. according to the device described in any one of claim 9 to 11, it is characterised in that when second convolution module includes During one the second convolutional layer, described device also includes：4th acquisition module and the second determining module；

4th acquisition module, for obtaining each convolution corresponding to first convolutional layer in first convolution module Core；

Second determining module, size is maximum in first convolutional layer obtained for determining the 4th acquisition module The size of convolution kernel is the convolution kernel size of second convolutional layer.
15. according to the device described in any one of claim 9 to 11, it is characterised in that when second convolution module includes During at least two second convolutional layers, described device also includes：3rd determining module；

3rd determining module, for the convolution kernel in each first convolutional layer in first convolution module and Preset rules determine the convolution kernel of each second convolutional layer in second convolution module；

Wherein, the preset rules comprise at least at least one of following rule：

The convolution kernel in each first convolutional layer in first convolution module is ranked up according to preset order, sorted The volume as each second convolutional layer in second convolution module successively of the convolution kernel in first convolutional layer afterwards Product core；

Or the convolution kernel size of each second convolutional layer is in first convolution module in second convolution module The average value of convolution kernel size in first convolutional layer of predetermined number.
16. according to the device described in any one of claim 9 to 11, it is characterised in that the default network model is full convolution Network, described device also include：Sort module；

The sort module, classified pixel-by-pixel for characteristic pattern after the processing that is obtained to the 3rd acquisition module, Obtain carrying out the image after image segmentation.
A kind of 17. image processing apparatus, it is characterised in that including：

Processor；

For storing the memory of processor-executable instruction；

Wherein, the processor is configured as：

Processing image is treated using the first convolution module of default network model and carries out feature extraction, obtains the pending image Fisrt feature figure；The size of the fisrt feature figure is less than the pending image, and first convolution module is included at least One the first convolutional layer；

The fisrt feature figure is up-sampled using the second convolution module of the default network model, obtains second feature Figure, the size of the second feature figure are more than the fisrt feature figure；Second convolution module includes the second convolutional layer, described The step-length of the convolution kernel movement of the second convolutional layer is proper fraction in second convolution module；

According to the second feature figure, characteristic pattern after the processing equal with the pending picture size is obtained.
18. a kind of computer-readable recording medium, is stored thereon with computer instruction, it is characterised in that the instruction is by processor Following steps are realized during execution：

Processing image is treated using the first convolution module of default network model and carries out feature extraction, obtains the pending image Fisrt feature figure；The size of the fisrt feature figure is less than the pending image, and first convolution module is included at least One the first convolutional layer；

The fisrt feature figure is up-sampled using the second convolution module of the default network model, obtains second feature Figure, the size of the second feature figure are more than the fisrt feature figure；Second convolution module includes the second convolutional layer, described The step-length of the convolution kernel movement of the second convolutional layer is proper fraction in second convolution module；

According to the second feature figure, characteristic pattern after the processing equal with the pending picture size is obtained.