CN109165645A - A kind of image processing method, device and relevant device - Google Patents

Abstract

The embodiment of the invention discloses a kind of image processing method, device and relevant devices, method includes: to obtain the target image comprising target object and references object, the Pixel Dimensions of detected target object in the target image, as object pixel size, and the Pixel Dimensions of references object in the target image are detected, as reference pixel size；The reference actual size for obtaining references object according to object pixel size, reference pixel size and refers to actual size, determines the target actual size of target object.Using the present invention, the size of business object is automatically determined, improve the efficiency of measurement target object size.

Description

A kind of image processing method, device and relevant device

Technical field

The present invention relates to field of computer technology more particularly to a kind of image processing methods, device and relevant device.

Background technique

In agricultural insurance, if peasant household is the poultry of cultivation or is that the crops planted are insured, and are sent out by object is protected When raw natural calamity or accident, it is necessary to initiate Claims Resolution request to insurance institution to obtain insurance indemnity.In order to get agriculture The loss amount at family and determine amount for which loss settled, the insurance business person of insurance institution needs to the in-site measurement of accident occurs by guarantor couple The volume size of elephant, and then indemnity is determined according to the volume size measured.But it is protected by business personnel to in-site measurement The volume of object can expend a large amount of human cost, and can spend business personnel's a large amount of time.

Summary of the invention

The embodiment of the present invention provides a kind of image processing method, device and relevant device, can automatically determine industry The size of business object, improves the efficiency of measurement business object size.

On the one hand the embodiment of the present invention provides a kind of image processing method, comprising:

The target image comprising target object and references object is obtained, detects the target object in the target image Pixel Dimensions, as object pixel size, and detect Pixel Dimensions of the references object in the target image, as Reference pixel size；

The reference actual size for obtaining the references object, according to the object pixel size, the reference pixel size Actual size is referred to described, determines the target actual size of the target object.

On the one hand the embodiment of the present invention provides a kind of image processing apparatus, comprising:

Module is obtained, for obtaining the target image comprising the target object and references object；

First detection module, for detecting Pixel Dimensions of the target object in the target image, as target Pixel Dimensions；

Second detection module, for detecting Pixel Dimensions of the references object in the target image, as reference Pixel Dimensions；

Determining module, for obtaining the reference actual size of the references object, according to the object pixel size, described Reference pixel size and it is described refer to actual size, determine the target actual size of the target object.

Wherein, the first detection module, comprising:

First convolution unit, for based on the convolutional layer in the first complete convolutional neural networks model, to the target figure As carrying out process of convolution, the first convolution characteristic pattern being composed of the first convolution characteristic information is obtained；

Search unit, for searching for multiple first interest region in the first convolution characteristic pattern；

First convolution unit is also used to carry out the first convolution characteristic information for including in each first interest region Pondization processing, obtains first structure characteristic information, identify the first structure characteristic information that includes in each first interest region with First matching degree of multiple attribute type features in the first complete convolutional neural networks model；

First convolution unit is also used in corresponding multiple first matching degrees in each first interest region, will most Big first matching degree is as the corresponding confidence level in the first interest region；

First convolution unit is also used in the corresponding confidence level in multiple first interest region, by maximum confidence Corresponding first interest region is as first object region, and using the size in the first object region as the object pixel Size.

Wherein, described search unit, comprising:

Score computation subunit, for the sliding window on the first convolution characteristic pattern, in each window based on not Multiple candidate regions are determined with the anchor point frame under scale；

The score computation subunit is also used to calculate separately the target of the first convolution characteristic information in each candidate region Prospect score；

Score determines subelement, for the target prospect score according to each candidate region, determines the multiple One interest region.

Wherein, the score determines subelement, comprising:

First determines subelement, and the candidate region for the target prospect score to be greater than score threshold is determined as first Auxiliary area；

Described first determines subelement, is also used in first auxiliary area, will have maximum target prospect score The first auxiliary area be determined as the second auxiliary area；

It deletes and retains subelement, for calculating separately between second auxiliary area and remaining first auxiliary area The first auxiliary area that overlapping area is greater than area threshold is deleted, and retains overlapping area and be less than or equal to by overlapping area First auxiliary area of the area threshold；

Described first determines subelement, is also used to for second auxiliary area and the first auxiliary area retained being determined as First interest region.

Wherein, second detection module, comprising:

Converting unit, for the target image to be converted to target gray image；

Area determination unit, for using the connection region in the target gray image as the first reference zone；

The converting unit is also used to respectively input the gray level image in multiple first reference zones in disaggregated model, Identify the first probability in each first reference zone comprising the references object；

Selecting unit, for being chosen in the multiple first reference zone and meeting matching item according to first probability First reference zone of part, as the second reference zone；

Size determination unit determines the reference pixel size for the size according to second reference zone.

Wherein, the area determination unit, comprising:

Gradient computation subunit, for calculating the corresponding first gradient figure of the target gray image according to gradient operator, And closed operation is carried out to the first gradient figure, obtain the second gradient map；

First detection sub-unit, for detecting the connection region of the first gradient figure and second gradient map respectively, And the connection region that will test out is become a full member processing, and multiple auxiliary reference regions are obtained；

First detection sub-unit is also used to have and the auxiliary reference region in the target gray image The region of same position information, as first reference zone.

Wherein, the selecting unit, comprising:

First extracts subelement, for extracting maximum first probability in multiple first probability；

Second determines subelement, if being less than or equal to probability threshold value for maximum first probability, according to described the The size of the Aspect Ratio of location information, first reference zone where one reference zone, first reference zone, really The weight of fixed first reference zone, is determined as maximum first reference zone of weight to meet the first of the matching condition Reference zone, and maximum first reference zone of the weight is determined as second reference zone；

Described second determines subelement, will be described if being also used to maximum first probability is greater than the probability threshold value Corresponding first reference zone of the first probability of maximum is determined as meeting the first reference zone of the matching condition, and by described in most Corresponding first reference zone of big first probability is determined as second reference zone.

Wherein, the size determination unit, comprising:

Second extracts subelement, for extracting institute when the references object belongs to the first fixed references object of shape The gray-scale Image Edge information in the second reference zone is stated, auxiliary gradient image is obtained；

Second detection sub-unit, for detecting the full curve in the auxiliary gradient image, as first object curve, First object diameter is determined according to the first object curve；

Second detection sub-unit is also used to the first object diameter being determined as the reference pixel size.

Wherein, the size determination unit, further includes:

Subelement is clustered, it, will be described for when the references object belongs to unfixed second references object of shape In target image, there is the region with the second reference zone same position information, as third reference zone；

The cluster subelement is also used to the color according to the image in the third reference zone, joins to the third Image in the domain of examination district carries out color cluster processing, obtains cluster result region；

Size determines subelement, for the size according to the cluster result region, determines the reference pixel size.

Wherein, second detection module, further includes:

Second convolution unit, for when the references object belongs to unfixed second references object of shape, based on covering Convolutional layer in the convolutional neural networks model of code region, carries out process of convolution to the target image, obtains by the second convolution spy Reference ceases the second convolution characteristic pattern being composed；

Recognition unit, it is emerging to each second for searching for multiple second interest region in the second convolution characteristic pattern The the second convolution characteristic information for including in interesting region carries out pond processing, obtains the second structure feature information；

The recognition unit is also used to according to the second feature information for including in each second interest region, described in identification It include the second probability of second references object in each second interest region；

Computing unit for maximum second probability corresponding second interest region to be determined as auxiliary mark region, and is counted Calculate the binary cover of each pixel in the auxiliary mark region；

The computing unit is also used to belong to all pixels group corresponding to the binary cover of foreground mask and is combined into target Subgraph, and the reference pixel size is sized to according to the target subgraph.

Wherein, second detection module, further includes:

Third convolution unit, for being based on second when the references object belongs to the first fixed references object of shape Convolutional layer in complete convolutional neural networks model carries out process of convolution to the target image, obtains by third convolution feature The third convolution characteristic pattern that information is composed；

The third convolution unit is also used to search for multiple third interest region in the third convolution characteristic pattern, right The third convolution characteristic information for including in each third interest region carries out pond processing, obtains third structure feature information；

The third convolution unit is also used to according to the third feature information for including in each third interest region, identification It include the third probability of first references object in each third interest region；

The third convolution unit is also used in the corresponding third probability in multiple third interest region, will have maximum The third interest region of third probability is determined as the second target area；

Extraction unit obtains edge gradient image for extracting the marginal information of the image in second target area；

The third convolution unit is also used to detect the full curve in the edge gradient image, as the second target Curve determines the second aimed dia according to second aim curve, and second aimed dia is determined as the reference Pixel Dimensions.

Wherein, the acquisition module, is specifically used for:

Service request associated with the target object is received, obtaining according to the service request includes the target pair As the target image with the references object.

Then described image processing unit further include:

Display module, for by the attribute classification of the corresponding attribute type feature of the confidence level in the first object region, As the first object region corresponding label information, according to the corresponding label information in the first object region and described Target actual size determines target service data associated with the service request；

The display module is also used to show the target service data, and stores the target service data；The mesh Mark business datum includes: the sign information of business Claims Resolution amount and the target object；

Sending module, for the target service data to be sent to service terminal associated with the service request.

On the one hand the embodiment of the present invention provides a kind of electronic equipment, comprising: processor and memory；

The processor is connected with memory, wherein for storing program code, the processor is used for the memory Said program code is called, to execute such as the method in the embodiment of the present invention.

On the one hand the embodiment of the present invention provides a kind of computer storage medium, the computer storage medium is stored with meter Calculation machine program, the computer program include program instruction, and described program is instructed when being executed by a processor, executed such as the present invention Method in embodiment.

The embodiment of the present invention is by obtaining the target image comprising target object and references object, and detected target object is in mesh Pixel Dimensions in logo image as object pixel size, and detect the Pixel Dimensions of references object in the target image, as Reference pixel size；The reference actual size for obtaining references object, according to object pixel size, reference pixel size and with reference to real Border size determines the target actual size of target object.It is above-mentioned it is found that from the image comprising target object and references object Detected target object Pixel Dimensions in the picture and references object Pixel Dimensions in the picture respectively, then obtain reference pair The full-size(d) of elephant can determine the full-size(d) of target object under a proportional relationship, so as to automatically determine target The full-size(d) of object avoids the size for measuring business object in a manual manner, improves the efficiency of measurement business object size.

Detailed description of the invention

In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with It obtains other drawings based on these drawings.

Fig. 1 a is a kind of system architecture diagram of image processing method provided in an embodiment of the present invention；

Fig. 1 b- Fig. 1 d is a kind of schematic diagram of a scenario of image processing method provided in an embodiment of the present invention；

Fig. 2 is a kind of flow diagram of image processing method provided in an embodiment of the present invention；

A kind of Fig. 3 flow diagram for detecting reference pixel size provided in an embodiment of the present invention；

Fig. 4 is a kind of flow diagram of image processing method provided in an embodiment of the present invention；

Fig. 5 is a kind of structural schematic diagram of image processing apparatus provided in an embodiment of the present invention；

Fig. 6 is the structural schematic diagram of a kind of electronic equipment provided in an embodiment of the present invention.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other Embodiment shall fall within the protection scope of the present invention.

A referring to Figure 1 is a kind of system architecture diagram of image processing method provided in an embodiment of the present invention, server 10a There is provided service for user terminal cluster, user terminal cluster may include: user terminal 10b, user terminal 10c ..., user Terminal 10d.It is following to request to be illustrated for scene so that in agricultural insurance, peasant household initiates Claims Resolution to insurance institution, when user (can To be user 10e, 10f or 10g) when needing to initiate Claims Resolution request to insurance institution, user can (can be with based on user terminal It is user terminal 10b, user terminal 10c or user terminal 10d) picture of the shooting comprising target object and references object, it uses After family terminal gets picture, it is sent to server 10a.Server 10a detects target object and reference pair in picture respectively The Pixel Dimensions of elephant, and according to the actual size and proportionate relationship of references object, determine the full-size(d) of target object, into And sign information associated with target object and Claims Resolution amount are determined according to the full-size(d) of target object.Server 10a It by determining sign information and Claims Resolution amount associated storage to database, and is sent to the user terminal, so that user knows this The sign information and amount for which loss settled obtained of the target object of secondary Claims Resolution request.Certainly, if user terminal itself can be with Target object and the Pixel Dimensions of references object in picture are detected, target object can also be directly determined in the user terminal Sign information and Claims Resolution amount need to only be sent to server 10a for storing determining sign information and Claims Resolution amount.Under Fig. 1 b- Fig. 1 d is stated by taking a user terminal 10b and server 10a as an example, illustrates how to determine target object and references object Pixel Dimensions and target object full-size(d).

Wherein, user terminal may include mobile phone, tablet computer, laptop, palm PC, intelligent sound, movement Internet device (MID, mobile internet device), wearable is set POS (Point Of Sales, point of sale) machine Standby (such as smartwatch, Intelligent bracelet etc.) etc..

B referring to Figure 1 is a kind of schematic diagram of a scenario of image processing method provided in an embodiment of the present invention.When target pair When accident occurs as (can be poultry, livestock etc.), peasant household initiates Claims Resolution request to insurance company.First in user terminal 10b Interface 20a in, fill in essential information of reporting a case to the security authorities, such as odd numbers when purchase insurance, the reason of accident and time occurs.At interface In 20b, shooting includes the photo of target object, in order to accurately calculate the full-size(d) of target object, before shooting photo, also A references object need to be placed on the upper surface of the side of target object or target object, references object can be identity card, public affairs Hand over the standard card or unitary coin of card equidimension standard.After the completion of preparation, one can be shot and contain target pair As the picture with references object.Target object is in the center of picture in picture in order to make shooting, facilitates subsequent detection target The size of object, when shooting photo, can prompt user to be directed at target object and references object as shown in the 20c of interface Preset rectangle frame shooting.

As illustrated in figure 1 c, when user terminal 10b gets the picture 20d of shooting, picture 20d is sent to server 10a, stores the identification model 20e that can detecte the Pixel Dimensions of target object in picture in server 10a, and can be with Detect the identification model 20g of the Pixel Dimensions of references object in picture.Picture 20d is inputted in identification model 20e, based on identification Convolutional layer in model 20e executes convolution algorithm to picture 20d, and available convolution characteristic pattern includes in the convolution characteristic pattern Be target object in picture 20d convolution characteristic information.It is, in general, that the size of obtained convolution characteristic pattern is to be less than figure The boundary of each convolution characteristic pattern can be filled upper numerical value for the size of unified convolution characteristic pattern by the size of piece 20d " 0 ", so that the size of convolution characteristic pattern is consistent with the size of picture 20d, and the numerical value " 0 " filled will not influence subsequent calculating As a result.With according to preset step-length (step-length can be 1) sliding window (window size can be 3 × 3) on convolution characteristic pattern, In each window, multiple candidates are determined according to preset anchor point frame (anchor point frame size can be 16 × 16) and multiple scales Frame, the multiple dimensioned length-width ratio that can be adjustment anchor point frame is respectively 0.5,1,2, and length-width ratio anchor point frame adjusted is distinguished 0.5 times, 1 times, 2 times of scaling.It is recognised that multiple candidate frames can be corresponded in each sliding window.According to each candidate Convolution characteristic information included in frame calculates the prospect score of each candidate frame, the corresponding candidate frame of the higher explanation of prospect score In comprising target object probability it is higher.Prospect score is greater than the candidate frame of default score threshold as the first interest region. Due to that can have a large amount of overlapping area between multiple first interest region, can by those corresponding with highest prospect score The first interest region between one interest region with a large amount of overlapping areas is deleted, and is only retained corresponding with highest prospect score The first interest region between there is a small amount of overlapping area or not no the first interest region of overlapping area.It has been determined multiple Behind one interest region, the subsequent convolution characteristic information in each first interest region carries out pondization processing and full junction Reason, the characteristic information that connects that treated entirely based on the classifier calculated in identification model 20e with it is multiple included in classifier The matching degree of attribute type feature, matching degree is higher, illustrates the attribute classification of object included in corresponding first interest region Be corresponded in classifier attribute type feature the other probability of Attribute class it is higher.Each first interest region can obtain more A matching degree, by using maximum matching degree as the confidence level in the first interest region, and by the corresponding attribute of maximum matching degree Label information of the classification as the first interest region is the corresponding label information in a first interest region, and corresponding One confidence level.Wherein, label information is attribute classification possessed by object in the first interest region, for example, attribute classification can To be the livestocks such as pig, sheep, ox.In the corresponding confidence level in multiple first interest region, choose corresponding to maximum confidence The first interest region as first object region 20f, and the first object region 20f carry label information " pig ", the label Information " pig " is exactly the attribute classification of target object.It is assured that target object is being schemed by the size of first object region 20f Pixel Dimensions in piece 20d, such as can Pixel Dimensions using the length of target area 20f as target object in picture 20d. Above-mentioned identification model 20e is based on algorithm of target detection RFCN (Region-based Fully Convolutional Networks, the complete convolutional network based on region) training, algorithm of target detection can also include: RCNN (Regions with CNN, region convolutional neural networks), FAST RCNN (fast region convolutional neural networks), FASTER RCNN (faster region convolutional neural networks), above-mentioned algorithm of target detection may be implemented in picture 20d and identify One target area, and identify the attribute classification of target object in first object region.

As illustrated in figure 1 c, picture 20d is inputted in identification model 20g, based on the convolutional layer in identification model 20g, to figure Piece 20d executes convolution algorithm, and available convolution characteristic pattern, include in the convolution characteristic pattern is references object in picture 20d Convolution characteristic information.Equally, in the Boundary filling numerical value " 0 " of convolution characteristic pattern, so that the size and picture of convolution characteristic pattern The size of 20d is consistent, and the numerical value " 0 " filled will not influence subsequent calculated result.It is searched in above-mentioned convolution characteristic image Second interest region, the detailed process in the second interest region of search can join the correlation step in above-mentioned the first interest of search region Suddenly.Pondization processing and full connection processing, classifier calculated are carried out respectively to the convolution characteristic information in each second interest region The other matching probability of Attribute class of full connection treated characteristic information and references object (is to calculate each second interest region In include references object probability), using maximum matching probability corresponding second interest region as target auxiliary area.Based on knowledge Recurrence mask layer in other model 20g calculates the binary cover of each pixel in target auxiliary area.Binary cover includes prospect Mask and background mask, foreground mask indicate that corresponding pixel belongs to references object, and background mask indicates that corresponding pixel belongs to Background.Using the combination of pixels for belonging to foreground mask as target subgraph 20h, it is assured that according to the size of target subgraph 20h Pixel Dimensions of the references object in picture 20d, such as the diagonal line length of target subgraph 20h can schemed as references object Pixel Dimensions in piece 20d.Above-mentioned identification model 20g is based on splitting object algorithm Mask-RCNN (Mask Region- Based Fully Convolutional Networks, the complete convolutional network based on mask region) training, Mask-RCNN can accurately obtain the location and shape in picture of references object, for measuring the reference of small volume Object when accurate Pixel Dimensions can guarantee subsequent determining indemnity, has smaller error.Wherein it is determined that target is sub The sequencing of Figure 20 h and determining first object region 20f do not limit, and it is to pass through identification that target object, which is attribute classification, What model 20e was determined.

As shown in Figure 1 d, server 10a is according to Pixel Dimensions of target object " pig " in picture 20d, the picture of references object Plain size, the actual size of references object can determine the full-size(d) (including true body is long and true body is wide) of target object, Such as the true body of shown pig body is long in the 30a of interface are as follows: 80cm, according to true body is long and the Attribute class of target object Not Que Ding target object weight: 70kg, and then this reason is determined according to the weight of target object, the attribute classification of target object Pay for the indemnity of request are as follows: 900 yuan.The body length of target object, weight, indemnity can be sent to use by server 10a Family terminal 10b, the display data received from server 10a in the interface 30b of user terminal.By in detection image The Pixel Dimensions of target object and references object, and then determine the full-size(d) of target object.It can automatically determine target The full-size(d) of object avoids measuring size in a manual manner, improves the efficiency of measurement target object size.

Wherein, the detailed process of target object and the Pixel Dimensions of references object may refer to the following figure in detection image Embodiment corresponding to 2 to Fig. 4.

Further, Fig. 2 is referred to, is a kind of flow diagram of image processing method provided in an embodiment of the present invention. As shown in Fig. 2, described image processing method may include:

Step S101 obtains the target image comprising the target object and references object, detects the target object and exist Pixel Dimensions in the target image, as object pixel size.

It is asked specifically, terminal device (can be the server in above-mentioned Fig. 1 d) receives business associated with target object (such as settlement of insurance claim request in above-mentioned Fig. 1 a) is asked, obtaining target image according to service request (can be in above-mentioned Fig. 1 c Picture 20d), it include that target object (can be the pig body in above-mentioned Fig. 1 c) and references object (can be in the target image Card in above-mentioned Fig. 1 c), wherein the object for full-size(d) to be calculated is known as target object, for making ratio with target object Compared with object be known as references object.Based on the volume in the first complete convolutional neural networks model for constructing completion in terminal device Lamination, terminal device carry out process of convolution to target image, that is, the sub-fraction characteristic information randomly selected in target image is made For sample (convolution kernel), successively slip over all target images using this sample as a window, that is, above-mentioned sample and Target image does convolution algorithm, to obtain the convolution characteristic information about target object in target image.After process of convolution, The available convolution characteristic pattern of terminal device, include in the convolution characteristic pattern is the convolution feature of target object in target image The convolution characteristic information of target object is known as the first convolution characteristic information by information, terminal device, and will be by the first convolution feature The convolution characteristic image that information is composed is known as the first convolution characteristic pattern.Based on RPN (Region Proposal Network, Choose network in region) algorithm, multiple interest regions, referred to as the first interest region are searched in the first convolution characteristic pattern.Below with It is illustrated for one the first interest region, is counted in a like fashion if there is multiple first interest region terminal device that can adopt Calculate the label information of object in the confidence level and each first interest in each first interest region.To the first interest region Zhong Bao The first convolution characteristic information for containing carries out pond processing, is aggregate statistics the first convolution characteristic information, can be with after aggregate statistics Obtain the static structure characteristic information about target object in target image, referred to as first structure characteristic information.Pondization is handled In order to reduce subsequent calculation amount and shared weight.Based on the classifier in the first complete convolutional neural networks model, identification It is multiple in the complete convolutional neural networks model of first structure characteristic information and first in first interest region after aggregate statistics The matching degree of attribute type feature, referred to as the first matching degree, that is, each first interest region will correspond to multiple first With degree, and each first interest region corresponds to the attribute in the quantity and the first complete convolutional neural networks model of the first matching degree The quantity of type feature is identical, wherein each attribute type feature corresponds to an attribute classification.In each first interest region pair In multiple first matching degrees answered, terminal device is using maximum first matching degree as the confidence level in the first interest region, and terminal Equipment is using the attribute classification of the corresponding attribute type feature of maximum first matching degree as the mark of object in the first interest region Sign information.For example, including attribute classification " pig " corresponding attribute type feature in the first complete convolutional neural networks model；Attribute class Not " sheep " corresponding attribute type feature；And the corresponding attribute type feature of attribute classification " ox ", classifier identify that first structure is special The first matching degree that reference ceases between A attribute type feature corresponding with above-mentioned 3 attribute classifications is available: first structure is special It is 0.1 that reference, which ceases A and first matching degree of attribute classification " pig ",；The first of first structure characteristic information A and attribute classification " sheep " Matching degree is 0.7；The first matching degree of first structure characteristic information A and attribute classification " ox " is 0.9.Therefore first structure feature The confidence level in information A corresponding first interest region be 0.9, and in above-mentioned first interest region object label information are as follows: ox.

Using aforesaid way, terminal device can be respectively that each first interest region determines corresponding confidence level and mark Information is signed, next maximum confidence corresponding first interest region can be determined as to first object region (such as above-mentioned figure First object region 20f in 1c), and using the Pixel Dimensions of first object region in the target image as the ruler of object pixel It is very little, such as by the pixel length in first object region or using the elemental area in first object region as object pixel size.On The first complete convolutional neural networks model is stated based on algorithm of target detection RFCN training and come.

Based on RPN algorithm, searched in the first convolution characteristic pattern multiple first interest region detailed process may is that it is pre- Window size (such as window size is 3*3) is set, and according to preset step-length, (step-length can on the first convolution characteristic pattern for terminal device Think 1) successively sliding window, in each window, according to preset anchor point frame (anchor point frame size can be 16 × 16) with And multiple scales determine multiple candidate frames, the multiple dimensioned length-width ratio that can be adjustment anchor point frame or the size for scaling anchor point frame. It is recognised that multiple candidate frames can be corresponded in each sliding window.Terminal device calculates the first volume in each candidate region The target prospect score of product spy's information, is the probability by belonging to foreground area in the corresponding candidate region of classifier calculated, if Target prospect score is higher, illustrate the candidate region belong to foreground area probability it is higher.Since every target image all can Corresponding a large amount of candidate region (quantity can exceed that 2000), and these candidate regions are there are a large amount of overlapping region, because This can reject the candidate regions of a part overlapping using NMS (non maximum suppression, non-maxima suppression) algorithm Domain.The detailed process of NMS may is that candidate region of the terminal device by target prospect score greater than preset score threshold first As the first auxiliary area, in the corresponding target prospect score of all first auxiliary areas, selection has maximum target prospect First auxiliary area of score is determined as the second auxiliary area.Using formula (1) (calculating IOU) or formula (2) can be used (calculating overlap) calculates separately the overlapping area between the second auxiliary area and each first auxiliary area,

Wherein, A indicates that the area of the second auxiliary area, B indicate that the area of the first auxiliary area, IOU are equal to the first auxiliary The intersection of area and the second secondary surface divided by the first service floor area and the second service floor area union；It is auxiliary that overlap is equal to first The long-pending intersection with the second secondary surface of principal surface is divided by the second service floor area.Whether IOU or formula that formula (1) calculates (2) overlap calculated, numerical value is higher to illustrate that overlapping area is bigger.Overlapping area is greater than preset area threshold The first auxiliary area delete, corresponding the first auxiliary area for just retaining overlapping area and being less than or equal to area threshold, by the Two auxiliary areas and the first auxiliary area remained are as the first interest region.The above process is exactly RPN algorithm search The all processes in the first interest region, and in order to reduce subsequent calculation amount, one NMS algorithm of nesting is gone back in RPN algorithm, For deleting the excessive candidate region of overlapping area.

Step S102 detects Pixel Dimensions of the references object in the target image, as reference pixel size.

The label information of target object and target is determined by first object region in first object region has been determined After the object pixel size of object, terminal device continues to test the Pixel Dimensions of references object in target image, referred to as reference image Plain size.Whether terminal device detection target image is color image first, if target image is color image, it is necessary first to will Target image gray processing, and the image after gray processing is known as target gray image；If target image originally gray level image, Just be not used as any processing, but at this time will target image be known as target gray image.Color image is converted into gray level image It is that 3 channels (channel R, the channel G, channel B) of color image are converted into only one single channel image.Terminal is set It is standby to be based on edge detection, identify the connection region in target gray image, referred to as the first reference zone.Wherein, edge detection can It, can be by continuous and closure contour line to identify target object and the lines of outline of references object in target gray image Region corresponding to item is as connection region.It is illustrated by taking first reference zone as an example below and how to determine corresponding One probability, terminal device input the gray level image in the first reference zone in trained disaggregated model in advance, the classification mould Classifier in type is two classifiers, which can calculate general comprising references object in the first reference zone Rate, referred to as the first probability.For example, references object is card；So two classifiers, which can identify in the first reference zone, includes First probability of card, certain first probability is higher, illustrates that corresponding first reference zone is more possible to comprising references object. When classifier in train classification models, it is only necessary to two class samples (positive sample and negative sample), positive sample be include reference The reference zone of object, negative sample are the reference zones not comprising references object.

If there is multiple first reference zones, terminal device can determine each first reference zone pair using aforesaid way The first probability answered.The first probability obtained according to classifier calculated, terminal device select full in multiple first reference zones First reference zone of sufficient matching condition, and using the first reference zone chosen as the second reference zone.According to second The Pixel Dimensions of reference zone in the target image determine the size of reference pixel, for example, by the pixel of the second reference zone it is long or Person is using the elemental area of the second reference zone as object pixel size.

Step S103 obtains the reference actual size of the references object, according to the object pixel size, the reference Pixel Dimensions and it is described refer to actual size, determine the target actual size of the target object.

Specifically, terminal device obtains the actual size (referred to as referring to actual size) of references object, according to target object Object pixel size, the reference actual size of the reference pixel size of references object and references object, proportionally relationship (proportionate relationship specifically: the ratio and object pixel size and target between reference pixel size and reference actual size are practical Ratio between size is identical), calculate the actual size of target object, referred to as target actual size.Wherein it is possible to according to formula (3) target actual size is calculated:

Wherein, A1 indicates reference pixel size；A2 indicates to refer to actual size；B1 indicates object pixel size；B2 is indicated Target actual size.

Optionally, subsequent target service data further to be determined according to the target actual size of target object, and The attribute classification of target object plays decisive role to target service data again, therefore terminal device also needs to determine target object Attribute, determine that the other detailed process of Attribute class is: terminal device obtain first object region corresponding to label information, it is above-mentioned Label information is determined according to the attribute classification of the corresponding attribute type feature of confidence level in first object region.Terminal device According to the corresponding label information in first object region and target actual size, target industry associated with service request is determined Business data.For example, target service data can be business Claims Resolution amount, target pair when service request is settlement of insurance claim request The sign informations such as weight, the volume of elephant, it is, determining the signs such as the weight of target object according to the actual size of target object Information, and then business Claims Resolution amount is determined according to the label information of target object in sign information and first object region.

It optionally, can be with displaying target business datum, target service data after terminal device has determined target service data It include: business Claims Resolution amount, and the sign information (weight, volume etc.) with target object.Terminal device is by target service number According to transmission as the associated service terminal of service request (the user terminal 10b in the embodiment as corresponding to above-mentioned Fig. 1 d).

Further, Fig. 3 is referred to, a kind of Fig. 3 process for detecting reference pixel size provided in an embodiment of the present invention is shown It is intended to.As shown in figure 3, the detailed process of detection reference pixel size includes the following steps S201- step S204, and step S201- step S204 is a specific embodiment of step S102 in embodiment corresponding to Fig. 2:

The target image is converted to target gray image by step S201, and by the connection in the target gray image Lead to region as the first reference zone.

Specifically, target image is converted to gray level image, referred to as target gray image by terminal device.According to preset ladder Operator is spent, bench operator can be sobel gradient operator, canny gradient operator, laplace gradient operator etc..It is calculated using gradient Son can rapidly calculate the gradient of each pixel in target image, combine the image that the gradient of each pixel obtains and claim For gradient map, first gradient figure is known as by the gradient map that target image obtains.Terminal device carries out closed operation to first gradient figure, Obtained gradient image is known as the second gradient map, and closed operation is that first gradient figure is first expanded post-etching, after executing closed operation The minuscule hole in first gradient figure can be filled up and connect the adjacent object in first gradient figure.First gradient figure and It include the lines of outline of all objects (including target object and references object) in target image in two gradient maps, terminal device will The region that continuous and closure lines of outline is identified in first gradient figure is known as being referred to as connection region, equally by the second gradient The region that continuous and closure lines of outline is identified in figure is referred to as connection region.Since above-mentioned connection region may be not The regional graphics of rule also need terminal device that processing of becoming a full member is made in determining connection region, so that the connection area that becomes a full member that treated Domain is the regional graphics (for example, rectangle) of rule.To become a full member treated, connection region is known as auxiliary reference region.Due to auxiliary Reference zone corresponds to first gradient figure and the second gradient map, therefore also needs to be mapped in target gray figure, be by There is the region with auxiliary reference region same position information in target gray figure, as the first reference zone.For example, auxiliary Coordinate of the reference zone in first gradient figure is respectively (3,3,20,20), wherein the first item in above-mentioned coordinate indicates auxiliary The starting abscissa of reference zone；The starting ordinate in Section 2 expression auxiliary reference region；Section 3 indicates auxiliary reference area The length in domain；The width in Section 4 expression auxiliary reference region.Will be in target gray figure, coordinate " (3,3,20,20) " mark Region as the first reference zone.

Step S202 inputs the gray level image in multiple first reference zones in disaggregated model respectively, identifies each the It include the first probability of the references object in one reference zone.

Specifically, being illustrated how terminal device calculates the first reference zone by taking first reference zone as an example below Corresponding first probability.Terminal device inputs the gray level image in the first reference zone in trained disaggregated model in advance, Classifier in the disaggregated model is two classifiers, which can calculate in the first reference zone comprising reference The probability of object, referred to as the first probability.For example, references object is card；So two classifiers can identify the first reference area It include the first probability of card in domain, certain first probability is higher, illustrates that corresponding first reference zone is more possible to include References object.Wherein, disaggregated model can be based on made of the convolutional neural networks training in deep learning.If having multiple One reference zone can be based on disaggregated model, identify corresponding first probability of each first reference zone respectively.

Step S203 chooses in the multiple first reference zone according to first probability and meets matching condition First reference zone, as the second reference zone.

Specifically, terminal device extracts maximum first probability in multiple first probability.Maximum first probability is detected, if Maximum first probability is less than or equal to preset probability threshold value, illustrates that the first probability determined by disaggregated model cannot function as screening The precondition of first reference zone, therefore terminal device is needed according to the location information, every where each first reference zone The size of the Aspect Ratio of a first reference zone, each first reference zone determines the weight of each first reference zone, In, if the position of the first reference zone closer to the center of target image, then weight is bigger；If the length and width of the first reference zone Than setting length-width ratio condition closer to pre-, then weight is bigger；If the size of the first reference zone closer to pre- size condition, So weight is bigger.After the corresponding weight of each first reference zone is calculated, maximum first reference zone of weight is made For the first reference zone for meeting matching condition, and the first reference zone of matching condition will be met as the second reference zone, It is using maximum first reference zone of weight as the second reference zone.

If maximum first probability is greater than preset probability threshold value, illustrate that the first probability determined by disaggregated model can be used as The precondition of the first reference zone is screened, therefore corresponding first reference zone of maximum first probability is determined as by terminal device Meet the first reference zone of matching condition, and the first reference zone of matching condition will be met as the second reference zone, i.e., It is using corresponding first reference zone of maximum first matching probability as the second reference zone.No matter above-mentioned it is found that maximum first Which condition probability meets, and first reference zone is only had chosen from multiple first reference zones as the second reference area Domain, remaining first reference zone need not be referring to subsequent arithmetic.

Step S204 determines the reference pixel size according to the size of second reference zone.

Specifically, references object is divided into fixed the first references object and unfixed second reference pair of shape of shape As.For example, coin is exactly the first fixed references object of shape, and card is (since there may be crimpings or folding for paper card Trace etc.) it is exactly unfixed second references object of shape.Different references object, corresponding different calculation.Work as reference pair When as belonging to fixed the first references object of shape, same terminal device is based on gradient operator, extracts in the second reference zone The marginal information of gray level image, available gradient image corresponding with gray level image in the second reference zone, referred to as assists Gradient image.It is converted based on unique step hough, terminal device determines full curve and the center of circle in auxiliary gradient image, will Above-mentioned full curve is known as first object curve.According to first object curve and the center of circle, can determine and first object curve Corresponding diameter, referred to as first object diameter.Using first object diameter as the reference pixel size of references object.

When references object belongs to unfixed second references object of shape, as the target gray where the second reference zone Image maps back target image, is that terminal device will have and the second reference zone same position information in the target image Region, as third reference zone.According to the color of image in third reference zone, terminal device is in third reference zone The color of image carries out clustering processing, to eliminate the dash area on third reference zone boundary, improves subsequent the second reference of calculating The precision of the reference pixel size of object.Treated that region is known as cluster result region by color cluster for terminal device, according to The size in cluster result region determines the reference pixel size of references object.Such as by the pixel in cluster result region it is long or Using the elemental area in cluster result region or the diagonal line length in cluster result region as object pixel size.

The embodiment of the present invention is by obtaining the target image comprising target object and references object, and detected target object is in mesh Pixel Dimensions in logo image as object pixel size, and detect the Pixel Dimensions of references object in the target image, as Reference pixel size；The reference actual size for obtaining references object, according to object pixel size, reference pixel size and with reference to real Border size determines the target actual size of target object.It is above-mentioned it is found that from the image comprising target object and references object Detected target object Pixel Dimensions in the picture and references object Pixel Dimensions in the picture respectively, then obtain reference pair The full-size(d) of elephant can determine the full-size(d) of target object under a proportional relationship, so as to automatically determine target The full-size(d) of object avoids measuring size in a manual manner, improves the efficiency of measurement target object size.

It is a kind of flow diagram of image processing method provided in an embodiment of the present invention please also refer to Fig. 4, at image Reason method includes the following steps:

Step S301 obtains the target image comprising the target object and references object, detects the target object and exist Pixel Dimensions in the target image, as object pixel size.

Wherein, the specific implementation of step S301 may refer to the step S101 in embodiment corresponding to above-mentioned Fig. 2.

Step S302 is rolled up when the references object belongs to unfixed second references object of shape based on mask region Convolutional layer in product neural network model carries out process of convolution to the target image, obtains by the second convolution characteristic information group Second convolution characteristic pattern made of conjunction.

Specifically, when references object belongs to unfixed second references object of shape, based on being constructed in terminal device Convolutional layer in the mask region convolutional neural networks model (the identification model 20g in such as above-mentioned Fig. 1 c) of completion, terminal device Process of convolution is carried out to target image, that is, randomly selects the sub-fraction characteristic information in target image as sample (convolution Core), all target images are successively slipped over using this sample as a window, that is, above-mentioned sample and target image are rolled up Product operation, to obtain the convolution characteristic information about the second references object in target image.After process of convolution, terminal device Available convolution characteristic pattern, include in the convolution characteristic pattern is the convolution feature letter of the second references object in target image Convolution characteristic information obtained above is known as the second convolution characteristic information by breath, terminal device, and convolution obtained above is special Sign image is known as the second convolution characteristic pattern.

Step S303 searches for multiple second interest region in the second convolution characteristic pattern, to each second region of interest The the second convolution characteristic information for including in domain carries out pond processing, obtains the second structure feature information.

Specifically, being based on RPN algorithm, terminal device searches for multiple interest regions in the second convolution characteristic pattern, referred to as the Two interest regions, wherein the detailed process based on RPN algorithm search interest region may refer in above-mentioned Fig. 2 corresponding embodiment Step S101.It is illustrated by taking a second interest region as an example below, if there is multiple second interest region terminal device can be with Adopt the second structure feature information calculated in each second interest region in a like fashion.To including in the second interest region Second convolution characteristic information carries out pond processing, is aggregate statistics the second convolution characteristic information, available after aggregate statistics About the static structure characteristic information of the second references object in target image, referred to as the second structure feature information.

Step S304 identifies that described each second is emerging according to the second feature information for including in each second interest region It include the second probability of second references object in interesting region.

Specifically, based on the classifier in the convolutional neural networks model of mask region, above-mentioned classifier can export each The matching probability (referred to as the second probability) of second structure feature information and the second references object, that is, terminal device identification are each It include the second probability of the second references object in second interest region.Certain second probability is higher, illustrates corresponding second interest Region is more possible to comprising the second references object.When classifier in train classification models, it is only necessary to two class sample (positive samples Sheet and negative sample), positive sample is the interest region for including the second references object, and negative sample is not comprising the second references object Interest region.

Maximum second probability corresponding second interest region is determined as auxiliary mark region, and calculates institute by step S305 State the binary cover of each pixel in auxiliary mark region.

Specifically, since there are multiple second interest regions, and each second interest region corresponds to second probability, eventually End equipment is using maximum second probability corresponding second interest region as auxiliary mark region.It has been determined that auxiliary mark region is i.e. complete At the position detection to the second references object in target image, since the second references object is irregular figure, in order to mention Height calculates the precision of the Pixel Dimensions of the second references object, and terminal device is also needed further to be partitioned into auxiliary mark region The specific corresponding pixel of second references object.Based on the recurrence mask layer in the convolutional neural networks model of mask region, calculate The binary cover of each pixel in target auxiliary area.Binary cover includes foreground mask and background mask, and foreground mask indicates Corresponding pixel belongs to the second references object, and background mask indicates that corresponding pixel belongs to background.

All pixels group corresponding to the binary cover for belonging to foreground mask is combined into target subgraph, and root by step S306 The reference pixel size is sized to according to the target subgraph.

Specifically, image made of the combination of pixels for belonging to foreground mask is by terminal device is known as target subgraph (as above State the target subgraph 20h in Fig. 1 c), according to the size of target subgraph, determine the reference pixel size of the second references object.Such as By the pixel length of target subgraph or by the elemental area of target subgraph or using the diagonal line length of target subgraph as target picture Plain size.Above-mentioned mask region convolutional neural networks model based on splitting object algorithm Mask-RCNN training and come, Mask- RCNN can accurately obtain the location and shape of the second references object in the target image, for measuring the second of small volume References object when more accurate Pixel Dimensions can guarantee the actual size of subsequent determining target object, has smaller error.

Optionally, when references object belongs to unfixed first references object of shape, based on being constructed in terminal device Complete the second complete convolutional neural networks model in convolutional layer, terminal device to target image carry out process of convolution, i.e., with Machine chooses the sub-fraction characteristic information in target image as sample (convolution kernel), successively using this sample as a window All target images are slipped over, that is, above-mentioned sample and target image do convolution algorithm, to obtain the pass in target image In the convolution characteristic information of the first references object.After process of convolution, the available convolution characteristic pattern of terminal device, the convolution feature Include in figure is the convolution characteristic information of the first references object in target image, and convolution characteristic information obtained above is known as Third convolution characteristic information, and convolution characteristic image obtained above is known as third convolution characteristic pattern.Based on RPN algorithm, eventually End equipment searches for multiple interest regions, referred to as third interest region in third convolution characteristic pattern, wherein being based on RPN algorithm search The detailed process in interest region may refer to the step S101 in above-mentioned Fig. 2 corresponding embodiment.Respectively in third interest region The third convolution characteristic information for including carries out pond processing, is aggregate statistics third convolution characteristic information, can after aggregate statistics To obtain the static structure characteristic information about the first references object in target image, referred to as third structure feature information.It is based on Classifier in second complete convolutional neural networks model, terminal device identify each third feature information and the first references object Matching probability (referred to as third probability), that is to say identification each third interest region in comprising the first references object third it is general Rate.Certain third probability is higher, illustrates that corresponding third interest region is more possible to comprising the first references object.In training point When classifier in class model, it is only necessary to two class samples (positive sample and negative sample), positive sample be include the first references object Interest region, negative sample is the interest region not comprising the first references object.Above-mentioned second complete convolutional neural networks model Based on algorithm of target detection RFCN training come.Due to there are multiple third interest region, and each third interest region pair A third probability is answered, terminal device is using maximum third probability corresponding third interest region as the second target area.Terminal Image in second target area is converted to gray level image by equipment, is equally based on gradient operator, is extracted above-mentioned gray level image Marginal information, available gradient image corresponding with image in the second target area, referred to as edge gradient image.Based on etc. Step-length hough transformation, terminal device can determine full curve and the center of circle in edge gradient image, by above-mentioned full curve Referred to as the second aim curve.Terminal device can determine corresponding with the second aim curve according to the second aim curve and the center of circle Diameter, referred to as the second aimed dia.Using the second aimed dia as the reference pixel size of references object.

Step S307 obtains the reference actual size of the references object, according to the object pixel size, the reference Pixel Dimensions and it is described refer to actual size, determine the target actual size of the target object.

Wherein, the specific implementation of step S307 may refer to the step S103 in embodiment corresponding to above-mentioned Fig. 2.

The embodiment of the present invention is by obtaining the target image comprising target object and references object, and detected target object is in mesh Pixel Dimensions in logo image as object pixel size, and detect the Pixel Dimensions of references object in the target image, as Reference pixel size；The reference actual size for obtaining references object, according to object pixel size, reference pixel size and with reference to real Border size determines the target actual size of target object.It is above-mentioned it is found that from the image comprising target object and references object Detected target object Pixel Dimensions in the picture and references object Pixel Dimensions in the picture respectively, then obtain reference pair The full-size(d) of elephant can determine the full-size(d) of target object under a proportional relationship, so as to automatically determine target The full-size(d) of object avoids measuring size in a manual manner, improves the efficiency of measurement target object size.

Further, Fig. 5 is referred to, is a kind of structural schematic diagram of image processing apparatus provided in an embodiment of the present invention. As shown in figure 5, image processing apparatus 1 may include: to obtain module 11, first detection module 12, the second detection module 13, determine Module 14.

Module 11 is obtained, the target image comprising the target object and references object is obtained；

Specifically, obtaining module 11 receives service request associated with target object, target is obtained according to service request Image includes target object and references object in the target image, wherein the object for full-size(d) to be calculated is known as mesh Object is marked, for being known as references object with the object that target object is made comparisons.

First detection module 12, for detecting Pixel Dimensions of the target object in the target image, as mesh Mark Pixel Dimensions；

Specifically, the convolutional layer in the first complete convolutional neural networks model completed based on building, first detection module 12 pairs of target images carry out process of convolution, the available convolution characteristic pattern of first detection module 12, include in the convolution characteristic pattern Be target object in target image convolution characteristic information, first detection module 12 claims the convolution characteristic information of target object For the first convolution characteristic information, and the convolution characteristic image being composed of the first convolution characteristic information is known as the first convolution spy Sign figure.Based on RPN (network is chosen in Region Proposal Network, region) algorithm, searched in the first convolution characteristic pattern Multiple interest regions, referred to as the first interest region.Pond is carried out to the first convolution characteristic information for including in the first interest region Processing, is aggregate statistics the first convolution characteristic information, available about target object in target image after aggregate statistics Static structure characteristic information, referred to as first structure characteristic information.Based on the classifier in the first complete convolutional neural networks model, It identifies in the first structure characteristic information and the first complete convolutional neural networks model in the first interest region after aggregate statistics The matching degree of multiple attribute type features, referred to as the first matching degree, that is, each first interest region will correspond to multiple One matching degree, and in the quantity and the first complete convolutional neural networks model of corresponding first matching degree in each first interest region The quantity of attribute type feature is identical, wherein each attribute type feature corresponds to an attribute classification.In each first region of interest In corresponding multiple first matching degrees in domain, first detection module 12 is using maximum first matching degree as the confidence in the first interest region Degree, and terminal device is using the attribute classification of the corresponding attribute type feature of maximum first matching degree as in the first interest region The label information of object.

Second detection module 13, for detecting Pixel Dimensions of the references object in the target image, as ginseng Examine Pixel Dimensions；

Specifically, the second detection module 13 continues to test the Pixel Dimensions of references object in target image, referred to as reference image Plain size.Second detection module 13 first detects whether target image is color image, if target image is color image, first It needs target image gray processing, and the image after gray processing is known as target gray image；If target image is originally grey Spend image, be just not used as any processing, but at this time will target image be known as target gray image.Second detection module 13 is based on Edge detection identifies the connection region in target gray image, referred to as the first reference zone.Wherein, edge detection can identify Target object and the lines of outline of references object in target gray image out, can be right by continuous and closure lines of outline institute The region answered is as connection region.Second detection module 13 is trained in advance by the gray level image input in the first reference zone In disaggregated model, the classifier in the disaggregated model is two classifiers, which can calculate the first reference area It include the probability of references object in domain, referred to as the first probability, certain first probability is higher, illustrates corresponding first reference zone just It is more possible to comprising references object.

Determining module 14, for obtaining the reference actual size of the references object, according to the object pixel size, institute It states reference pixel size and described with reference to actual size, determines the target actual size of the target object.

Specifically, determining module 14 obtains the actual size (referred to as referring to actual size) of references object, according to target pair The reference actual size of the object pixel size of elephant, the reference pixel size of references object and references object, is proportionally closed System's (proportionate relationship specifically: the ratio and object pixel size and target between reference pixel size and reference actual size are real Ratio between the size of border is identical), calculate the actual size of target object, referred to as target actual size.

Wherein, it is real that module 11, first detection module 12, the second detection module 13, the concrete function of determining module 14 are obtained Existing mode may refer to the step S101- step S103 in above-mentioned Fig. 2 corresponding embodiment, be not discussed here.

Fig. 5 is referred to, first detection module 12 may include: the first convolution unit 121, search unit 122.

First convolution unit 121, for based on the convolutional layer in the described first complete convolutional neural networks model, to described Target image carries out process of convolution, obtains the first convolution characteristic pattern being composed of the first convolution characteristic information；

Search unit 122, for searching for multiple first interest region in the first convolution characteristic pattern；

First convolution unit 121 is also used to the first convolution characteristic information for including in each first interest region Pond processing is carried out, first structure characteristic information is obtained, identifies the first structure feature letter for including in each first interest region First matching degree of breath and multiple attribute type features in the described first complete convolutional neural networks model；

First convolution unit 121 is also used in corresponding multiple first matching degrees in each first interest region, will Maximum first matching degree is as the corresponding confidence level in the first interest region；

First convolution unit 121 is also used in the corresponding confidence level in multiple first interest region, by maximum confidence Corresponding first interest region is spent as first object region, and using the size in the first object region as the target picture Plain size.

Wherein, the first convolution unit 121, search unit 122 concrete function implementation to may refer to above-mentioned Fig. 2 corresponding Step S101 in embodiment, is not discussed here.

Refer to Fig. 5, search unit 122 may include: that score computation subunit 1221, score determine subelement 1222.

Score computation subunit 1221 is used for the sliding window on the first convolution characteristic pattern, in each window base Anchor point frame under different scale determines multiple candidate regions；

The score computation subunit 1221 is also used to calculate separately the first convolution characteristic information in each candidate region Target prospect score；

Score determines subelement 1222, for the target prospect score according to each candidate region, determines described more A first interest region.

Wherein, score computation subunit 1221, score determine that the concrete function implementation of subelement 1222 may refer to Step S101 in above-mentioned Fig. 2 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, score determines that subelement 1222 may include: the first determining subelement 12221, delete reservation son list Member 12222.

First determines subelement 12221, and the candidate region for the target prospect score to be greater than to score threshold determines For the first auxiliary area；

Described first determines subelement 12221, is also used in first auxiliary area, will have maximum target prospect First auxiliary area of score is determined as the second auxiliary area；

It deletes and retains subelement 12222, for calculating separately second auxiliary area and remaining first auxiliary area Between overlapping area, by overlapping area be greater than area threshold the first auxiliary area delete, and retain overlapping area be less than or Person is equal to the first auxiliary area of the area threshold；

Described first determines subelement 12221, is also used to second auxiliary area and the first auxiliary area retained It is determined as first interest region.

Wherein, first determine subelement 12221, delete retain subelement 12222 concrete function implementation can join See the step S101 in above-mentioned Fig. 2 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, the second detection module 13 may include: converting unit 131, area determination unit 132, selecting unit 133, size determination unit 134.

Converting unit 131, for the target image to be converted to target gray image；

Area determination unit 132, for using the connection region in the target gray image as the first reference zone；

The converting unit 131 is also used to the gray level image in multiple first reference zones inputting disaggregated model respectively In, identify the first probability in each first reference zone comprising the references object；

Selecting unit 133, for being chosen in the multiple first reference zone and meeting matching according to first probability First reference zone of condition, as the second reference zone；

Size determination unit 134 determines the reference pixel size for the size according to second reference zone.

Wherein, the specific function of converting unit 131, area determination unit 132, selecting unit 133, size determination unit 134 The mode of being able to achieve may refer to the step S102 in above-mentioned Fig. 2 corresponding embodiment, be not discussed here.

Fig. 5 is referred to, area determination unit 132 may include: gradient computation subunit 1321, the first detection sub-unit 1322。

Gradient computation subunit 1321, for calculating corresponding first ladder of the target gray image according to gradient operator Degree figure, and closed operation is carried out to the first gradient figure, obtain the second gradient map；

First detection sub-unit 1322, for detecting the connection area of the first gradient figure and second gradient map respectively Domain, and the connection region that will test out is become a full member processing, and multiple auxiliary reference regions are obtained；

First detection sub-unit 1322 is also used to have and the auxiliary reference in the target gray image The region of region same position information, as first reference zone.

Wherein, gradient computation subunit 1321, the first detection sub-unit 1322 concrete function implementation may refer to Step S201 in above-mentioned Fig. 3 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, selecting unit 133 may include: that the first extraction subelement 1331, second determines subelement 1332.

First extracts subelement 1331, for extracting maximum first probability in multiple first probability；

Second determines subelement 1332, if being less than or equal to probability threshold value for maximum first probability, according to institute State the ruler of location information where the first reference zone, the Aspect Ratio of first reference zone, first reference zone It is very little, it determines the weight of first reference zone, is determined as maximum first reference zone of weight to meet the matching condition The first reference zone, and maximum first reference zone of the weight is determined as second reference zone；

Described second determines subelement 1332, will if being also used to maximum first probability is greater than the probability threshold value Corresponding first reference zone of maximum first probability is determined as meeting the first reference zone of the matching condition, and by institute It states corresponding first reference zone of maximum first probability and is determined as second reference zone.

Wherein, the first extraction subelement 1331, second determines that the concrete function implementation of subelement 1332 may refer to Step S203 in above-mentioned Fig. 3 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, size determination unit 134 may include: the second extraction subelement 1341, the second detection sub-unit 1342。

Second extracts subelement 1341, for mentioning when the references object belongs to the first fixed references object of shape The gray-scale Image Edge information in second reference zone is taken, auxiliary gradient image is obtained；

Second detection sub-unit 1342, for detecting the full curve in the auxiliary gradient image, as first object Curve determines first object diameter according to the first object curve；

Second detection sub-unit 1342 is also used to the first object diameter being determined as the reference pixel ruler It is very little.

Wherein, second extraction subelement 1341, the second detection sub-unit 1342 concrete function implementation may refer to Step S204 in above-mentioned Fig. 3 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, size determination unit 134 may include: the second extraction subelement 1341, the second detection sub-unit 1342；It can also include: that cluster subelement 1343, size determine subelement 1344.

Subelement 1343 is clustered, it, will be for when the references object belongs to unfixed second references object of shape In the target image, there is the region with the second reference zone same position information, as third reference zone；

The cluster subelement 1343, is also used to the color according to the image in the third reference zone, to described Image in three reference zones carries out color cluster processing, obtains cluster result region；

Size determines subelement 1344, for the size according to the cluster result region, determines the reference pixel ruler It is very little.

Wherein, cluster that subelement 1343, that size determines that the concrete function implementation of subelement 1344 may refer to is above-mentioned Step S204 in Fig. 3 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, the second detection module 13 may include: converting unit 131, area determination unit 132, selecting unit 133, size determination unit 134 can also include: the second convolution unit 135, recognition unit 136, computing unit 137.

Second convolution unit 135, for being based on when the references object belongs to unfixed second references object of shape Convolutional layer in the convolutional neural networks model of mask region carries out process of convolution to the target image, obtains by the second convolution The second convolution characteristic pattern that characteristic information is composed；

Recognition unit 136, for searching for multiple second interest region in the second convolution characteristic pattern, to each second The the second convolution characteristic information for including in interest region carries out pond processing, obtains the second structure feature information；

The recognition unit 136 is also used to identify institute according to the second feature information for including in each second interest region State the second probability in each second interest region comprising second references object；

Computing unit 137, for maximum second probability corresponding second interest region to be determined as auxiliary mark region, and Calculate the binary cover of each pixel in the auxiliary mark region；

The computing unit 137 is also used to belong to all pixels group corresponding to the binary cover of foreground mask and is combined into Target subgraph, and the reference pixel size is sized to according to the target subgraph.

Wherein, the second convolution unit 135, recognition unit 136, computing unit 137 concrete function implementation can join See the step S302- step S306 in above-mentioned Fig. 4 corresponding embodiment, is not discussed here.

Fig. 5 is referred to, the second detection module 13 may include: converting unit 131, area determination unit 132, selecting unit 133, size determination unit 134, the second convolution unit 135, recognition unit 136, computing unit 137 can also include: third volume Product unit 138, extraction unit 139.

Third convolution unit 138, for being based on institute when the references object belongs to the first fixed references object of shape The convolutional layer in the second complete convolutional neural networks model is stated, process of convolution is carried out to the target image, obtains being rolled up by third The third convolution characteristic pattern that product characteristic information is composed；

The third convolution unit 138 is also used to search for multiple third interest region in the third convolution characteristic pattern, Pond processing is carried out to the third convolution characteristic information for including in each third interest region, obtains third structure feature information；

The third convolution unit 138 is also used to be known according to the third feature information for including in each third interest region It include the third probability of first references object in not described each third interest region；

The third convolution unit 138 is also used in the corresponding third probability in multiple third interest region, will have most The third interest region of big third probability is determined as the second target area；

Extraction unit 139 obtains features of edge gradient maps for extracting the marginal information of the image in second target area Picture；

The third convolution unit 138 is also used to detect the full curve in the edge gradient image, as the second mesh Curve is marked, the second aimed dia is determined according to second aim curve, and second aimed dia is determined as the ginseng Examine Pixel Dimensions.

Wherein, third convolution unit 138, extraction unit 139 concrete function implementation to may refer to above-mentioned Fig. 4 corresponding Step S306 in embodiment, is not discussed here.

Fig. 5 is referred to, module 11 is obtained and is specifically used for:

Service request associated with the target object is received, obtaining according to the service request includes the target pair As the target image with the references object.

Then described image detection device further include:

Display module 15, for by the Attribute class of the corresponding attribute type feature of the confidence level in the first object region Not, as the first object region corresponding label information, according to the corresponding label information in the first object region, Yi Jisuo Target actual size is stated, determines target service data associated with the service request.

The display module 15, is also used to displaying target business datum, and stores the target service data；The target Business datum includes: the sign information of business Claims Resolution amount and the target object；

Sending module 16, it is whole for the target service data to be sent to business associated with the service request End.

Wherein, the concrete function implementation for obtaining module 11 may refer to the step in above-mentioned Fig. 2 corresponding embodiment S101 is not discussed here；Display module 15, sending module 16 concrete function implementation may refer to above-mentioned Fig. 2 Step S103 in corresponding embodiment, is not discussed here.

The Pixel Dimensions of detected target object in the picture are distinguished from the image comprising target object and references object, And references object Pixel Dimensions in the picture, then the full-size(d) of references object is obtained, mesh can be determined under a proportional relationship The full-size(d) of object is marked, so as to automatically determine the full-size(d) of target object, avoids measuring in a manual manner Size improves the efficiency of measurement target object size.

Further, Fig. 6 is referred to, is the structural schematic diagram of a kind of electronic equipment provided in an embodiment of the present invention.Such as Fig. 6 Shown, the image processing apparatus 1 in above-mentioned Fig. 6 can be applied to the electronic equipment 1000, and the electronic equipment 1000 can be with It include: processor 1001, network interface 1004 and memory 1005, in addition, the electronic equipment 1000 can also include: user Interface 1003 and at least one communication bus 1002.Wherein, communication bus 1002 is logical for realizing the connection between these components Letter.Wherein, user interface 1003 may include display screen (Display), keyboard (Keyboard), and optional user interface 1003 is also It may include standard wireline interface and wireless interface.Network interface 1004 optionally may include the wireline interface, wireless of standard Interface (such as WI-FI interface).Memory 1005 can be high speed RAM memory, be also possible to non-labile memory (non- Volatile memory), a for example, at least magnetic disk storage.Memory 1005 optionally can also be that at least one is located at Storage device far from aforementioned processor 1001.As shown in fig. 6, as can in a kind of memory 1005 of computer storage medium To include operating system, network communication module, Subscriber Interface Module SIM and equipment control application program.

In electronic equipment 1000 shown in Fig. 6, network interface 1004 can provide network communication function；And user interface 1003 are mainly used for providing the interface of input for user；And processor 1001 can be used for calling what is stored in memory 1005 to set Standby control application program, to realize:

The target image comprising the target object and references object is obtained, detects the target object in the target figure Pixel Dimensions as in, as object pixel size, and detect Pixel Dimensions of the references object in the target image, As reference pixel size；

The reference actual size for obtaining the references object, according to the object pixel size, the reference pixel size Actual size is referred to described, determines the target actual size of the target object.

In one embodiment, the processor 1001 is executing the detection target object in the target image Pixel Dimensions specifically execute following steps when as object pixel size:

Based on the convolutional layer in the first complete convolutional neural networks model, process of convolution is carried out to the target image, is obtained To the first convolution characteristic pattern being composed of the first convolution characteristic information；

Multiple first interest region is searched in the first convolution characteristic pattern, to including in each first interest region First convolution characteristic information carries out pond processing, obtains first structure characteristic information；

Identify the first structure characteristic information and the described first complete convolutional Neural net for including in each first interest region First matching degree of multiple attribute type features in network model；

In corresponding multiple first matching degrees in each first interest region, using maximum first matching degree as the first interest The corresponding confidence level in region；

In the corresponding confidence level in multiple first interest region, using maximum confidence corresponding first interest region as One target area, and using the size in the first object region as the object pixel size.

In one embodiment, the processor 1001 searches for multiple first in the first convolution characteristic pattern in execution When interest region, following steps are specifically executed:

The sliding window on the first convolution characteristic pattern is determined based on the anchor point frame under different scale in each window Multiple candidate regions；

The target prospect score for calculating separately the first convolution characteristic information in each candidate region, according to each candidate The target prospect score in region determines the multiple first interest region.

In one embodiment, the processor 1001 is obtained in execution according to the target prospect of each candidate region Point, when determining the multiple first interest region, specifically execute following steps:

The candidate region that the target prospect score is greater than score threshold is determined as the first auxiliary area；

In first auxiliary area, it is auxiliary that the first auxiliary area with maximum target prospect score is determined as second Help region；

The overlapping area between second auxiliary area and remaining first auxiliary area is calculated separately, by overlapping area The first auxiliary area greater than area threshold is deleted, and it is first auxiliary less than or equal to the area threshold to retain overlapping area Help region；

Second auxiliary area and the first auxiliary area retained are determined as first interest region.

In one embodiment, the processor 1001 is executing the detection references object in the target image Pixel Dimensions specifically execute following steps when as reference pixel size:

The target image is converted into target gray image, and using the connection region in the target gray image as First reference zone；

Gray level image in multiple first reference zones is inputted in disaggregated model respectively, identifies each first reference zone Interior the first probability comprising the references object；

According to first probability, the first reference area for meeting matching condition is chosen in the multiple first reference zone Domain, as the second reference zone；

According to the size of second reference zone, the reference pixel size is determined.

In one embodiment, the processor 1001 execute using the connection region in the target gray image as When the first reference zone, following steps are specifically executed:

According to gradient operator, the corresponding first gradient figure of the target gray image is calculated, and to the first gradient figure Closed operation is carried out, the second gradient map is obtained；

The connection area detected the connection region of the first gradient figure and second gradient map respectively, and will test out Domain is become a full member processing, and multiple auxiliary reference regions are obtained；

There to be the region with the auxiliary reference region same position information in the target gray image, as institute State the first reference zone.

In one embodiment, the processor 1001 is being executed according to first probability, in the multiple first ginseng The first reference zone that selection meets matching condition in the domain of examination district specifically executes following steps when as the second reference zone:

Maximum first probability is extracted in multiple first probability；

If maximum first probability is less than or equal to probability threshold value, the position where first reference zone The size of information, the Aspect Ratio of first reference zone, first reference zone, determines first reference zone Maximum first reference zone of weight is determined as meeting the first reference zone of the matching condition by weight, and by the power Maximum first reference zone of weight is determined as second reference zone；

If maximum first probability is greater than the probability threshold value, maximum first probability corresponding first is referred to Region is determined as meeting the first reference zone of the matching condition, and by corresponding first reference area of maximum first probability Domain is determined as second reference zone.

In one embodiment, the processor 1001 is executing the size according to second reference zone, determines institute When stating reference pixel size, following steps are specifically executed:

When the references object belongs to the first fixed references object of shape, the ash in second reference zone is extracted Image edge information is spent, auxiliary gradient image is obtained；

The full curve in the auxiliary gradient image is detected, it is bent according to the first object as first object curve Line determines first object diameter；

The first object diameter is determined as the reference pixel size.

In one embodiment, the processor 1001 is executing the size according to second reference zone, determines institute When stating reference pixel size, following steps are specifically executed:

When the references object belongs to unfixed second references object of shape, will have in the target image With the region of the second reference zone same position information, as third reference zone；

According to the color of the image in the third reference zone, color is carried out to the image in the third reference zone Clustering processing obtains cluster result region；

According to the size in the cluster result region, the reference pixel size is determined.

In one embodiment, the processor 1001 is executing the detection references object in the target image Pixel Dimensions specifically execute following steps when as reference pixel size:

When the references object belongs to unfixed second references object of shape, it is based on mask region convolutional neural networks Convolutional layer in model carries out process of convolution to the target image, obtain being composed of the second convolution characteristic information the Two convolution characteristic patterns；

Multiple second interest region is searched in the second convolution characteristic pattern, to including in each second interest region Second convolution characteristic information carries out pond processing, obtains the second structure feature information；

According to the second feature information for including in each second interest region, each second interest region Zhong Bao is identified The second probability containing second references object；

Maximum second probability corresponding second interest region is determined as auxiliary mark region, and calculates the auxiliary mark The binary cover of each pixel in region；

All pixels group corresponding to the binary cover for belonging to foreground mask is combined into target subgraph, and according to the target Subgraph is sized to the reference pixel size.

In one embodiment, the processor 1001 is detecting pixel of the references object in the target image Size specifically executes following steps when as reference pixel size:

When the references object belongs to the first fixed references object of shape, it is based on the second complete convolutional neural networks mould Convolutional layer in type carries out process of convolution to the target image, obtains the third being composed of third convolution characteristic information Convolution characteristic pattern；

Multiple third interest region is searched in the third convolution characteristic pattern, to including in each third interest region Third convolution characteristic information carries out pond processing, obtains third structure feature information；

According to the third feature information for including in each third interest region, each third interest region Zhong Bao is identified Third probability containing first references object；

In the corresponding third probability in multiple third interest region, there will be the third interest region of maximum third probability true It is set to the second target area；

The marginal information for extracting the image in second target area, obtains edge gradient image；

The full curve in the edge gradient image is detected, it is bent according to second target as the second aim curve Line determines the second aimed dia, and second aimed dia is determined as the reference pixel size.

In one embodiment, the processor 1001 is obtaining the target image comprising target object and references object When, specifically execute following steps:

Service request associated with the target object is received, obtaining according to the service request includes the target pair As the target image with the references object；

The processor 1001 also executes following steps:

By the attribute classification of the corresponding attribute type feature of the confidence level in the first object region, as first mesh Region corresponding label information is marked, according to the corresponding label information in the first object region and the target actual size, really Fixed target service data associated with the service request；

Displaying target business datum, and store the target service data；The target service data include: business Claims Resolution The sign information of amount and the target object；

The target service data are sent to service terminal associated with the service request.

The embodiment of the present invention is by obtaining the target image comprising target object and references object, and detected target object is in mesh Pixel Dimensions in logo image as object pixel size, and detect the Pixel Dimensions of references object in the target image, as Reference pixel size；The reference actual size for obtaining references object, according to object pixel size, reference pixel size and with reference to real Border size determines the target actual size of target object.It is above-mentioned it is found that from the image comprising target object and references object Detected target object Pixel Dimensions in the picture and references object Pixel Dimensions in the picture respectively, then obtain reference pair The full-size(d) of elephant can determine the full-size(d) of target object under a proportional relationship, so as to automatically determine target The full-size(d) of object avoids measuring size in a manual manner, improves the efficiency of measurement target object size.

It should be appreciated that real corresponding to executable Fig. 2 to the Fig. 4 above of electronic equipment 1000 described in the embodiment of the present invention The description in example to described image processing method is applied, also can be performed in embodiment corresponding to Fig. 5 above and described image processing is filled 1 description is set, details are not described herein.In addition, being described to using the beneficial effect of same procedure, also no longer repeated.

In addition, it need to be noted that: the embodiment of the invention also provides a kind of computer storage medium, and the meter Computer program performed by the image processing apparatus 1 being mentioned above, and the computer journey are stored in calculation machine storage medium Sequence includes program instruction, when the processor executes described program instruction, is able to carry out the corresponding implementation of Fig. 2 to Fig. 4 above Therefore description in example to described image processing method will be repeated no longer here.In addition, having to using same procedure Beneficial effect description, is also no longer repeated.For undisclosed skill in computer storage medium embodiment according to the present invention Art details please refers to the description of embodiment of the present invention method.

Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage medium In, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access Memory, RAM) etc..

The above disclosure is only the preferred embodiments of the present invention, cannot limit the right model of the present invention with this certainly It encloses, therefore equivalent changes made in accordance with the claims of the present invention, is still within the scope of the present invention.

Claims (15)

1. a kind of image processing method characterized by comprising

The target image comprising target object and references object is obtained, picture of the target object in the target image is detected Plain size as object pixel size, and detects Pixel Dimensions of the references object in the target image, as reference Pixel Dimensions；

The reference actual size for obtaining the references object, according to the object pixel size, the reference pixel size and institute It states with reference to actual size, determines the target actual size of the target object.

2. the method according to claim 1, wherein the detection target object is in the target image Pixel Dimensions, as object pixel size, comprising:

Based on the convolutional layer in the first complete convolutional neural networks model, process of convolution is carried out to the target image, obtain by The first convolution characteristic pattern that first convolution characteristic information is composed；

Multiple first interest region is searched in the first convolution characteristic pattern, to include in each first interest region first Convolution characteristic information carries out pond processing, obtains first structure characteristic information；

Identify the first structure characteristic information and the described first complete convolutional neural networks mould for including in each first interest region First matching degree of multiple attribute type features in type；

In corresponding multiple first matching degrees in each first interest region, using maximum first matching degree as the first interest region Corresponding confidence level；

In the corresponding confidence level in multiple first interest region, using maximum confidence corresponding first interest region as the first mesh Region is marked, and using the size in the first object region as the object pixel size.

3. according to the method described in claim 2, it is characterized in that, described search for multiple in the first convolution characteristic pattern One interest region, comprising:

The sliding window on the first convolution characteristic pattern is determined based on the anchor point frame under different scale multiple in each window Candidate region；

The target prospect score for calculating separately the first convolution characteristic information in each candidate region, according to each candidate region Target prospect score, determine the multiple first interest region.

4. according to the method described in claim 3, it is characterized in that, the target prospect according to each candidate region obtains Point, determine the multiple first interest region, comprising:

The candidate region that the target prospect score is greater than score threshold is determined as the first auxiliary area；

In first auxiliary area, the first auxiliary area with maximum target prospect score is determined as the second auxiliary region Domain；

The overlapping area between second auxiliary area and remaining first auxiliary area is calculated separately, overlapping area is greater than First auxiliary area of area threshold is deleted, and retains the first auxiliary region that overlapping area is less than or equal to the area threshold Domain；

Second auxiliary area and the first auxiliary area retained are determined as first interest region.

5. the method according to claim 1, wherein the detection references object is in the target image Pixel Dimensions, as reference pixel size, comprising:

The target image is converted into target gray image, and using the connection region in the target gray image as first Reference zone；

Gray level image in multiple first reference zones is inputted in disaggregated model respectively, identifies packet in each first reference zone The first probability containing the references object；

According to first probability, the first reference zone for meeting matching condition is chosen in the multiple first reference zone, As the second reference zone；

According to the size of second reference zone, the reference pixel size is determined.

6. according to the method described in claim 5, it is characterized in that, the connection region by the target gray image is made For the first reference zone, comprising:

According to gradient operator, the corresponding first gradient figure of the target gray image is calculated, and the first gradient figure is carried out Closed operation obtains the second gradient map；

The connection region detected the connection region of the first gradient figure and second gradient map respectively, and will test out turns Positive processing, obtains multiple auxiliary reference regions；

To have in the target gray image and the region of the auxiliary reference region same position information, as described the One reference zone.

7. according to the method described in claim 5, it is characterized in that, described according to first probability, the multiple first The first reference zone for meeting matching condition is chosen in reference zone, as the second reference zone, comprising:

Maximum first probability is extracted in multiple first probability；

Position letter if maximum first probability is less than or equal to probability threshold value, where first reference zone The size of breath, the Aspect Ratio of first reference zone, first reference zone, determines the power of first reference zone Maximum first reference zone of weight is determined as meeting the first reference zone of the matching condition by weight, and by the weight Maximum first reference zone is determined as second reference zone；

If maximum first probability is greater than the probability threshold value, by corresponding first reference zone of maximum first probability It is determined as meeting the first reference zone of the matching condition, and corresponding first reference zone of maximum first probability is true It is set to second reference zone.

8. according to the method described in claim 5, it is characterized in that, the size according to second reference zone, determines The reference pixel size, comprising:

When the references object belongs to the first fixed references object of shape, the grayscale image in second reference zone is extracted As marginal information, auxiliary gradient image is obtained；

The full curve in the auxiliary gradient image is detected, it is true according to the first object curve as first object curve Determine first object diameter；

The first object diameter is determined as the reference pixel size.

9. according to the method described in claim 5, it is characterized in that, the size according to second reference zone, determines The reference pixel size, comprising:

When the references object belongs to unfixed second references object of shape, will have and institute in the target image The region for stating the second reference zone same position information, as third reference zone；

According to the color of the image in the third reference zone, color cluster is carried out to the image in the third reference zone Processing, obtains cluster result region；

According to the size in the cluster result region, the reference pixel size is determined.

10. the method according to claim 1, wherein the detection references object is in the target image In Pixel Dimensions, as reference pixel size, comprising:

When the references object belongs to unfixed second references object of shape, it is based on mask region convolutional neural networks model In convolutional layer, to the target image carry out process of convolution, obtain the volume Two being composed of the second convolution characteristic information Product characteristic pattern；

Multiple second interest region is searched in the second convolution characteristic pattern, to include in each second interest region second Convolution characteristic information carries out pond processing, obtains the second structure feature information；

According to the second feature information for including in each second interest region, identify to include institute in each second interest region State the second probability of the second references object；

Maximum second probability corresponding second interest region is determined as auxiliary mark region, and calculates the auxiliary mark region In each pixel binary cover；

All pixels group corresponding to the binary cover for belonging to foreground mask is combined into target subgraph, and according to the target subgraph It is sized to the reference pixel size.

11. the method according to claim 1, wherein the detection references object is in the target image In Pixel Dimensions, as reference pixel size, comprising:

When the references object belongs to the first fixed references object of shape, based in the second complete convolutional neural networks model Convolutional layer, to the target image carry out process of convolution, obtain the third convolution being composed of third convolution characteristic information Characteristic pattern；

Multiple third interest region is searched in the third convolution characteristic pattern, to the third for including in each third interest region Convolution characteristic information carries out pond processing, obtains third structure feature information；

According to the third feature information for including in each third interest region, identify to include institute in each third interest region State the third probability of the first references object；

In the corresponding third probability in multiple third interest region, will there is the third interest region of maximum third probability to be determined as Second target area；

The marginal information for extracting the image in second target area, obtains edge gradient image；

The full curve in the edge gradient image is detected, it is true according to second aim curve as the second aim curve Fixed second aimed dia, and second aimed dia is determined as the reference pixel size.

12. according to the method described in claim 2, it is characterized in that, described obtain the mesh comprising target object and references object Logo image, comprising:

Receive associated with target object service request, according to the service request obtain comprising the target object with The target image of the references object；

Then the method also includes:

By the attribute classification of the corresponding attribute type feature of the confidence level in the first object region, as the first object area Domain corresponding label information, according to the corresponding label information in the first object region and the target actual size, determine with The associated target service data of service request；

It shows the target service data, and stores the target service data；The target service data include: business Claims Resolution The sign information of amount and the target object；

The target service data are sent to service terminal associated with the service request.

13. a kind of image processing apparatus characterized by comprising

Module is obtained, for obtaining the target image comprising the target object and references object；

First detection module, for detecting Pixel Dimensions of the target object in the target image, as object pixel Size；

Second detection module, for detecting Pixel Dimensions of the references object in the target image, as reference pixel Size；

Determining module, for obtaining the reference actual size of the references object, according to the object pixel size, the reference Pixel Dimensions and it is described refer to actual size, determine the target actual size of the target object.

14. a kind of electronic equipment characterized by comprising processor and memory；

The processor is connected with memory, wherein the memory is for storing program code, and the processor is for calling Said program code, to execute such as the described in any item methods of claim 1-12.

15. a kind of computer storage medium, which is characterized in that the computer storage medium is stored with computer program, described Computer program includes program instruction, and described program is instructed when being executed by a processor, executed such as any one of claim 1-12 The method.