CN108230257A

CN108230257A - Image processing method, device, electronic equipment and storage medium

Info

Publication number: CN108230257A
Application number: CN201711131131.3A
Authority: CN
Inventors: 苏鑫; 秦红伟
Original assignee: Beijing Sensetime Technology Development Co Ltd
Current assignee: Beijing Sensetime Technology Development Co Ltd
Priority date: 2017-11-15
Filing date: 2017-11-15
Publication date: 2018-06-29

Abstract

The application embodiment discloses image processing method, device, electronic equipment and storage medium, and method therein includes：The equalization processing based on gray scale is carried out to pending image, the equalization processing based on gray scale is used to increase the grey level region that the pixel value of the pending image is distributed；Pending image after equalization processing is supplied to neural network, and the handling result of the pending image is exported through the neural network.

Description

Image processing method, device, electronic equipment and storage medium

Technical field

This application involves computer vision technique, more particularly, to a kind of image processing method, image processing apparatus, nerve The training method of network, the training device of neural network, electronic equipment and computer readable storage medium.

Background technology

Image procossing generally includes：Recognition of face, face location detection (the external frame detection of such as face), the inspection of human body key point Survey, face critical point detection and In vivo detection processing etc..

The factors such as the dark either backlight of ambient light or complex illumination, can cause the quality of image to be declined.For There is the image being short of to a certain degree in picture quality, how to promote the accuracy of processing result image, be one and merit attention Technical problem.

Invention content

The application embodiment provides a kind of image processing techniques scheme.

According to the application embodiment one aspect, a kind of image processing method is provided, this method includes：It treats Processing image carries out the equalization processing based on gray scale, and the equalization processing based on gray scale is used to increase the pending figure The grey level region that the pixel value of picture is distributed；Pending image after equalization processing is supplied to neural network, and pass through The neural network exports the handling result of the pending image.

In one embodiment of the application, it is described the equalization processing based on gray scale is carried out to pending image before, packet It includes：Judge whether the grey level region that the pixel value of pending image is distributed is less than first threshold, treated if so, performing The step of processing image carries out the equalization processing based on gray scale.

In the another embodiment of the application, grey level region that the pixel value for judging pending image is distributed Whether it is less than before first threshold, including：Judge whether the corresponding pixel quantity of any pixel value reaches in pending image Second threshold, if so, corresponding pixel value to be included in the range in the grey level region, if it is not, then not by corresponding picture Plain value is included in the range in grey level region.

In the application a further embodiment, the pending image includes：Gray level image or non-gray level image.

It is described the equalization processing based on gray scale is carried out to pending image to include in the application a further embodiment： To carrying out histogram equalization processing by the gray level image as pending image.

It is described the equalization processing based on gray scale is carried out to pending image to include in the application a further embodiment： It is gray level image, and transformed gray level image is equal into column hisgram using by the non-greyscale image transitions as pending image Weighing apparatusization processing.

It is described the equalization processing based on gray scale is carried out to pending image to include in the application a further embodiment： To be brightness coloration YUV image by the non-greyscale image transitions as pending image, to the Y channel pictures of the YUV image Element value carries out histogram equalization processing, and the YUV image after equalization processing is converted to the form of the non-gray level image.

In the application a further embodiment, the non-gray level image includes：RGB RGB image.

In the application a further embodiment, the handling result that the pending image is exported through the neural network Including at least one following：The face recognition result of the pending image is exported through the neural network；Through the nerve net Network exports the face location testing result of the pending image；The face of the pending image is exported through the neural network Critical point detection result；The human body critical point detection result of the pending image is exported through the neural network；Through the god The position of human body testing result of the pending image is exported through network；The pending image is exported through the neural network Human action testing result；The gestures detection result of the pending image is exported through the neural network；Through the nerve net Network exports the hand position testing result of the pending image；The live body that the pending image is exported through the neural network is examined Survey result.

In the application a further embodiment, the method further includes：The training neural network, and the training nerve The process of network includes：The image pattern concentrated according to training image obtains the input picture of neural network to be trained；By described in Input picture is supplied to neural network to be trained, and the processing knot of neural network output the to be trained input picture described in warp Fruit；The labeled data for the image pattern concentrated using the training image carries out the neural network to be trained as tutorial message Supervised learning, so that the handling result convergence of the neural network output to be trained.

In the application a further embodiment, described image sample includes：Gray level image sample or non-gray-scale map are decent This.

In the application a further embodiment, the image pattern concentrated according to training image obtains nerve net to be trained The input picture of network includes at least one following：It is concentrated from training image and obtains image pattern, using described image sample as treating The input picture of training neural network；It is concentrated from training image and obtains image pattern, described image sample carried out based on gray scale Equalization processing, the image pattern after the equalization processing is by the input picture as neural network to be trained.

It is described that the equalization processing based on gray scale is carried out to described image sample in the application yet another embodiment, Image pattern after the equalization processing is included by the input picture as neural network to be trained：To gray level image sample into Column hisgram equalization processing, the gray level image sample after the histogram equalization processing is by as neural network to be trained Input picture.

It is described that the equalization processing based on gray scale is carried out to described image sample in the application yet another embodiment, Image pattern after the equalization processing is included by the input picture as neural network to be trained：By non-gray level image sample Gray level image sample is converted to, and transformed gray level image sample is subjected to histogram equalization processing, the histogram is equal Weighing apparatusization treated gray level image sample is by the input picture as neural network to be trained.

It is described that the equalization processing based on gray scale is carried out to described image sample in the application yet another embodiment, Image pattern after the equalization processing is included by the input picture as neural network to be trained：By non-gray level image sample YUV image sample is converted to, histogram equalization processing is carried out, and will be balanced to the Y channels pixel value of the YUV image sample Changing treated, YUV image sample be converted to the form of the non-gray level image sample, the transformed non-gray-scale map of form Decent by the input picture as neural network to be trained.

In the application a further embodiment, the non-gray level image sample includes：RGB image sample.

In the application a further embodiment, the neural network includes：For extracting the first nerves net of characteristics of image Network and the nervus opticus network for identified/detected target object, the pending image after the equalization processing are input to It is described to be used to extract the image that the first nerves network of characteristics of image is formed for extracting in the first nerves network of characteristics of image Feature is input in the nervus opticus network for identified/detected target object.

It is described to include for extracting the first nerves network of characteristics of image in the application yet another embodiment：Convolution Neural network, the nervus opticus network for identified/detected target object include：Convolutional neural networks or including at least The full Connection Neural Network of one layer of full articulamentum.

In terms of according to the other in which of the application embodiment, a kind of training method of neural network is provided, it is described Method includes：The image pattern concentrated according to training image obtains the input picture of neural network to be trained；The input is schemed As being supplied to neural network to be trained, and through the handling result of neural network the to be trained output input picture；With the instruction Practice the labeled data of the image pattern in image set for tutorial message, treat trained neural network and exercise supervision study, so as to treat The handling result convergence of training neural network output.

In one embodiment of the application, described image sample includes：Gray level image sample or non-gray-scale map are decent This.

In the another embodiment of the application, the image pattern concentrated according to training image obtains nerve net to be trained The input picture of network includes at least one following：It is concentrated from training image and obtains image pattern, using described image sample as treating The input picture of training neural network；It is concentrated from training image and obtains image pattern, described image sample carried out based on gray scale Equalization processing, the image pattern after the equalization processing is by the input picture as neural network to be trained.

In terms of according to the other in which of the application embodiment, a kind of image processing apparatus is provided, the device is main Including：Equalization processing module, it is described based on the equal of gray scale for carrying out the equalization processing based on gray scale to pending image Weighing apparatusization handles to increase the grey level region that the pixel value of the pending image is distributed；Obtain processing result image mould Block, for the pending image after equalization processing to be supplied to neural network, and through waiting to locate described in neural network output Manage the handling result of image.

In one embodiment of the application, described device further includes：Judgment module, for judging the picture of pending image Whether the grey level region that plain value is distributed is less than first threshold, if so, triggering the equalization processing module performs institute It states and the equalization processing based on gray scale is carried out to pending image.

In the another embodiment of the application, described device further includes：Grey level regions module is determined, for judging to treat Whether the corresponding pixel quantity of any pixel value reaches second threshold in processing image, if so, corresponding pixel value is received Enter the range in the grey level region, if it is not, corresponding pixel value not to be included in then the range in the grey level region.

In the application a further embodiment, described device further includes：The training device of neural network, the neural network Training device include：Input picture unit is obtained, training nerve is treated for being obtained according to the image pattern that training image is concentrated The input picture of network；Input picture handling result unit is obtained, for the input picture to be supplied to nerve net to be trained Network, and through the handling result of neural network the to be trained output input picture；Supervised learning unit, for being schemed with the training The labeled data of image pattern in image set is tutorial message, treats trained neural network and exercises supervision study, so as to wait to train The handling result convergence of neural network output.

According to the application embodiment in another aspect, providing a kind of training device of neural network, described device packet It includes：Input picture unit is obtained, for obtaining the input figure of neural network to be trained according to the image pattern that training image is concentrated Picture；Input picture handling result unit is obtained, for the input picture to be supplied to neural network to be trained, and through waiting to train Neural network exports the handling result of the input picture；Supervised learning unit, for the image concentrated with the training image The labeled data of sample is tutorial message, treats trained neural network and exercises supervision study, so that neural network to be trained output Handling result convergence.

According to another aspect of the application embodiment, a kind of electronic equipment is provided, including：Memory, for depositing Store up computer program；Processor, for performing the computer program stored in the memory, and the computer program is held During row, each step in the application method embodiment is realized.

According to another aspect of the application embodiment, a kind of computer readable storage medium is provided, is stored thereon with Computer program when the computer program is executed by processor, performs each step in the application method embodiment.

According to another aspect of the application embodiment, a kind of computer program is provided, which is handled When device performs, each step in the application method embodiment is performed.

Based on image processing method, device, electronic equipment and computer readable storage medium that the application provides, this Shen Please by carrying out the equalization processing based on gray scale to pending image, it can increase what the pixel value of pending image was distributed Grey level region, in this way, the application can be eliminated to a certain extent due to half-light, backlight, overexposure and complex illumination Etc. the soft edge of target object caused by factors in pending image and grain details phenomena such as not knowing, so as to The influences of the factors to the quality of pending image such as half-light, backlight, overexposure and complex illumination are reduced to a certain extent, And then the picture for being showed the pending image after equalization processing is more clear, for example, after equalization processing being made Pending image in target object profile and grain details etc. be more clear it is apparent；The application is by will be at equalization Pending image after reason is supplied to neural network, and image is carried out for the pending image after equalization processing by neural network Processing, since the picture that the pending image after equalization processing is showed is more clear, be conducive to neural network and hold Row image processing operations.It follows that the image processing techniques that the application provides is conducive to improve the image procossing of neural network As a result accuracy.

Below by drawings and embodiments, the technical solution of the application is described in further detail.

Description of the drawings

The attached drawing of a part for constitution instruction describes presently filed embodiment, and is used to solve together with description Release the principle of the application.

With reference to attached drawing, according to following detailed description, the application can be more clearly understood, wherein：

Fig. 1 is the flow chart of one embodiment of image processing method of the application；

Fig. 2 is the flow chart of one embodiment of image processing method for being used to implement face location detection of the application；

Fig. 3 is the face location testing result schematic diagram that the neural network of the application exports；

Fig. 4 is the another face location testing result schematic diagram that the neural network of the application exports；

Fig. 5 is another face location testing result schematic diagram that the neural network of the application exports；

Fig. 6 is the flow chart of one embodiment of training method of the neural network of the application；

Fig. 7 is the schematic diagram of one embodiment of image processing apparatus of the application；

Fig. 8 is the schematic diagram of one embodiment of training device of the neural network of the application；

Fig. 9 is the block diagram for the example devices for realizing the application embodiment.

Specific embodiment

The various exemplary embodiments of the application are described in detail now with reference to attached drawing.It should be noted that：Unless in addition have Body illustrates that the unlimited system of component and the positioned opposite of step, numerical expression and the numerical value otherwise illustrated in these embodiments is originally The range of application.Simultaneously, it should be appreciated that for ease of description, the sizes of the various pieces shown in attached drawing be not according to What practical proportionate relationship was drawn.

It is illustrative to the description only actually of at least one exemplary embodiment below, is never used as to the application And its application or any restrictions that use.It can for technology, method and equipment known to person of ordinary skill in the relevant It can be not discussed in detail, but in the appropriate case, the technology, method and apparatus should be considered as part of specification.It should It notices：Similar label and letter represents similar terms in following attached drawing, therefore, once in a certain Xiang Yi attached drawing It is defined, then in subsequent attached drawing does not need to that it is further discussed.

The embodiment of the present application can be applied in the electronic equipments such as terminal device, computer system and server, can It is operated together with numerous other general or specialized computing system environments or configuration.Suitable for terminal device, computer system, service The example of well-known terminal device, computing system, environment and/or configuration that the electronic equipments such as device are used together is included but not It is limited to：Personal computer system, server computer system, thin client, thick client computer, hand-held or laptop devices, based on micro- The system of processor, set-top box, programmable consumer electronics, NetPC Network PC, minicomputer system, mainframe computer System and distributed cloud computing technology environment including any of the above described system etc..

The electronic equipments such as terminal device, computer system, server can be in the department of computer science performed by computer system It is described under the general linguistic context of system executable instruction (such as program module).In general, program module can include routine, program, mesh Beacon course sequence, component, logic, data structure etc., they perform specific task or realize specific abstract data type.Meter Calculation machine systems/servers can be implemented in distributed cloud computing environment, and in distributed cloud computing environment, task is by by logical What the remote processing devices of communication network link performed.In distributed cloud computing environment, program module can be located at and include storage On the Local or Remote computing system storage medium of equipment.

Exemplary embodiment

The technical solution of the image procossing provided with reference to Fig. 1 to Fig. 9 to the application and the skill of training neural network Art scheme illustrates.

Fig. 1 is the flow chart of image processing method one embodiment of the application.As shown in Figure 1, the embodiment method packet It includes：Step S100 and step S110.

S100, the equalization processing based on gray scale is carried out to pending image, to increase the pixel value of pending image institute The grey level region of distribution.

In an optional example, the pending image in the application can be that the figures such as static picture or photo are presented Picture, or video frame in dynamic video etc. is presented.It is used for recognition of face, people in the image processing techniques of the application Face position detection (such as face external frame detection) or face critical point detection in the relevant application of face in the case of, should Pending image has generally comprised face, and the face in pending image can be positive face, or side face.The application's Image processing techniques is used for human body critical point detection, and either position of human body detection or human action detection etc. are related to human body Application in the case of, which has generally comprised human body.It is used for gesture in the image processing techniques of the application Detection or hand position detection etc. in the relevant application of hand in the case of, which has generally comprised hand.In this Shen Image processing techniques please be used for In vivo detection etc. in the relevant application of live body in the case of, which usually wraps Contain live body (such as living person, live the animal that cat or dog living etc. live for another example).

In an optional example, the pending image in the application can be gray level image, or non-gray-scale map Picture.Non- gray level image can be RGB (RGB) image etc..Certainly, non-gray level image may be other than RGB image its The image of his form, such as YUV (brightness coloration) image.The application does not limit the specific manifestation form of pending image.

In an optional example, due to factors such as half-light, backlight, overexposure and complex illuminations, often make to wait to locate The picture quality of reason image is affected, and therefore, the presence of these factors is generally not favored neural network and performs image procossing behaviour Make；For example, the picture at least part image-region in pending image under half-light scene (including part half-light scene etc.) Plain value (pixel value of such as gray level image) is usually concentrated in relatively low value region range, this causes in pending image extremely The texture gradient of small part image-region is smaller, so as to make the target object (such as face or human body) in pending image Feature it is more fuzzy, this frequently can lead at the pending image progress image under using neural network for half-light scene During reason, exist the handling result that can not obtain handling result or acquisition accuracy rate it is relatively low the phenomenon that；For another example due to excessive The overall light exposed under scene is very bright, and the background light under backlight scene is very bright, the light under complex illumination scene The factors such as the diversity of line often make the target objects such as profile and the detail textures of target object in pending image At least part feature is more fuzzy, therefore, is utilizing neural network directly against fields such as overexposure, backlight or complex illuminations When pending image under scape carries out image procossing, often there are neural networks can not obtain handling result or neural network Handling result accuracy rate it is relatively low the phenomenon that.The application by carrying out the equalization processing based on gray scale to pending image, The indexs such as contrast and/or the brightness of pending image can be made to obtain relatively reasonable adjustment, so as to a certain degree On avoid influence of the factors such as half-light, overexposure, backlight and complex illumination to the picture quality of pending image, Jin Eryou Conducive to the accuracy for the processing result image for improving neural network.

In an optional example, the application can carry out the equalization based on gray scale to the whole region of pending image Processing, the application can also carry out the equalization processing based on gray scale to the subregion of pending image, for example, for waiting to locate The region of other in addition to outer rim in reason image carries out the equalization processing based on gray scale.

In an optional example, the application may be used histogram equalization processing mode come to pending image into Equalization processing of the row based on gray scale, the application can also be carried out based on the equal of gray scale pending image using other modes Weighing apparatusization processing, for example, being carried out pending image based on gray scale by way of directly adjusting the indexs such as contrast and/brightness Equalization processing.

One of the equalization processing based on gray scale is carried out to pending image by the way of histogram equalization processing A optional example is：It is gray level image (following to can be described as pending gray level image) in pending image, the application can Directly to carry out histogram equalization processing to pending gray level image, the gray level image after histogram equalization processing is this Pending image after equalization processing in application.

For the pending gray level image of half-light scene or overexposure scene, after equalization processing, The pixel value of pending gray level image be no longer limited in a small range region (in such as relatively low pixel value range region or compared with In high pixel value range region), but large-scale multiple grey level region (such as 0-255 ranges are expanded to by equal proportion Region), and be uniformly distributed, it is closer to so as to which the pending gray level image after equalization processing be made usually to have with normal picture Contrast and the indexs such as brightness；For example, for the original pending gray level image of half-light scene, equalization processing Pending gray level image afterwards usually has the indexs such as higher contrast and brightness；For another example compared to overexposure scene Original pending gray level image for, pending gray level image after equalization processing usually has relatively low contrast and bright The indexs such as degree.

For the pending gray level image of backlight scene or complex illumination scene etc., in pending gray level image The pixel value for crossing dark and excessively bright region is no longer limited in a small range region, but expands to big model by equal proportion The multiple grey level regions (such as 0-255 range areas) enclosed, and be uniformly distributed, it is excessively dark in pending gray level image so as to make Or excessively bright region has the indexs such as the contrast being closer to normal picture and brightness after via equalization processing.

In an optional example, above-mentioned normal picture typically refers to, the preferable image of picture quality, for example, face wheel Clean up clear apparent, and the clear apparent image of face grain details.

Seen from the above description, the picture of the pending gray level image after the equalization processing of the application can be apparent, mesh Marking the features such as profile and the detail textures of object can be more obvious, and image processing operations are performed so as to be conducive to neural network, And then be conducive to improve the accuracy of the processing result image of neural network.

The another of the equalization processing based on gray scale is carried out to pending image by the way of histogram equalization processing One optional example is：In the case where pending image is RGB image (following to be known as pending RGB image), first, will treat Processing RGB image is converted to YUV image, then, histogram equalization processing is carried out to the Y channels pixel value of YUV image, it Afterwards, the YUV image after equalization processing is converted into RGB image, above-mentioned transformed RGB image is the equilibrium in the application Change treated pending image.Since Y channel tables show brightness, i.e. grayscale value, therefore, the application passes through in YUV image Y channels pixel value carry out histogram equalization processing, the YUV image after equalization processing can be made to have in terms of gray scale The indexs such as the contrast being closer to normal picture and brightness, and RGB image is converted to by the YUV image after equalization processing Afterwards, the indexs such as contrast and brightness of the indexs such as the overall contrast of transformed RGB image and brightness and normal picture are more It is close.It follows that the picture of the RGB image after equalization processing can be more clear, the profile and texture of target object are thin The features such as section can be more obvious, performs image processing operations so as to be conducive to neural network, and then be conducive to improve neural network Processing result image accuracy.

It should be strongly noted that it is the non-gray level image of the other forms in addition to RGB image in pending image In the case of, the application, which equally may be used, to be first converted to YUV image by pending image and (is such as converted directly into YUV image, for another example RGB image is first converted to, then YUV image etc. is converted to by RGB image), then, the Y channels pixel value of YUV image is carried out straight Square figure equalization processing and then the YUV image after equalization processing is converted to corresponding form non-gray level image side Formula, to obtain the pending image after equalization processing.It is carried out for the non-gray level image of the other forms in addition to RGB image The process of equalization processing based on gray scale, no longer enumerates explanation herein.

The equalization processing based on gray scale is carried out again to pending image by the way of histogram equalization processing One optional example is：In the case where pending image is RGB image (following to be known as pending RGB image), first, will treat Processing RGB image is converted to gray level image, then, histogram equalization processing, histogram is carried out to transformed gray level image Gray level image after equalization processing is the pending image after the equalization processing in the application.Since pending RGB schemes As being converted into gray level image, and the picture of the pending gray level image after equalization processing can be more clear, target object The features such as profile and detail textures can be more obvious, and therefore, the application is conducive to neural network and performs image processing operations, into And be conducive to improve the accuracy of the processing result image of neural network.

It should be strongly noted that it is the non-gray level image of the other forms in addition to RGB image in pending image In the case of, the application equally may be used first is converted to gray level image by pending image, and to transformed gray level image into The mode of column hisgram equalization processing, to obtain the pending image after equalization processing.For its in addition to RGB image The non-gray level image of his form carries out the process of the equalization processing based on gray scale, no longer enumerates explanation herein.

In an optional example, the application, can before carrying out pending image based on the equalization processing of gray scale Whether the grey level region being distributed with the pixel value for first judging pending image is less than first threshold, if less than the first threshold Value, then perform the operation that the equalization processing based on gray scale is carried out to pending image；And if being not less than first threshold, it can No longer to perform the operation that the equalization processing based on gray scale is carried out to pending image.Above-mentioned grey level region can be reflected Go out situations such as pending image is with the presence or absence of half-light, backlight, overexposure and complex illumination.Certainly, the application can not also Above-mentioned judgement operation is performed, and the equalization processing based on gray scale directly is carried out to pending image.

In an optional example, the application can determine that the pixel value of pending image is distributed by following manner Grey level region：Judge any pixel value in pending image (for example, the gray value of pixel, for another example the R of pixel leads to Average value of the numerical value of road numerical value, G channels numerical value and channel B etc.) whether corresponding pixel quantity reach second threshold, such as Fruit reaches second threshold, then above-mentioned pixel value is brought into the range in grey level region, and if not reaching second threshold, Above-mentioned pixel value will not then be brought into the range in grey level region.

S110, the pending image after equalization processing is supplied to neural network, and exported through the neural network and wait to locate Manage the handling result of image.

In an optional example, the network structure of the neural network in the application may be used but be not limited to AlexNet, Depth residual error network (Deep Residual Network, ResNet) or VGGNet (Visual Geometry Group Network, visual geometric group network) etc. network structure used by neural networks.The neural network of the application can include：With First nerves network in extraction characteristics of image and the nervus opticus network for identified/detected target object.In the application Equalization processing after pending image be input in the first nerves network for extracting characteristics of image, for extract scheme As feature first nerves network for input the characteristics of image that is formed of image by as input information, be input to know In the nervus opticus network of not/detected target object, connect by the nervus opticus network for being used for identified/detected target object for it The characteristics of image received is further processed, to obtain and export the corresponding handling result of the characteristics of image, the processing knot Fruit is the handling result of pending image.

In an optional example, in the application can be specially to roll up for extracting the first nerves network of characteristics of image Product neural network.The network structure for being used to extract the convolutional neural networks of characteristics of image in the application can be according to extraction image The actual demand flexible design of feature, the application are not intended to limit this for extracting the specific net of the convolutional neural networks of characteristics of image Network structure；For example, at least a roll is can include but is not limited to for extracting the convolutional neural networks of characteristics of image in the application Lamination and at least a linear R eLU (Rectified Linear Units correct linear unit) layer etc., schemes for extracting As the number of plies that the convolutional neural networks of feature are included is more, then network is deeper.For extracting the first nerves net of characteristics of image Network can be deep neural network.

In an optional example, the nervus opticus network for identified/detected target object in the application can have Body is convolutional neural networks (for example, convolution kernel is 1 convolutional neural networks), or includes at least one layer of full articulamentum Full Connection Neural Network.The nervus opticus network for identified/detected target object in the application can be answered according to practical Image processing requirements in are flexibly set, so as to which neural network be enable to export the image procossing knot needed for practical application Fruit；For example, neural network exports face recognition result for pending image；For another example neural network is directed to pending image Export face location testing result (the external frame testing result of such as face)；For another example neural network is defeated for pending image Go out face critical point detection result；For another example neural network exports human body critical point detection result for pending image；Example again Such as, neural network exports position of human body testing result for pending image；For another example neural network is defeated for pending image Go out human action testing result；For another example neural network exports gestures detection result for pending image；It is for another example neural Network exports hand position testing result for pending image；For another example neural network is for the output live body inspection of pending image Survey result etc..The application does not limit image processing operations and the specific network structure that neural network specifically performs.

The training process of the neural network of the application may refer to following descriptions for Fig. 6.

Fig. 2 is the flow chart of image processing method one embodiment that can realize face location detection of the application.Such as Method shown in Fig. 2 includes：Step S200, step S210 and step S220.

S200, the histogram equalization processing based on gray scale is carried out to pending image.

In an optional example, in the feelings that pending image is gray level image (following to be referred to as pending gray level image) Under condition, the application directly can carry out histogram equalization processing, the histogram equalization processing to the pending gray level image Pending gray level image afterwards be provided in following step S210 it is pending after the equalization processing of neural network Image.

In an optional example, in situation of the pending image for RGB image (following to be known as pending RGB image) Under, pending RGB image first can be converted to YUV image by the application, then the Y channels pixel value of YUV image be carried out straight Square figure equalization processing, later, RGB image is converted to by the YUV image after equalization processing, which is To be provided to the pending image after the equalization processing of neural network in following step S210.

In an optional example, in situation of the pending image for RGB image (following to be known as pending RGB image) Under, pending RGB image first can be converted to gray level image by the application, then to transformed gray level image into column hisgram Equalization processing, the gray level image after histogram equalization processing are to be provided to neural network in following step S210 Pending image after equalization processing.

S210, the pending image after equalization processing is supplied to neural network, by the neural network for input Pending image zooming-out characteristics of image.

In an optional example, the pending image after equalization processing is by least one layer of convolutional layer and at least After the respective handling of one eLU layers of linear R, neural network extracts characteristics of image from the pending image of input.

S220, neural network determine each face location according to the characteristics of image that it is extracted, and export face location detection As a result.

In an optional example, the face location testing result of neural network output can include：At least one face Position (center of the external frame of such as face), the external frame size of at least one face and the face of external frame are in corresponding positions Put confidence level of appearance etc..The application can finally determine pending image according to the confidence level that face occurs in corresponding position In face location.

In an optional example, the pending image under half-light scene shown in Fig. 3 is passing through the Nogata based on gray scale After figure equalization processing, image shown in Fig. 4, comparison diagram 3 and Fig. 4 are formed it is found that pair of the pending image under half-light scene It is improved than degree and brightness, the features such as the exterior contour of face and detail textures are apparent from significantly, in Fig. 4 After shown image is input in neural network, neural network can export it and be directed to the face location of image shown in Fig. 4 Testing result, the external frame of face which can be embodied as in Fig. 3.It is external by the face in Fig. 3 Frame is it is found that the neural network of the application can accurately detect the face location in the image under half-light scene.

In an optional example, the pending image under complex illumination scene shown in fig. 5 is by based on gray scale After histogram equalization processing, after being input in neural network, neural network can export it for image shown in fig. 5 Face location testing result, the external frame of face which can be embodied as in Fig. 5.By in Fig. 5 The external frame of face is it is found that the neural network of the application can accurately detect the face position in the image under complex illumination scene It puts.

Fig. 6 is the flow chart of training neural network method one embodiment of the application.The embodiment side as shown in Figure 6 Method mainly includes：Step S600, step S610 and step S620.

S600, the image pattern concentrated according to training image obtain the input picture of neural network to be trained.

In an optional example, the image pattern that the training image of the application is concentrated was not carried out usually based on gray scale Equalization processing image pattern.The training image collection has generally comprised the preferable image pattern of quality, and (such as facial contour is clear Clear and detail textures clearly facial image sample), which generally also includes the image pattern of half-light scene, exposes The image pattern of the excessive image pattern of light, the image pattern of backlight scene and complex illumination scene.

In an optional example, the training image collection of the application can include gray level image sample and RGB image sample Etc. diversified forms image pattern.The image pattern that training image is concentrated usually has labeled data, the mark number of image pattern According to can be face location labeled data, people corresponding to face labeled data, face key point labeled data, human body it is crucial Point labeled data, position of human body labeled data, human action labeled data, gesture labeled data, hand position labeled data or work Body labeled data etc..

In an optional example, the application can concentrate from training image and obtain image pattern, and the figure that will be got The decent input picture directly as neural network to be trained, the image pattern directly as input picture be usually quality compared with Good image pattern.The application can also be based on the image pattern after acquisition image pattern is concentrated from training image The equalization processing of gray scale, and using the image pattern after equalization processing as the input picture of neural network to be trained.It needs The image pattern for carrying out the equalization processing based on gray scale is usually image pattern (the image sample as shown in Figure 3 of half-light scene Originally), image pattern (figure as shown in Figure 5 of the image pattern of overexposure, the image pattern of backlight scene and complex illumination Decent).

In an optional example, needing to carry out the equalization processing based on gray scale to the image pattern got, and In the case that the image pattern got is gray level image (following to be referred to as gray level image sample), the application can be directly to this Gray level image sample carries out histogram equalization processing, and using the gray level image sample after the histogram equalization processing as treating The input picture of training neural network.

In an optional example, needing to carry out the equalization processing based on gray scale to the image pattern got, and In the case that the image pattern got is RGB image (following to be known as RGB image sample), the application can be first by RGB image Sample is converted to YUV image, then, histogram equalization processing is carried out to the Y channels pixel value of YUV image, later, will be balanced Change that treated that YUV image is converted to RGB image, the application can be using transformed RGB image as neural network to be trained Input picture.

In an optional example, needing to carry out the equalization processing based on gray scale to the image pattern got, and In the case that the image pattern got is RGB image (following to be known as RGB image sample), the application can be first by RGB image Sample is converted to gray level image, and histogram equalization processing is then carried out to transformed gray level image, and the application can incite somebody to action straight Input picture of the gray level image as neural network to be trained after square figure equalization processing.

S610, input picture is supplied to neural network, and the handling result of the input picture is exported through neural network.

In an optional example, input picture is by least one layer of convolutional layer and at least eLU layers of a linear R After respective handling, neural network to be trained extracts characteristics of image from input picture, and neural network to be trained is extracted according to it The characteristics of image gone out carries out subsequent image processing operations, such as performs the operation that face location is determined according to characteristics of image, and defeated Go out the handling result of input picture, such as export the position (center of the external frame of such as face of the external frame of at least one face Deng), the faces position detection result such as the confidence level that occurs in corresponding position of the external frame size of at least one face and face.

S620, using training image concentrate image pattern labeled data as tutorial message, treat trained neural network into Row supervised learning, so that the handling result convergence of neural network to be trained output.

In an optional example, the application is tutorial message by using the corresponding labeled data of input picture, next pair Neural network to be trained exercises supervision study, and the handling result that can reduce neural network output to be trained is corresponding with input picture Labeled data between difference.Neural network to be trained output handling result labeled data corresponding with input picture it Between difference when meeting pre-provisioning request, neural network to be trained successfully trains completion.In addition, in the instruction for treating trained neural network During white silk, the application can also judge whether the training iterations of neural network to be trained have reached predetermined iterations And training image collection sample data concentrates whether also there is image pattern not being read etc., is judging nerve net to be trained The training iterations of network have reached predetermined iterations or training image concentrates the image pattern for being not present and not being read In the case of, it determines to meet predetermined convergence condition, this training process terminates.At the end of training process, if treating training god Difference between handling result labeled data corresponding with input picture through network output is unsatisfactory for pre-provisioning request, then waits to train This failure to train of neural network.

The application is treated by using the image pattern without the equalization processing based on gray scale that training image is concentrated Training neural network is trained, meanwhile, treat training nerve using the image pattern by the equalization processing based on gray scale Network is trained, and the neural network successfully trained can be made either to carry out image for contrast and the preferable image of brightness It handles or carries out image procossing for the image of contrast and luminance range, can have preferable image processing effect, So as to be conducive to improve the accuracy of the processing result image of neural network.

Fig. 7 is the structure diagram of the application image processing apparatus one embodiment.As shown in fig. 7, the figure of the embodiment As processing unit mainly includes：Equalization processing module 700 and acquisition processing result image module 710.Optionally, the implementation The image processing apparatus of example can also include：Judgment module 720 and determining grey level regions module 730.Optionally, the reality Applying the image processing apparatus of example can also include：The training device of neural network.The concrete structure of the training device of neural network Description may refer to following descriptions for Fig. 8, and this will not be repeated here.

Equalization processing module 700 is used to carry out pending image the equalization processing based on gray scale, in the application Equalization processing based on gray scale is used to increase the grey level region that the pixel value of the pending image is distributed.

In an optional example, equalization processing module 700 can be based on the whole region of pending image The equalization processing of gray scale, equalization processing module 700 can also carry out based on gray scale the subregion of pending image Equalization processing, for example, equalization processing module 700 is carried out for the region of other in addition to outer rim in pending image Equalization processing based on gray scale.

In an optional example, the mode that histogram equalization processing may be used in equalization processing module 700 is come pair Pending image carries out the equalization processing based on gray scale, and equalization processing module 700 can also be treated using other modes Processing image carry out the equalization processing based on gray scale, for example, equalization processing module 700 by directly adjust contrast and/ The mode of the indexs such as brightness to pending image carries out the equalization processing based on gray scale.

Equalization processing module 700 is carried out based on gray scale pending image by the way of histogram equalization processing An optional example of equalization processing be：It is gray level image (following to can be described as pending gray level image) in pending image In the case of, equalization processing module 700 directly can carry out histogram equalization processing to pending gray level image, and histogram is equal Weighing apparatusization treated gray level image is after the equalization processing in the application pending image.

Equalization processing module 700 is carried out based on gray scale pending image by the way of histogram equalization processing Another optional example of equalization processing be：It is RGB image (following to be known as pending RGB image) in pending image In the case of, first, pending RGB image is converted to YUV image by equalization processing module 700, then, equalization processing module The Y channels pixel value of 700 pairs of YUV images carries out histogram equalization processing, and later, equalization processing module 700 will equalize YUV image that treated is converted to RGB image, after above-mentioned transformed RGB image is the equalization processing in the application Pending image.Since Y channel tables show brightness, i.e. grayscale value, therefore, equalization processing module 700 is by YUV image Y channels pixel value carry out histogram equalization processing, the YUV image after equalization processing can be made to have in terms of gray scale The indexs such as the contrast being closer to normal picture and brightness, and RGB image is converted to by the YUV image after equalization processing Afterwards, the indexs such as contrast and brightness of the indexs such as the overall contrast of transformed RGB image and brightness and normal picture are more It is close.

It should be strongly noted that it is the non-gray level image of the other forms in addition to RGB image in pending image In the case of, equalization processing module 700, which equally may be used, to be first converted to YUV image by pending image and (is such as converted directly into YUV image is first converted to RGB image, then is converted to YUV image etc. by RGB image for another example), then, equalization processing module The Y channels pixel value of 700 pairs of YUV images carries out histogram equalization processing, and later, equalization processing module 700 again will be balanced Changing treated, YUV image be converted to the mode of the non-gray level image of corresponding form, it is pending after equalization processing to obtain Image.Equalization processing module 700 is carried out for the non-gray level image of the other forms in addition to RGB image based on the equal of gray scale The process of weighing apparatusization processing, no longer enumerates explanation herein.

Equalization processing module 700 is carried out based on gray scale pending image by the way of histogram equalization processing Another optional example of equalization processing be：It is RGB image (following to be known as pending RGB image) in pending image In the case of, first, pending RGB image is converted to gray level image by equalization processing module 700, then, equalization processing mould Block 700 carries out histogram equalization processing to transformed gray level image, and the gray level image after histogram equalization processing is Pending image after equalization processing in the application.

It should be strongly noted that it is the non-gray level image of the other forms in addition to RGB image in pending image In the case of, equalization processing module 700 equally may be used first is converted to gray level image, and to transformed by pending image Gray level image carries out the mode of histogram equalization processing, to obtain the pending image after equalization processing.Equalization processing Module 700 carries out the process of the equalization processing based on gray scale for the non-gray level image of the other forms in addition to RGB image, No longer illustrate one by one herein.

Processing result image module 710 is obtained to be used to the pending image after equalization processing being supplied to neural network, And the handling result of pending image is exported through neural network.

Judgment module 720 is used to judge whether the grey level region that the pixel value of pending image is distributed is less than first Threshold value, if so, the module execution of triggering equalization processing carries out the equalization processing based on gray scale to pending image.

In an optional example, the application is carried out based on gray scale pending image in equalization processing module 700 Before equalization processing, can grey level region that the pixel value of pending image is distributed first be judged by judgment module 720 Whether it is less than first threshold, if less than first threshold, then performs equalization processing module 700 and base is carried out to pending image In the operation of the equalization processing of gray scale；And if not less than first threshold, it can hold no longer equalization processing module 700 Row carries out pending image the operation of the equalization processing based on gray scale.

Grey level regions module 730 is determined for judging the corresponding pixel quantity of any pixel value in pending image Whether second threshold is reached, if so, corresponding pixel value to be included in the range in the grey level region, if it is not, will not then Corresponding pixel value is included in the range in the grey level region.

In an optional example, it is pending to determine that grey level regions module 730 can be determined by following manner The grey level region that the pixel value of image is distributed：It is any in pending image to determine that grey level regions module 730 judges Pixel value (for example, the gray value of pixel, for another example the numerical value of the R channels numerical value of pixel, G channels numerical value and channel B is flat Mean value etc.) whether corresponding pixel quantity reach second threshold, if reaching second threshold, it is determined that grey level region mould Above-mentioned pixel value is brought into the range in grey level region by block 730, and if not reaching second threshold, it is determined that gray level Above-mentioned pixel value will not be brought into the range in grey level region by other regions module 730.

Fig. 8 is the structure diagram of training device one embodiment of the application neural network.As shown in figure 8, the implementation The training device of the neural network of example mainly includes：It obtains input picture unit 800, obtain input picture handling result unit 810 and supervised learning unit 820.Each unit is illustrated respectively below.

It obtains image pattern of the input picture unit 800 for being concentrated according to training image and obtains neural network to be trained Input picture.

In an optional example, obtain the image pattern that training image used in input picture unit 800 is concentrated and lead to Chang Weiwei carried out the image pattern of the equalization processing based on gray scale.The training image collection has generally comprised quality and has preferably schemed Decent (such as facial contour is clear and detail textures clearly facial image sample), which generally also includes secretly The image pattern of light field scape, over-exposed image pattern, the image pattern of backlight scene and complex illumination scene image Sample.

In an optional example, gray scale can be included by obtaining training image concentration used in input picture unit 800 The image pattern of the diversified forms such as image pattern and RGB image sample.The image pattern that training image is concentrated usually has mark Data are noted, the labeled data of image pattern can be labeled data, the face of the people corresponding to face location labeled data, face Key point labeled data, human body key point labeled data, position of human body labeled data, human action labeled data, gesture mark Data, hand position labeled data or live body labeled data etc..

In an optional example, acquisition image pattern can be concentrated from training image by obtaining input picture unit 800, and By the image pattern got directly as the input picture of neural network to be trained.Directly as the image pattern of input picture The usually preferable image pattern of quality.Image sample can also be obtained being concentrated from training image by obtaining input picture unit 800 After this, the equalization processing based on gray scale is carried out to the image pattern, and using the image pattern after equalization processing as waiting to instruct Practice the input picture of neural network.Obtaining input picture unit 800 needs to carry out the image sample of the equalization processing based on gray scale This be usually the image pattern of half-light scene, the image pattern of overexposure, backlight scene image pattern and complex illumination Image pattern.

In an optional example, need to be based on the image pattern got obtaining input picture unit 800 The equalization processing of gray scale, and it is (following to be referred to as ash for gray level image to obtain the image pattern that input picture unit 800 is got Spend image pattern) in the case of, obtain input picture unit 800 directly can carry out histogram equalization to the gray level image sample Change is handled, and using the gray level image sample after the histogram equalization processing as the input picture of neural network to be trained.

In an optional example, need to be based on the image pattern got obtaining input picture unit 800 The equalization processing of gray scale, and it is (following to be known as RGB figures for RGB image to obtain the image pattern that input picture unit 800 is got Decent) in the case of, obtain input picture unit 800 first can be converted to YUV image by RGB image sample, then, obtain Input picture unit 800 carries out histogram equalization processing to the Y channels pixel value of YUV image, later, obtains input picture list YUV image after equalization processing is converted to RGB image by member 800, and obtaining input picture unit 800 can will be transformed Input picture of the RGB image as neural network to be trained.

In an optional example, need to be based on the image pattern got obtaining input picture unit 800 The equalization processing of gray scale, and it is (following to be known as RGB figures for RGB image to obtain the image pattern that input picture unit 800 is got Decent) in the case of, obtain input picture unit 800 first can be converted to gray level image by RGB image sample, then, obtain Input picture unit 800 is taken to carry out histogram equalization processing to transformed gray level image, obtaining input picture unit 800 can Using the input picture by the gray level image after histogram equalization processing as neural network to be trained.

Input picture handling result unit 810 is obtained to be mainly used for input picture being supplied to neural network to be trained, and The handling result of the input picture is exported through neural network to be trained.

In an optional example, obtain input picture handling result unit 810 and be supplied to the defeated of neural network to be trained Enter image by least one layer of convolutional layer and at least after the respective handling of eLU layers of a linear R, neural network to be trained from Characteristics of image is extracted in input picture, neural network to be trained is carried out according to the characteristics of image that it is extracted at subsequent image Reason operation, such as performs the operation that face location is determined according to characteristics of image, and export the handling result of input picture, so as to make to obtain Input picture handling result unit 810 is taken to get the handling result of input picture, such as obtains input picture handling result unit 810 get the position (center of the external frame of such as face of the external frame of at least one face of neural network output to be trained Deng), the faces position detection result such as the confidence level that occurs in corresponding position of the external frame size of at least one face and face.

The labeled data of image pattern that supervised learning unit 820 is used to concentrate using training image is treated as tutorial message Training neural network exercises supervision study, so that the handling result that neural network to be trained exports restrains.

In an optional example, supervised learning unit 820 is guidance by using the corresponding labeled data of input picture Information exercises supervision study to treat trained neural network, can reduce the handling result of neural network output to be trained with it is defeated Enter the difference between the corresponding labeled data of image.Supervised learning unit 820 is at the place for determining neural network output to be trained When difference between corresponding with the input picture labeled data of reason result meets pre-provisioning request, then it is assumed that neural network to be trained into Work(training is completed.In addition, in the training process for treating trained neural network, supervised learning unit 820 can also judge to wait to instruct Whether the training iterations of white silk neural network have reached predetermined iterations and training image collection sample data concentration It is no also to there is image pattern not being read etc., it is had reached in the training iterations for judging neural network to be trained predetermined In the case that iterations or training image are concentrated there is no the image pattern not being read, supervised learning unit 820 determines Meet predetermined convergence condition, this training process terminates.At the end of training process, if the place of neural network to be trained output Difference between reason result labeled data corresponding with input picture is unsatisfactory for pre-provisioning request, then supervised learning unit 820 is thought This failure to train of neural network to be trained.

The training device of neural network by using training image concentrate without the equalization processing based on gray scale Image pattern is treated trained neural network and is trained, meanwhile, utilize the image pattern by the equalization processing based on gray scale It treats trained neural network to be trained, the neural network successfully trained can be made either preferable for contrast and brightness Image carries out image procossing and still carries out image procossing for the image of contrast and luminance range, can have preferable Image processing effect, so as to be conducive to improve the accuracy of the processing result image of neural network.

Example devices

Fig. 9 shows the example devices 900 for being adapted for carrying out the application, and equipment 900 can be the control being configured in automobile System/electronic system, mobile terminal (for example, intelligent mobile phone etc.), personal computer (PC, for example, desktop computer or Notebook computer etc.), tablet computer and server etc..In Fig. 9, equipment 900 includes one or more processor, communication Portion etc., one or more of processors can be：One or more central processing unit (CPU) 901 and/or, one Or multiple images processor (GPU) 913 etc., processor can be executable in read-only memory (ROM) 902 according to being stored in Instruction or performed from the executable instruction that storage section 908 is loaded into random access storage device (RAM) 903 it is various appropriate Action and processing.Communication unit 912 can include but is not limited to network interface card, and the network interface card can include but is not limited to IB (Infiniband) network interface card.Processor can communicate to perform with read-only memory 902 and/or random access storage device 930 Executable instruction is connected with communication unit 912 by bus 904 and communicated by communication unit 912 with other target devices, so as to Complete the corresponding steps in the application.

In an optional example, processor is mainly included by performing the step of executable instruction is completed：It treats Processing image carries out the equalization processing based on gray scale, and the equalization processing based on gray scale is used to increase the pending figure The grey level region that the pixel value of picture is distributed；Pending image after equalization processing is supplied to neural network, and pass through The neural network exports the handling result of the pending image.In an optional example, processor can by performing The step of execute instruction is completed mainly includes：The defeated of neural network to be trained is obtained according to the image pattern that training image is concentrated Enter image；The input picture is supplied to neural network to be trained, and export the input picture through neural network to be trained Handling result；The labeled data for the image pattern concentrated using the training image treats trained neural network as tutorial message Exercise supervision study, so that the handling result convergence of neural network to be trained output.

Processor may refer to phase in above method embodiment by performing the specific steps that executable instruction is completed Description is closed, is no longer described in detail herein.In addition, in RAM 903, can also be stored with the various programs needed for device operation with And data.CPU901, ROM902 and RAM903 are connected with each other by bus 904.In the case where there is RAM903, ROM902 is Optional module.RAM903 stores executable instruction or executable instruction is written into ROM902 at runtime, and executable instruction makes Central processing unit 901 performs the step included by above-mentioned method for segmenting objects.Input/output (I/O) interface 905 is also connected to Bus 904.Communication unit 912 can be integrally disposed, may be set to be with multiple submodule (for example, multiple IB network interface cards), and It is connect respectively with bus.

I/O interfaces 905 are connected to lower component：Importation 906 including keyboard, mouse etc.；It is penetrated including such as cathode The output par, c 907 of spool (CRT), liquid crystal display (LCD) etc. and loud speaker etc.；Storage section 908 including hard disk etc.； And the communications portion 909 of the network interface card including LAN card, modem etc..Communications portion 909 via such as because The network of spy's net performs communication process.Driver 910 is also according to needing to be connected to I/O interfaces 905.Detachable media 911, such as Disk, CD, magneto-optic disk, semiconductor memory etc. are mounted on driver 910, as needed in order to be read from thereon Computer program be installed in as needed in storage section 908.

It should be strongly noted that framework as shown in Figure 9 is only a kind of optional realization method, in concrete practice process In, can the component count amount and type of above-mentioned Fig. 9 be selected, be deleted, be increased or be replaced according to actual needs；In different function Component setting on, can also be used it is separately positioned or integrally disposed and other implementations, for example, GPU and CPU separate setting, for another example GPU, can be integrated on CPU, communication unit separates setting, also can be integrally disposed in CPU or GPU is first-class by reason.These are replaceable Embodiment each fall within the protection domain of the application.

Particularly, it according to presently filed embodiment, may be implemented as calculating below with reference to the process of flow chart description Machine software program, for example, the application embodiment includes a kind of computer program product, it can it includes machine is tangibly embodied in The computer program on medium is read, computer program was included for the program code of the step shown in execution flow chart, program generation Code can include the instruction corresponding to each step in the corresponding method for performing the application and providing, for example, for pending figure Instruction as carrying out the equalization processing based on gray scale, the equalization processing based on gray scale are used to increase the pending image The grey level region that pixel value is distributed；And for the pending image after equalization processing to be supplied to neural network, And the instruction of the handling result of the pending image is exported through neural network.

In such embodiment, which can be downloaded and pacified from network by communications portion 909 It fills and/or is mounted from detachable media 911.When the computer program is performed by central processing unit (CPU) 901, perform Realize the above-metioned instruction of each step of method in the application.

The present processes and device, electronic equipment and computer-readable storage medium may be achieved in many ways Matter.For example, can by any combinations of software, hardware, firmware or software, hardware, firmware come realize the present processes and Device, electronic equipment and computer readable storage medium.The said sequence of the step of for method merely to illustrate, The step of the present processes, is not limited to sequence described in detail above, unless specifically stated otherwise.In addition, at some In embodiment, the application can be also embodied as recording program in the recording medium, these programs include being used to implement basis The machine readable instructions of the present processes.Thus, the application also covers storage for performing the journey according to the present processes The recording medium of sequence.The description of the present application in order to example and description for the sake of and provide, and be not exhaustively or will The application is limited to disclosed form.Many modifications and variations are obvious for the ordinary skill in the art.Choosing It is the principle and practical application in order to more preferably illustrate the application to select and describe embodiment, and makes the ordinary skill people of this field Member it will be appreciated that the application so as to design the various embodiments with various modifications suitable for special-purpose.

Claims

1. a kind of image processing method, which is characterized in that the method includes：

The equalization processing based on gray scale is carried out to pending image, the equalization processing based on gray scale is described for increasing The grey level region that the pixel value of pending image is distributed；

Pending image after equalization processing is supplied to neural network, and export the pending figure through the neural network The handling result of picture.

2. according to the method described in claim 1, it is characterized in that, described carry out the equalization based on gray scale to pending image Before processing, including：

Judge whether the grey level region that the pixel value of pending image is distributed is less than first threshold, if so, performing institute State the step of equalization processing based on gray scale is carried out to pending image.

3. the according to the method described in claim 2, it is characterized in that, ash that the pixel value for judging pending image is distributed Spend whether rank region is less than before first threshold, including：

Judge whether the corresponding pixel quantity of any pixel value reaches second threshold in pending image, if so, will be corresponding Pixel value be included in the range in the grey level region, if it is not, then corresponding pixel value is not included in the grey level area The range in domain.

4. according to the method in any one of claims 1 to 3, which is characterized in that the pending image includes：Gray-scale map Picture or non-gray level image.

5. according to the method described in claim 4, it is characterized in that, described carry out the equalization based on gray scale to pending image Processing includes：

To carrying out histogram equalization processing by the gray level image as pending image.

6. a kind of training method of neural network, which is characterized in that the method includes：

The image pattern concentrated according to training image obtains the input picture of neural network to be trained；

The input picture is supplied to neural network to be trained, and the place of the input picture is exported through neural network to be trained Manage result；

The labeled data for the image pattern concentrated using the training image is treated trained neural network and exercised supervision as tutorial message Study, so that the handling result convergence of neural network to be trained output.

7. a kind of image processing apparatus, which is characterized in that including：

Equalization processing module, it is described based on the equal of gray scale for carrying out the equalization processing based on gray scale to pending image Weighing apparatusization handles to increase the grey level region that the pixel value of the pending image is distributed；

Processing result image module is obtained, for the pending image after equalization processing to be supplied to neural network, and through institute State the handling result that neural network exports the pending image.

8. a kind of training device of neural network, which is characterized in that described device includes：

Input picture unit is obtained, for obtaining the input figure of neural network to be trained according to the image pattern that training image is concentrated Picture；

Input picture handling result unit is obtained, for the input picture to be supplied to neural network to be trained, and through waiting to instruct Practice the handling result that neural network exports the input picture；

Supervised learning unit, the labeled data of image pattern for being concentrated using the training image treat instruction as tutorial message Practice neural network to exercise supervision study, so that the handling result that neural network to be trained exports restrains.

9. a kind of electronic equipment, including：

Memory, for storing computer program；

Processor, for performing the computer program stored in the memory, and the computer program is performed, and is realized Method described in any one of the claims 1-6.

10. a kind of computer readable storage medium, is stored thereon with computer program, which is executed by processor When, realize the method described in any one of the claims 1-6.