CN109559300A

CN109559300A - Image processing method, electronic equipment and computer readable storage medium

Info

Publication number: CN109559300A
Application number: CN201811378299.9A
Authority: CN
Inventors: 宋涛; 李想之
Original assignee: Shanghai Sensetime Intelligent Technology Co Ltd
Current assignee: Shanghai Sensetime Intelligent Technology Co Ltd
Priority date: 2018-11-19
Filing date: 2018-11-19
Publication date: 2019-04-02

Abstract

The embodiment of the present application discloses a kind of image processing method, electronic equipment and computer readable storage medium, and wherein method includes: to obtain destination image data to be processed；Based on default neural network model, multi-scale prediction is carried out to the destination image data using N number of network output layer as prediction interval, obtains the first prediction result, first prediction result includes at least one prediction block, and the N is the integer greater than 1；Substep Screening Treatment is carried out at least one described prediction block, obtains target prediction as a result, it is possible to increase pneumonia diagnosis efficiency and accuracy based on chest x-ray piece.

Description

Image processing method, electronic equipment and computer readable storage medium

Technical field

The present invention relates to field of image processings, and in particular to a kind of image processing method, electronic equipment and computer-readable Storage medium.

Background technique

In the world, pneumonia accounts for 15% or more of all 5 years old or less death of child.Although pneumonia is common, Accurate Diagnosis Pneumonia is a difficult task, it is needed by well-trained specialist examination chest x-ray piece (coherent X- Radiation, CXR), and pass through clinical medical history, vital sign and experimental check confirmation.Pneumonia is usually expressed as impermeable on CXR The increased region of lightness, because there are many other diseases for lung, also shows as CXR however, the diagnosis of pneumonia is very complicated on CXR On opacity increase.In addition, the problem of each shift of clinician all suffers from reading great amount of images, diagnosis of pneumonia needs Radiologist has the judgement of expert level, however good radiologist concentrates on a few regions, other regions It is difficult to enjoy the diagnosis of expert level, causes the death rate of pneumonia higher, it is seen then that currently based on the pneumonia diagnosis efficiency of CXR It is not high with accuracy.

Summary of the invention

The embodiment of the present application provides a kind of image processing method, electronic equipment and computer readable storage medium, can be with Improve pneumonia diagnosis efficiency and accuracy based on CXR.

The embodiment of the present application first aspect provides a kind of image processing method, comprising:

Obtain destination image data to be processed；

Based on default neural network model, the destination image data is carried out using N number of network output layer as prediction interval Multi-scale prediction obtains the first prediction result, and first prediction result includes at least one prediction block, and the N is greater than 1 Integer；

Substep Screening Treatment is carried out at least one described prediction block, obtains target prediction result.

In a kind of optional embodiment, before acquisition destination image data to be processed, the method is also wrapped It includes: raw image data to be processed is converted to the destination image data for meeting target component.

In a kind of optional embodiment, described be converted to raw image data to be processed meets target component Destination image data includes:

The raw image data is converted to the destination image data of default gray value and/or preset image sizes.

It is described based on default neural network model in a kind of optional embodiment, using N number of network output layer as in advance It surveys layer and multi-scale prediction is carried out to the destination image data, obtain the first prediction result, first prediction result includes extremely After a few prediction block, further includes:

Obtain the anchor point in the default neural network model as the characteristic pattern of N number of network output layer of the prediction interval Parameter；

Multi-scale prediction is carried out to the destination image data in the prediction interval based on the anchor parameter, described in acquisition N number of anchor point data that each pixel on the characteristic pattern of prediction interval generates；

The matched target anchor point of prediction block is determined in the anchor point data according to preset matching rule；

According to default weighting algorithm obtain recurrence loss parameter of the prediction block based on the matched target anchor point and Classification Loss parameter, first prediction result include that the prediction block and the prediction block are based on the matched target anchor The recurrence loss parameter and Classification Loss parameter of point.

In a kind of optional embodiment, described according to preset matching rule, determination is described pre- in the anchor point data Surveying the matched target anchor point of frame includes:

The friendship of the prediction block and the anchor point data and ratio are obtained, the friendship is determined in the anchor point data and ratio is most Big anchor point data are the matched target anchor point of the prediction block.

In a kind of optional embodiment, the preset matching rule includes default anchor point numerical value M, and the basis is default Matching rule determines that the matched target anchor point of prediction block includes: in the anchor point data

Obtain the friendship of the prediction block and the anchor point data and ratio, by it is described friendship and than according to descending sequence into Row sequence determines in the sequence and hands over and be the matched target anchor point of the prediction block than corresponding anchor point data for first M.

It is described that substep Screening Treatment is carried out at least one described prediction block in a kind of optional embodiment, it obtains Target prediction result includes:

Substep Screening Treatment is carried out at least one described prediction block based on two step non-maxima suppression methods, described in acquisition Target prediction result.

In a kind of optional embodiment, the two step non-maxima suppression methods that are based on are at least one described prediction Frame carries out substep Screening Treatment, and obtaining the target prediction result includes:

The first Overlapping parameters and coordinate of the prediction block are obtained, first Overlapping parameters is deleted and is greater than first threshold Prediction block, using the mean value of the coordinate of the prediction block of the deletion and the coordinate of the prediction block of reservation as the seat of new prediction block Mark；

The second Overlapping parameters of the new prediction block are obtained, second Overlapping parameters is deleted and is less than the pre- of second threshold Survey frame.

In a kind of optional embodiment, the method also includes:

Obtain the class probability of the prediction block；

If the prediction block that first Overlapping parameters are greater than first threshold is two, which comprises

It deletes first Overlapping parameters to be greater than in the prediction block of first threshold, the lower prediction block of class probability.

In a kind of optional embodiment, the basis presets weighting algorithm and obtains the prediction block based on the matching Target anchor point recurrence loss parameter and Classification Loss parameter include:

Friendship based on the prediction block and the anchor point data and ratio calculating target weight value, according to the target weight value It is calculated using the loss function of weighting and obtains the recurrence loss parameter of the prediction block based on the matched target anchor point；

It is calculated according to the target weight value using the cross entropy loss function of weighting and obtains the prediction block based on described The Classification Loss parameter of matched target anchor point.

In a kind of optional embodiment, the method also includes:

The recurrence loss parameter and the Classification Loss parameter are inputted into optimization module with preset ratio, learned based on default Strategy is practised to be updated the parameter of the prediction interval.

In a kind of optional embodiment, the default neural network model includes visual geometric group network model.

In a kind of optional embodiment, the target prediction result further includes the coordinate of the prediction block, the side Method further include:

Processing is marked to the raw image data according to the target prediction result, output includes the mark The prediction address of raw image data after reason.

The embodiment of the present application second aspect provides a kind of electronic equipment, comprising: obtains module, prediction module and screening mould Block, in which:

The acquisition module, for obtaining destination image data to be processed；

The prediction module, for being based on default neural network model, using N number of network output layer as prediction interval to described Destination image data carries out multi-scale prediction, obtains the first prediction result, first prediction result includes at least one prediction Frame, the N are the integer greater than 1；

The screening module obtains target prediction knot for carrying out substep Screening Treatment at least one described prediction block Fruit.

In a kind of optional embodiment, the electronic equipment further include:

Image conversion module, for raw image data to be processed to be converted to the target image number for meeting target component According to.

In a kind of optional embodiment, described image conversion module is specifically used for turning the raw image data It is changed to the destination image data of default gray value and/or preset image sizes.

In a kind of optional embodiment, the prediction module includes first unit, second unit and third unit, In:

The first unit is used for:

After the prediction module obtains above-mentioned first prediction result, conduct in the default neural network model is obtained The anchor parameter of the characteristic pattern of N number of network output layer of the prediction interval；

The second unit is used for, and determines that the prediction block is matched in the anchor point data according to preset matching rule Target anchor point；

The third unit is used for, and is obtained the prediction block according to default weighting algorithm and is based on the matched target anchor point Recurrence loss parameter and Classification Loss parameter, first prediction result include that the prediction block and the prediction block are based on The recurrence loss parameter and Classification Loss parameter of the matched target anchor point.

In a kind of optional embodiment, the second unit is specifically used for, and obtains the prediction block and the anchor point The friendships of data and ratio determine the friendship in the anchor point data and are the matched mesh of the prediction block than maximum anchor point data Mark anchor point.

In a kind of optional embodiment, the preset matching rule includes default anchor point numerical value M；

The second unit is specifically used for, and obtains the friendship of the prediction block and the anchor point data and ratio, simultaneously by the friendship Than being ranked up according to descending sequence, determines in the sequence and hand over and be described pre- than corresponding anchor point data for first M Survey the matched target anchor point of frame.

In a kind of optional embodiment, the screening module is specifically used for, and is based on two step non-maxima suppression methods Substep Screening Treatment is carried out at least one described prediction block, obtains the target prediction result.

In a kind of optional embodiment, the screening module is specifically used for:

In a kind of optional embodiment, the acquisition module is also used to, and obtains the class probability of the prediction block；

The screening module is specifically used for:

If the prediction block that first Overlapping parameters are greater than first threshold is two, deletes first Overlapping parameters and be greater than In the prediction block of first threshold, the lower prediction block of class probability.

In a kind of optional embodiment, the third unit is specifically used for:

In a kind of optional embodiment, the electronic equipment further includes optimization module, and the optimization module is used for: will The recurrence loss parameter and the Classification Loss parameter are obtained with preset ratio, and according to the recurrence loss parameter and described Classification Loss parameter is updated based on parameter of the default learning strategy to the prediction interval.

In a kind of optional embodiment, the target prediction result further includes the coordinate of the prediction block；

The electronic equipment further includes reporting modules, is used for according to the target prediction result to the raw image data It is marked processing, prediction address of the output comprising the label treated raw image data.

The embodiment of the present application third aspect provides another electronic equipment, including processor and memory, the storage Device is for storing one or more programs, and one or more of programs are configured to be executed by the processor, described program Including for executing the step some or all of as described in the embodiment of the present application first aspect either method.

The embodiment of the present application fourth aspect provides a kind of computer readable storage medium, the computer readable storage medium For storing the computer program of electronic data interchange, wherein the computer program executes computer as the application is real Some or all of apply described in a first aspect either method step.

The embodiment of the present application is by obtaining destination image data to be processed, based on default neural network model, with N number of net Network output layer as prediction interval to the destination image data carry out multi-scale prediction, obtain the first prediction result, described first Prediction result includes at least one prediction block, and the N is the integer greater than 1, then carries out substep at least one described prediction block Screening Treatment obtains target prediction as a result, it is possible to increase pneumonia diagnosis efficiency and accuracy based on CXR.

Detailed description of the invention

In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below There is attached drawing needed in technical description to be briefly described.

Fig. 1 is a kind of flow diagram of image processing method disclosed in the embodiment of the present application；

Fig. 2 is the flow diagram of another kind image processing method disclosed in the embodiment of the present application；

Fig. 3 is the structural schematic diagram of a kind of electronic equipment disclosed in the embodiment of the present application；

Fig. 4 is the structural schematic diagram of another kind electronic equipment disclosed in the embodiment of the present application.

Specific embodiment

In order to enable those skilled in the art to better understand the solution of the present invention, below in conjunction in the embodiment of the present application Attached drawing, the technical scheme in the embodiment of the application is clearly and completely described, it is clear that described embodiment is only A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art Every other embodiment obtained without creative efforts, shall fall within the protection scope of the present invention.

Description and claims of this specification and term " first " in above-mentioned attached drawing, " second " etc. are for distinguishing Different objects, are not use to describe a particular order.In addition, term " includes " and " having " and their any deformations, it is intended that It is to cover and non-exclusive includes.Such as the process, method, system, product or equipment for containing a series of steps or units do not have It is defined in listed step or unit, but optionally further comprising the step of not listing or unit, or optionally also wrap Include other step or units intrinsic for these process, methods, product or equipment.

Referenced herein " embodiment " is it is meant that a particular feature, structure, or characteristic described can wrap in conjunction with the embodiments Containing at least one embodiment of the present invention.Each position in the description occur the phrase might not each mean it is identical Embodiment, nor the independent or alternative embodiment with other embodiments mutual exclusion.Those skilled in the art explicitly and Implicitly understand, embodiment described herein can be combined with other embodiments.

Electronic equipment involved by the embodiment of the present application can permit other multiple terminal devices and access.Above-mentioned electricity Sub- equipment includes terminal device, in the specific implementation, above-mentioned terminal device is including but not limited to such as with touch sensitive surface (example Such as, touch-screen display and/or touch tablet) mobile phone, laptop computer or tablet computer etc it is other portable Formula equipment.It is to be further understood that in certain embodiments, the equipment is not portable communication device, but has and touch The desktop computer of sensing surface (for example, touch-screen display and/or touch tablet).

The concept of deep learning in the embodiment of the present application is derived from the research of artificial neural network.Multilayer sense containing more hidden layers Know that device is exactly a kind of deep learning structure.Deep learning, which forms more abstract high level by combination low-level feature, indicates Attribute class Other or feature, to find that the distributed nature of data indicates.

Deep learning is a kind of based on the method for carrying out representative learning to data in machine learning.Observation (such as a width Image) various ways can be used to indicate, such as vector of each pixel intensity value, or be more abstractively expressed as a series of Side, region of specific shape etc..And use certain specific representation methods be easier from example learning tasks (for example, face Identification or human facial expression recognition).The benefit of deep learning is feature learning and the layered characteristic with non-supervisory formula or Semi-supervised It extracts highly effective algorithm and obtains feature by hand to substitute.Deep learning is a new field in machine learning research, motivation Be to establish, simulation human brain carries out the neural network of analytic learning, the mechanism that it imitates human brain explains data, such as image, Sound and text.

It describes in detail below to the embodiment of the present application.

Referring to Fig. 1, Fig. 1 is a kind of flow diagram of image procossing disclosed in the embodiment of the present application, as shown in Figure 1, The image processing method can be executed by above-mentioned electronic equipment, be included the following steps:

101, raw image data to be processed is converted to the destination image data for meeting target component.

Before executing image procossing by neural network model, first raw image data to be processed can be returned One change pretreatment, be converted to the destination image data for meeting target component, then execute step 102, can also directly acquire to The raw image data of processing is as destination image data, then executes step 102, that is, do not need image conversion can also carry out it is pre- It surveys.The main purpose of image preprocessing is to eliminate information unrelated in image, restores useful real information, enhancing is for information about Detectability and to the maximum extent simplify data, to improve the reliability of feature extraction, image segmentation, matching and identification.

The raw image data to be processed mentioned in the embodiment of the present application can be to be obtained by various medical image equipments The medical image obtained, especially chest x-ray piece (coherent X-radiation, CXR) have diversity, in the picture may be used It can be presented as the diversity of the features such as gray value of image, picture size, raw image data in the embodiment of the present application can be with It is an at least CXR image.

Above-mentioned target component can be understood as the parameter of description characteristics of image, i.e., for making above-mentioned raw image data in system The regulation parameter of one style.For example, above-mentioned target component may include: big for describing image resolution ratio, image grayscale, image The parameter of the features such as small.

Electronic equipment in the embodiment of the present application can store above-mentioned target component, on above-mentioned target component may include Default gray value and/or preset image sizes are stated, the raw image data can be converted to default gray value and/or are preset The destination image data of picture size.

Specifically, above-mentioned raw image data can be normalized within the scope of the gray value codomain of 0-255, it is being related to When the training process of the neural network model in the embodiment of the present application, the gray value mean value (ratio obtained in training set can be counted As mean value be 125), and, the picture size of the above-mentioned raw image data of input is converted into above-mentioned preset image sizes, than Such as 800 from 1024x 1024 to 800x, the matrix for being merged into three channels (i.e. 3x 800x 800) can also be repeated, it is last every A channel all subtracts above-mentioned gray value mean value, the purpose for the arrangement is that in order to utilize the model trained in advance in natural image Carry out initialization feature network with parameter, makes that the detection algorithm in the embodiment of the present application is restrained faster and generalization is stronger.

Wherein, in order to expand data volume, the generalization ability of model is improved, can be used positive sample data (containing pneumonia) In prediction block (Ground Truth box, GT box) the pneumonia region that includes cut out to be fused to normal image data New training sample is generated in (non-pneumonia), is then added in original data set.For example, in the training process, former There is the training set of 6000 pictures, the data set generated in advance with the method for above-mentioned synthesis totally 7000 picture, in this way and original The training set merging come just has 13000 pictures as new training set.

By pre-processing to raw image data, its diversity can be reduced, neural network model can provide more Stable judgement.

102, based on default neural network model, using N number of network output layer as prediction interval to above-mentioned destination image data Multi-scale prediction is carried out, obtains the first prediction result, above-mentioned first prediction result includes at least one prediction block, and above-mentioned N is big In 1 integer.

In the embodiment of the present application, above-mentioned default neural network model can be using visual geometric group network mould trained in advance Type (Visual Geometry Group Network, the VGG) backbone of model as network.

The method used in the embodiment of the present application includes a kind of algorithm of target detection, specifically can be understood as designing one kind One stage detection framework.The One stage detection algorithm mentioned in the embodiment of the present application does not need candidate region and determines (region proposal) stage directly generates the class probability and position coordinate value of object, can be direct by single detection Final testing result is obtained, therefore has detection speed faster, than more typical algorithm such as YOLO, SSD, Retina- Net。

Specifically, above-mentioned multi-scale prediction can be understood as selecting multiple network output layers as prediction interval to above-mentioned target Image data is predicted, primarily to prediction obtains position and its class probability of different size of prediction block.

K-means algorithm is hard clustering algorithm, is the representative of the typically objective function clustering method based on prototype, it is Data point obtains the tune of interative computation using the method that function seeks extreme value to certain objective function of distance as optimization of prototype Whole rule.For K-means algorithm using Euclidean distance as similarity measure, it is to seek corresponding a certain initial cluster center vector V most Optimal sorting class, so that evaluation index J is minimum, algorithm is using error sum of squares criterion function as clustering criteria function.

In the embodiment of the present application, it can be designed on different prediction intervals according to the GT box in training set using K-means Anchor point data (anchor) size simultaneously matches with the receptive field of corresponding prediction interval.It intuitively says, receptive field is exactly visual impression By the size in region.In convolutional neural networks, the definition of receptive field is the characteristic pattern of each layer of convolutional neural networks output The area size that pixel (pixel) on (feature map) maps on the original image.For example, raw image data is The characteristic pattern size of 800x800, this layer are 50x50, it can be assumed that a pixel roughly correspond to original image 800/50 × 800/50 region, i.e., can probably see so big region, and actual calculation method will consider neural network Convolution kernel size and the step-length of convolution etc..

Wherein, above-mentioned N number of network output layer is selected as prediction interval, such as N=4, it can using 4 different predictions Layer, the anchor size being arranged on the characteristic pattern of each prediction interval can be different, in this way on the characteristic pattern of expression prediction interval Each pixel on generate 4 anchor of different sizes.By based on the multiple dimensioned pre- of above-mentioned default neural network model It surveys, obtains the first prediction result, above-mentioned first prediction result can be understood as the knot after the matching of above-mentioned anchor and GT box Fruit, including at least one prediction block, i.e., the pneumonia region of tentative prediction in pneumonia detection application.Obtaining above-mentioned first prediction As a result step 103 can be executed after.

In a kind of optional embodiment, can also include: after above-mentioned steps 102

What is obtained through the above steps can also include not only the recurrence loss parameter and classification including the prediction block Loss parameter, can be used as the Optimal Parameters of above-mentioned default neural network model, and above-mentioned steps are related to default neural network mould The training and learning process of type, specifically may refer to the associated description in embodiment shown in Fig. 2.

The embodiment of the present application designs a kind of detection framework for being directed to pneumonia detection, optimizes anchor size and prediction interval It chooses, keeps its prediction and matching more acurrate efficiently；In addition, it is optional, it is very high in order to alleviate the position confidence level of GT box of prediction But the lower imbalance problem of the probabilistic confidence of classification, it devises recurrence loss and Classification Loss based on weighting and calculates, Above-mentioned first prediction result obtained can be made more accurate.

103, substep Screening Treatment is carried out at least one above-mentioned prediction block, obtains target prediction result.

After obtaining above-mentioned first prediction result, prediction block therein can be screened, obtain target prediction knot Fruit.Specifically, can be realized by the screening process of substep, for example screen, can will be greater than during the first screening in two steps The prediction block of first threshold removes, and the coordinate of remaining prediction block takes the average value of these prediction blocks, and second step is again with one A lower second threshold removes the prediction block of coincidence, wherein above-mentioned first threshold is greater than above-mentioned second threshold.

Specifically, can realize above-mentioned Screening Treatment by two step non-maxima suppression methods in the embodiment of the present application.Its In, above-mentioned non-maxima suppression (Non-Maximum Suppression, NMS), as the term suggests it is exactly to inhibit not to be maximum Element, it can be understood as local maxima search, general this locally represents is a neighborhood, neighborhood there are two changeable parameters, First is that the dimension of neighborhood, second is that the size of neighborhood.

Existing NMS operation is all single step operation, only takes the coordinate of the highest prediction block of classification confidence level to be used as and retains, Have ignored the contribution for the frame that those classification confidences are low but positioning confidence level is high.Reasoning rank of the embodiment of the present application in step 102 In the last handling process of section, for pneumonia problem, a kind of two stage NMS method is proposed, overlapping frame can filtered out The stability of prediction block is improved simultaneously.

Specifically, can be removed using a kind of two stage NMS strategy, first stage by standard of a higher threshold value Extra frame, then the frame of coincidence is filtered by the NMS of second stage with a lower threshold value.

In a kind of possible embodiment, above-mentioned two steps non-maxima suppression method mainly may include:

The first Overlapping parameters and coordinate of above-mentioned prediction block are obtained, above-mentioned first Overlapping parameters is deleted and is greater than first threshold Prediction block, using the mean value of the coordinate of the prediction block of above-mentioned deletion and the coordinate of the prediction block of reservation as the seat of new prediction block Mark；

The second Overlapping parameters of above-mentioned new prediction block are obtained, above-mentioned second Overlapping parameters is deleted and is less than the pre- of second threshold Survey frame.

Specifically, above-mentioned Overlapping parameters can be understood as the parameter of description Duplication, it is in alternative embodiments, above-mentioned Overlapping parameters may be to hand over and than (Intersection-over-Union, IoU), be one used in target detection Concept is the overlapping rate of the candidate frame (candidate bound) generated and former indicia framing (ground truth bound), i.e., The ratio of their intersection and union.Most ideally completely overlapped, i.e., ratio is 1.

For example, Overlapping parameters (overlap) can be first passed through greater than 0.7 to remove the prediction block of high superposed, and The coordinate of the prediction block of reservation is the mean value for being removed frame and itself coordinate, more stable prediction block available in this way； Second step setting overlap is greater than 0.05 to remove the prediction block of overlapping (because being not heavy under pneumonia detection truth It closes).

Screening Treatment is carried out by the above method, the above-mentioned target prediction result comprising final prediction block can be obtained.With When the image procossing of pneumonia diagnosis, above-mentioned target prediction result can be understood as the pneumonia position predicted and hospital and physical examination Center can realize pneumonia prediction based on patient chest X-ray, improve diagnostic level, reduction is failed to pinpoint a disease in diagnosis.

The embodiment of the present application is by obtaining destination image data to be processed, based on default neural network model, with N number of net Network output layer as prediction interval to the destination image data carry out multi-scale prediction, obtain the first prediction result, described first Prediction result includes at least one prediction block, and the N is the integer greater than 1, then carries out substep at least one described prediction block Screening Treatment obtains target prediction as a result, the specific pneumonia position of CXR image prediction can be based on, improves the pneumonia based on CXR and examine Disconnected efficiency and accuracy.

Referring to Fig. 2, Fig. 2 is the flow diagram of another kind image processing method disclosed in the embodiment of the present application, Fig. 2 is It is advanced optimized on the basis of Fig. 1.The main body for executing the embodiment of the present application step can be used for medicine shadow to be a kind of As the electronic equipment of processing.As shown in Fig. 2, the image processing method includes the following steps:

201, raw image data to be processed is converted to the destination image data for meeting target component.

202, based on default neural network model, using N number of network output layer as prediction interval to above-mentioned destination image data Multi-scale prediction is carried out, obtains the first prediction result, above-mentioned first prediction result includes at least one prediction block, and above-mentioned N is big In 1 integer.

Wherein, above-mentioned steps 201 and step 202 can refer to the step 101 and step 102 of embodiment illustrated in fig. 1 respectively In specific descriptions, details are not described herein again.

203, the anchor point in above-mentioned default neural network model as the characteristic pattern of N number of network output layer of prediction interval is obtained Parameter.

Above-mentioned anchor parameter can be understood as the parameter of description anchor point size, can specifically include anchor point size and/or anchor Point ratio.The setting of prediction interval can be previously stored in electronic equipment, the anchor including prediction layer number N and features described above figure Point parameter.For example, when being preferable to provide 4 prediction intervals and realizing, the anchor size that is arranged on the characteristic pattern of first prediction interval For [0.1,0.1], the anchor size being arranged on the characteristic pattern of second prediction interval is [0.2,0.2], third prediction interval The anchor size being arranged on characteristic pattern is [0.4,0.4], and the anchor size being arranged on the characteristic pattern of the 4th prediction interval is [0.8,0.8].The ratio of anchor on each layer is { [1,1], [0.3,1], [0.6,1], [1.2,1] }, is indicated in this way It can produce 4 anchor of different sizes on each pixel on the characteristic pattern of prediction interval.

After the anchor parameter for determining prediction interval, the characteristic pattern for getting above-mentioned prediction interval, step 204 can be executed.

204, multi-scale prediction is carried out to above-mentioned destination image data in above-mentioned prediction interval based on above-mentioned anchor parameter, obtained N number of anchor point data that each pixel on the characteristic pattern of above-mentioned prediction interval generates.

Wherein, the prediction interval of above-mentioned default neural network model can choose the network output of N number of different scale (size) Layer (convolutional layer), it is a kind of multi-scale prediction that N number of prediction interval, which can be used for predicting the pneumonia prediction block of different scale size respectively, Frame.Such as when carrying out convolution, to the image of input extract respectively third and fourth, the obtained characteristic pattern of five layers of convolution, then They are zoomed into same size and a kind of performance of more sizes, wherein the characteristic pattern that deeper convolutional layer extracts more is taken out As the feature extracted is more advanced.

202 prediction block (GT box) in above-mentioned destination image data can be obtained through the above steps, by above-mentioned Step 204 can obtain N number of anchor point data (anchor) of the generation of each pixel on the characteristic pattern of above-mentioned prediction interval.It can be with Understand, the step 203- step 206 in the embodiment of the present application can be default neural network model and carry out prediction process The step of middle execution, it can be the processing in order to obtain this prediction result, it is understood that for the default neural network mould Type, because its parameter obtained can be used as the basis of optimization, is learnt and is instructed in the process for carrying out autonomous learning and optimization Practice.

Correspondingly, when preferred N=4, the anchor size being arranged on the characteristic pattern of first prediction interval be [0.1, 0.1], the anchor size being arranged on the characteristic pattern of second prediction interval is [0.2,0.2], on the characteristic pattern of third prediction interval The anchor size of setting is [0.4,0.4], the anchor size being arranged on the characteristic pattern of the 4th prediction interval be [0.8, 0.8].The ratio of anchor on each layer is { [1,1], [0.3,1], [0.6,1], [1.2,1] }, in this way expression prediction interval Characteristic pattern on each pixel on can produce 4 anchor of different sizes.

In the embodiment of the present application, what above-mentioned pixel represented is the meaning on a kind of coordinate, for example, with 50x50's Characteristic pattern illustrates, it is assumed that its receptive field is all 16*16 size, it is assumed that representing original image 16*16 in each pixel The receptive field of size, the region predicted so is also within the scope of this either with or without pneumonia, because not knowing true lung here Scorching prediction block width and be how many at high proportion, so provided with four kinds of sizes anchor, always have one in this way with lung Scorching true frame is closer, can simple subsequent backpropagation learning process, here it is a pixels to be responsible for 4 The prediction of anchor, and other several characteristic patterns can also execute step same as described above.

After tentatively obtaining above-mentioned prediction block and above-mentioned anchor point data, step 205 can be executed.

205, the above-mentioned matched target anchor point of prediction block is determined in above-mentioned anchor point data according to preset matching rule.

In general, the position confidence level of prediction block and class probability confidence level are uneven, for example, predicted position is very accurate Prediction block, but its class probability is lower, and such case is easy for being filtered in post-processing.Therefore the embodiment of the present application exists During anchor and GT box are matched, electronic equipment is stored with above-mentioned preset matching rule, can be used above-mentioned default It is matched, is specifically included with rule:

The friendship of above-mentioned prediction block and above-mentioned anchor point data and ratio are obtained, above-mentioned friendship is determined in above-mentioned anchor point data and ratio is most Big anchor point data are the matched target anchor point of above-mentioned prediction block.

Wherein, hand over and than (Intersection-over-Union, IoU), be used in target detection one it is general Read, be the overlapping rate of the candidate frame (candidate bound) generated and former indicia framing (ground truth bound), i.e., it Intersection and union ratio.Most ideally completely overlapped, i.e., ratio is 1.Using can in the embodiment of the present application With first by several anchor point data of each prediction block corresponding (having intersection), the maximum anchor point data of IoU as target anchor point ( The anchor being fitted on).

Because the IoU of some anchor and some GT box are maximum, that illustrates that this anchor with this GT box is most Matched, then this anchor can be used for the positional shift for being responsible for predicting this GT box.

It should be noted that each anchor can only match a GT box, but each GT box can have several Anchor is predicted.

Optionally, can also be stored in electronic equipment friendship and than threshold value, first determine whether the friendship of anchor point data and prediction block And compare whether greater than the friendship and than threshold value, if more than the friendship and prediction block more above-mentioned than determination in the anchor point data of threshold value is being greater than Matched target anchor point screens out the friendship of part and GT box and relatively low anchor first.Matching process therein can also Think the particular content of the step 205 description.

It is further alternative, default anchor point numerical value M can be set in preset matching rule；

The step is specific can include: obtains the friendship of above-mentioned prediction block and above-mentioned anchor point data and ratio, by above-mentioned friendship and ratio is pressed It is ranked up according to descending sequence, determines in above-mentioned sequence and hand over and be above-mentioned prediction block than corresponding anchor point data for first M Matched target anchor point.

Wherein, the matching strategy in the embodiment of the present application is that N number of prediction interval is put together consideration, matches more than one Target anchor point is to guarantee recall rate.When prediction under truth, there are a anchor point data to predict inclined or do not had It predicts, there are also other anchor point data as candidate, can reduce the case where detection is omitted in this way；And if some prediction block Scale size and position are all not good enough with pre-set anchor point Data Matching, can only be matched to a target anchor point, then M anchor point data are arranged to match this prediction block before IoU can be taken this when to sort.

Above-mentioned default anchor point numerical value M can be stored in advance in electronic equipment, i.e., each matched target anchor point of prediction block Quantity.For example, M=3 when handing over and being 0.35 than threshold value, can choose friendship and be greater than 0.35 as positive sample than threshold value IoU (the anchor point data no more than 0.35 are ignored), if the anchor number that the GT box is matched to sorts less than 3 according to IoU, Maximum 3 anchor of IoU are chosen as the anchor being matched to, i.e., above-mentioned target anchor point can guarantee enough call together in this way The rate of returning.

After determining the matched target anchor point of above-mentioned prediction block, step 206 can be executed.

206, recurrence loss ginseng of the above-mentioned prediction block based on above-mentioned matched target anchor point is obtained according to default weighting algorithm Several and Classification Loss parameter.

In the neural network of deep learning, the invention relates to objective function be divided into two parts: it is corresponding pre- Loss is lost in the position loss and classification confidence level for surveying circle, i.e., respectively above-mentioned recurrence loss (loss) parameter and Classification Loss Parameter.Specifically, the step can include:

Friendship based on above-mentioned prediction block and above-mentioned anchor point data and ratio calculating target weight value, according to above-mentioned target weight value It is calculated using the loss function of weighting and obtains above-mentioned recurrence loss parameter of the prediction block based on above-mentioned matched target anchor point；

It is calculated according to above-mentioned target weight value using the cross entropy loss function of weighting and obtains above-mentioned prediction block based on above-mentioned The Classification Loss parameter of matched target anchor point.

It wherein, can be in default neural network mould after determining above-mentioned prediction block and the matched target anchor point of prediction block The branch's calculating prediction block that returns of type is lost based on the positional shift of target anchor point, i.e., above-mentioned recurrence loss parameter, in classification point The class probability of each target anchor point is calculated in branch, i.e., above-mentioned Classification Loss parameter.

The above-mentioned target weight value of above-mentioned two branch can be obtained all in accordance with prediction block and the IoU size calculating of target anchor point , specifically, target weight value can be (1+IoU), the smooth based on the weighting of above-mentioned target weight value can be used L1loss calculates above-mentioned recurrence loss parameter.

It is exactly traditional intersection loss, the size of IoU is between 0 to 1 here if IoU is equal to 0.IoU is bigger to represent this The recurrence loss parameter and Classification Loss parameter that a anchor is calculated are bigger.

207, the first Overlapping parameters and coordinate of above-mentioned prediction block are obtained, above-mentioned first Overlapping parameters is deleted and is greater than the first threshold The prediction block of value, using the mean value of the coordinate of the prediction block of above-mentioned deletion and the coordinate of the prediction block of reservation as new prediction block Coordinate.

Wherein, step 206 and step 207 can refer to the specific descriptions in the step 103 of embodiment illustrated in fig. 1.

In object detection field, positive sample is exactly task target to be detected, for example, not of the same race in recognition of face The face at race's age, different expressions face, wear different decorations face etc. when face；And negative sample is then target Different background locating for object (this background does not include face), such as face will appear in different environments, street, indoor institute in a word Have and be likely to face occur in thinkable environment, negative sample is exactly the picture that these do not include face.The embodiment of the present application In positive sample can be understood as position existing for pneumonia (pneumonia region).

When predicting deduction (it will appear many prediction blocks including step 205), many at that time is all negative sample, I.e. class prediction is negative sample, then after filtering out such prediction block, remaining is exactly the frame for being predicted as positive sample, as upper Described, a prediction block might have several anchor and go to predict, will appear several prediction blocks during deduction in this way All repeat blocks are to a pneumonia region, but only one true pneumonia frame in fact here, so it is heavy to want method to filter these Multiple frame retains wherein best one prediction block (pneumonia frame) and is used as last prediction result, and removal here is overlapped frame can be with It is determined by above-mentioned first Overlapping parameters (overlap), above-mentioned first Overlapping parameters can be understood as the ginseng of description Duplication Number, in alternative embodiments, above-mentioned first Overlapping parameters may be IoU, also be stored with above-mentioned first in electronic equipment Threshold value, the first Overlapping parameters and above-mentioned first threshold by comparing above-mentioned prediction block, can delete above-mentioned first Overlapping parameters Greater than the prediction block of first threshold, achieve the effect that filter the more serious frame of overlapping phenomenon.

For example, a pneumonia region is gone to predict with 5 prediction blocks, need to remove wherein 4, that can be according to prediction pneumonia Probability sorting illustrate these prediction block overlapping phenomenon ratios if the IoU of maximum probability prediction block and other 4 frames is 0.8 It is more serious, directly that 4 can be frameed shift and removed.

Specifically, the scheme taken in the embodiment of the present application be this 4 frame shift except while, the prediction block of reservation is sat Mark is replaced with the coordinate average value of this 5 frames, more stable prediction block available in this way, then executes step 208.

Optionally, before step 207, this method further include: obtain the class probability of the prediction block；

If the prediction block that above-mentioned first Overlapping parameters are greater than first threshold is two, the method be can specifically include:

It deletes above-mentioned first Overlapping parameters to be greater than in the prediction block of first threshold, the above-mentioned lower prediction block of class probability.

In order to understand the content of image, can be classified with application image (image classification), this is using meter Calculation machine vision and machine learning algorithm extract the task of meaning from image.This operation can be simply an image distribution One label, such as cat, dog or elephant, or can also the advanced content to interpretation of images and return one it is human-readable Sentence.Wherein, above-mentioned class probability can be understood as the specific probability of image classification result, for example, in the embodiment of the present application certain Probability of one prediction block as pneumonia region.Above-mentioned class probability can also be judged based on above-mentioned Classification Loss parameter.

208, the second Overlapping parameters of above-mentioned new prediction block are obtained, above-mentioned second Overlapping parameters is deleted and is less than second threshold Prediction block.

Above-mentioned Overlapping parameters can be understood as the parameter of description Duplication, in alternative embodiments, above-mentioned overlapping ginseng Number or IoU.Electronic equipment can detecte above-mentioned new prediction block, to obtain the IoU of these prediction blocks.In addition, electronics Above-mentioned second threshold is also stored in equipment.

After obtaining new prediction block, because true pneumonia frame is almost without overlapping, it is possible to the step is executed, That is the NMS method of second step removes all overlapping frames with the second threshold (such as 0.05) of a very little, specifically, passing through The second Overlapping parameters and above-mentioned second threshold of more above-mentioned new prediction block can delete above-mentioned second Overlapping parameters less than The prediction block of two threshold values achievees the effect that filtering has the frame of overlapping.

Optionally, after step 208, this method further include: join above-mentioned recurrence loss parameter and above-mentioned Classification Loss Number inputs optimization module with preset ratio, is updated based on parameter of the default learning strategy to above-mentioned prediction interval.

Wherein, electronic equipment may include optimization module, obtain above-mentioned recurrence loss parameter and above-mentioned Classification Loss ginseng After number, it can be exported into above-mentioned optimization module with preset ratio, preset neural network model for this and carry out study update. Specifically, above-mentioned preset ratio can be 1:1, the optimizer selected can be RMSprop, and RMSprop is Geoff Hinton A kind of autoadapted learning rate method proposed.Gradient square Adagrad all before adding up, and RMSprop is only to count Corresponding average value is calculated, therefore can be relieved Adagrad algorithm learning rate and decline very fast problem.Optionally, warm- can be used The learning strategy of up, in preceding 20 frequency of training (epochs), with 0.0001*1.1^x's (wherein which epochs x indicates) Learning rate (learning rate) is learnt, later can in 60,100 and 160epochs, learning rate halves respectively.Its In, above-mentioned learning rate is used for the study schedule of Controlling model.

Optionally, this method can also include: to be marked according to above-mentioned target prediction result to above-mentioned raw image data Note processing, prediction address of the output comprising above-mentioned label treated raw image data.

Above-mentioned target prediction result includes the prediction block (coordinate) finally determined, can be prediction in a particular application Pneumonia frame, it can it is marked on CXR according to the target prediction result and provides body pneumonia position, and generate prediction address, In include marking the CXR image data in pneumonia region, the tight of pneumonia situation can also be optionally marked according to forecast analysis Weight grade, can export above-mentioned prediction address, and the auxiliary of pneumonia diagnosis is carried out as doctor.

Since diagnosis of pneumonia needs radiologist to have the judgement of expert level, however good radiologist A few regions are concentrated on, other regions are difficult to enjoy the diagnosis of expert level, and the death rate of pneumonia is caused constantly to increase.In recent years Come, with the appearance of deep learning algorithm, deep learning algorithm is obtained in fields such as medical image recognition, segmentation and detections Huge success.The Wu Enda team of Stanford University develops a CheXnet algorithm for pneumonia identification problem, which can To be more than or equal to the average level of radiologist by its diagnostic level of the automatic diagnosis of pneumonia of chest x-ray piece.But the algorithm It is a recognizer, can only predicts whether whole CXR has pneumonia, specific pneumonia position can not be obtained.

The embodiment of the present application can be applied to pneumonia detection field, area lower for medical level, can be based on remote Journey cloud platform carries out pneumonia detection to chest x-ray piece, and prediction determines specific pneumonia position, for doctor quick diagnosis provide it is auxiliary Assistant's section, is conducive to improve medical level of the hospital from far-off regions in terms of pneumonia diagnosis；It can also be physical examination in medical center Obtained chest x-ray piece carries out pneumonia prediction, improves diagnostic level, and reduction is failed to pinpoint a disease in diagnosis.

The embodiment of the present application meets the target image number of target component by being converted to raw image data to be processed According to, then based on default neural network model, destination image data is carried out using N number of network output layer as prediction interval multiple dimensioned pre- It surveys, obtains the first prediction result, including at least one prediction block, above-mentioned N is the integer greater than 1；Obtain above-mentioned default mind Through the anchor parameter in network model as the characteristic pattern of N number of network output layer of prediction interval, based on above-mentioned anchor parameter upper It states prediction interval and multi-scale prediction is carried out to above-mentioned destination image data, obtain each pixel on the characteristic pattern of above-mentioned prediction interval The N number of anchor point data generated determine the above-mentioned matched target anchor of prediction block according to preset matching rule in above-mentioned anchor point data Point obtains recurrence loss parameter of the above-mentioned prediction block based on above-mentioned matched target anchor point according to default weighting algorithm and classification is damaged Parameter is lost, the first Overlapping parameters and coordinate of above-mentioned prediction block are obtained, above-mentioned first Overlapping parameters is deleted and is greater than first threshold Prediction block, using the mean value of the coordinate of the prediction block of above-mentioned deletion and the coordinate of the prediction block of reservation as the seat of new prediction block Mark obtains the second Overlapping parameters of above-mentioned new prediction block, deletes the prediction block that above-mentioned second Overlapping parameters are less than second threshold, It can be based on the specific pneumonia position of CXR image prediction, improve pneumonia diagnosis efficiency and accuracy based on CXR.

It is above-mentioned that mainly the scheme of the embodiment of the present application is described from the angle of method side implementation procedure.It is understood that , in order to realize the above functions, it comprises execute the corresponding hardware configuration of each function and/or software mould for electronic equipment Block.Those skilled in the art should be readily appreciated that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure And algorithm steps, the present invention can be realized with the combining form of hardware or hardware and computer software.Some function actually with Hardware or computer software drive the mode of hardware to execute, the specific application and design constraint item depending on technical solution Part.Professional technician can be to specifically realizing described function using distinct methods, but this realization is not It is considered as beyond the scope of this invention.

The embodiment of the present application can carry out the division of functional module according to above method example to electronic equipment, for example, can With each functional module of each function division of correspondence, two or more functions can also be integrated in a processing module In.Above-mentioned integrated module both can take the form of hardware realization, can also be realized in the form of software function module.It needs It is noted that be schematical, only a kind of logical function partition to the division of module in the embodiment of the present application, it is practical real It is current that there may be another division manner.

Referring to Fig. 3, Fig. 3 is the structural schematic diagram of a kind of electronic equipment disclosed in the embodiment of the present application.As shown in figure 3, The electronic equipment 300 includes: to obtain module 310, prediction module 320 and screening module 330, in which:

The acquisition module 310, for obtaining destination image data to be processed；

The prediction module 320, for being based on default neural network model, using N number of network output layer as prediction interval pair The destination image data carries out multi-scale prediction, obtains the first prediction result, first prediction result includes at least one Prediction block, the N are the integer greater than 1；

The screening module 330 obtains target prediction for carrying out substep Screening Treatment at least one described prediction block As a result.

Optionally, above-mentioned electronic equipment 300 further include: image conversion module 340, for by original image number to be processed According to being converted to the destination image data for meeting target component.

Optionally, described image conversion module 340, specifically for the raw image data is converted to default gray value And/or the destination image data of preset image sizes.

Optionally, the prediction module 320 includes first unit 321, second unit 322 and third unit 323, in which:

The first unit 321 is used for:

The second unit 322 is used for, and the prediction block is determined in the anchor point data according to preset matching rule The target anchor point matched；

The third unit 323 is used for, and is obtained the prediction block according to default weighting algorithm and is based on the matched target The recurrence loss parameter and Classification Loss parameter of anchor point, first prediction result include the prediction block and the prediction block Recurrence loss parameter and Classification Loss parameter based on the matched target anchor point.

Optionally, the second unit 322 is specifically used for, and obtains the friendship of the prediction block and the anchor point data and ratio, The friendship is determined in the anchor point data and is the matched target anchor point of the prediction block than maximum anchor point data.

Optionally, the preset matching rule includes default anchor point numerical value M；

The second unit 322 is specifically used for, and obtains the friendship of the prediction block and the anchor point data and ratio, by the friendship And than being ranked up according to descending sequence, determines in the sequence and hand over and be described than corresponding anchor point data for first M The matched target anchor point of prediction block.

Optionally, the screening module 330 is specifically used for, based on two step non-maxima suppression methods to it is described at least one Prediction block carries out substep Screening Treatment, obtains the target prediction result.

Optionally, the screening module 330 is specifically used for:

Optionally, the acquisition module 310 is also used to, and obtains the class probability of the prediction block；

The screening module 330 is specifically used for:

Optionally, the third unit 323 is specifically used for:

Optionally, the electronic equipment 300 further includes optimization module 350, and the optimization module 350 is used for: described will be returned Return loss parameter and the Classification Loss parameter to obtain with preset ratio, and is damaged according to the recurrence loss parameter and the classification Parameter is lost to be updated based on parameter of the default learning strategy to the prediction interval.

Optionally, the default neural network model includes visual geometric group network model.

Optionally, the target prediction result further includes the coordinate of the prediction block；

The electronic equipment 300 further includes reporting modules 360, is used for according to the target prediction result to the original graph As data are marked processing, prediction address of the output comprising the label treated raw image data.

Electronic equipment in embodiment shown in Fig. 3 can execute the part or complete in Fig. 1 and/or embodiment illustrated in fig. 2 Portion's method.

Electronic equipment 300 shown in implementing Fig. 3, the available destination image data to be processed of electronic equipment 300, are based on Default neural network model carries out multi-scale prediction to the destination image data using N number of network output layer as prediction interval, obtains The first prediction result is obtained, first prediction result includes at least one prediction block, and the N is the integer greater than 1, then to described At least one prediction block carries out substep Screening Treatment, obtains target prediction as a result, the specific pneumonia position of CXR image prediction can be based on It sets, improves pneumonia diagnosis efficiency and accuracy based on CXR.

Referring to Fig. 4, Fig. 4 is the structural schematic diagram of another kind electronic equipment disclosed in the embodiment of the present application.Such as

Shown in Fig. 4, which includes processor 401 and memory 402, wherein electronic equipment 400 can be with Including bus 403, processor 401 and memory 402 can be connected with each other by bus 403, and bus 403 can be external components Interconnection standards (Peripheral Component Interconnect, abbreviation PCI) bus or expanding the industrial standard structure (Extended Industry Standard Architecture, abbreviation EISA) bus etc..Bus 403 can be divided into address Bus, data/address bus, control bus etc..Only to be indicated with a thick line in Fig. 4, it is not intended that only one convenient for indicating Bus or a type of bus.Wherein, electronic equipment 400 can also include input-output equipment 404, input-output equipment 404 may include display screen, such as liquid crystal display.Memory 402 is used to store one or more programs comprising instruction；Place Reason device 401 is used to call the part mentioned in the above-mentioned Fig. 1 and Fig. 2 embodiment of the instruction execution being stored in memory 402 or complete Portion's method and step.Above-mentioned processor 401 can correspond to the function of realizing each module in the electronic equipment 300 in Fig. 3.

Implement electronic equipment 400 shown in Fig. 4, the available destination image data to be processed of electronic equipment 400 is based on Default neural network model carries out multi-scale prediction to the destination image data using N number of network output layer as prediction interval, obtains The first prediction result is obtained, first prediction result includes at least one prediction block, and the N is the integer greater than 1, then to described At least one prediction block carries out substep Screening Treatment, obtains target prediction as a result, the specific pneumonia position of CXR image prediction can be based on It sets, improves pneumonia diagnosis efficiency and accuracy based on CXR.

The embodiment of the present application also provides a kind of computer readable storage medium, wherein the computer readable storage medium is deposited Storage is used for the computer program of electronic data interchange, which execute computer as remembered in above method embodiment Some or all of any image processing method of load step.

It should be noted that for the various method embodiments described above, for simple description, therefore, it is stated as a series of Combination of actions, but those skilled in the art should understand that, the present invention is not limited by the sequence of acts described because According to the present invention, some steps may be performed in other sequences or simultaneously.Secondly, those skilled in the art should also know It knows, the embodiments described in the specification are all preferred embodiments, and related actions and modules is not necessarily of the invention It is necessary.

In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, there is no the portion being described in detail in some embodiment Point, reference can be made to the related descriptions of other embodiments.

In several embodiments provided herein, it should be understood that disclosed device, it can be by another way It realizes.For example, the apparatus embodiments described above are merely exemplary, such as the division of the module (or unit), only Only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple module or components can be tied Another system is closed or is desirably integrated into, or some features can be ignored or not executed.Another point, it is shown or discussed Mutual coupling, direct-coupling or communication connection can be through some interfaces, the INDIRECT COUPLING or logical of device or module Letter connection can be electrical or other forms.

The module as illustrated by the separation member may or may not be physically separated, aobvious as module The component shown may or may not be physical module, it can and it is in one place, or may be distributed over multiple On network module.Some or all of the modules therein can be selected to realize the mesh of this embodiment scheme according to the actual needs 's.

It, can also be in addition, each functional module in each embodiment of the present invention can integrate in a processing module It is that modules physically exist alone, can also be integrated in two or more modules in a module.Above-mentioned integrated mould Block both can take the form of hardware realization, can also be realized in the form of software function module.

If the integrated module is realized in the form of software function module and sells or use as independent product When, it can store in a computer-readable access to memory.Based on this understanding, technical solution of the present invention substantially or Person says that all or part of the part that contributes to existing technology or the technical solution can body in the form of software products Reveal and, which is stored in a memory, including some instructions are used so that a computer equipment (can be personal computer, server or network equipment etc.) executes all or part of each embodiment the method for the present invention Step.And memory above-mentioned includes: USB flash disk, read-only memory (Read-Only Memory, ROM), random access memory The various media that can store program code such as (Random Access Memory, RAM), mobile hard disk, magnetic or disk.

Those of ordinary skill in the art will appreciate that all or part of the steps in the various methods of above-described embodiment is can It is completed with instructing relevant hardware by program, which can store in a computer-readable memory, memory It may include: flash disk, read-only memory, random access device, disk or CD etc..

The embodiment of the present application is described in detail above, specific case used herein to the principle of the present invention and Embodiment is expounded, and the above description of the embodiment is only used to help understand the method for the present invention and its core ideas； At the same time, for those skilled in the art can in specific embodiments and applications according to the thought of the present invention There is change place, in conclusion the contents of this specification are not to be construed as limiting the invention.

Claims

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

Obtain destination image data to be processed；

Based on default neural network model, more rulers are carried out to the destination image data using N number of network output layer as prediction interval Degree prediction obtains the first prediction result, and first prediction result includes at least one prediction block, and the N is whole greater than 1 Number；

2. image processing method according to claim 1, which is characterized in that it is described based on default neural network model, with N A network output layer carries out multi-scale prediction to the destination image data as prediction interval, obtains the first prediction result, described After first prediction result includes at least one prediction block, further includes:

It obtains in the default neural network model and joins as the anchor point of the characteristic pattern of N number of network output layer of the prediction interval Number；

Multi-scale prediction is carried out to the destination image data in the prediction interval based on the anchor parameter, obtains the prediction N number of anchor point data that each pixel on the characteristic pattern of layer generates；

Recurrence loss parameter and classification of the prediction block based on the matched target anchor point are obtained according to default weighting algorithm Loss parameter, first prediction result include the prediction block and the prediction block based on the matched target anchor point Return loss parameter and Classification Loss parameter.

3. image processing method according to claim 2, which is characterized in that it is described according to preset matching rule in the anchor Determine that the matched target anchor point of prediction block includes: in point data

The friendship of the prediction block and the anchor point data and ratio are obtained, the friendship is determined in the anchor point data and than maximum Anchor point data are the matched target anchor point of the prediction block.

4. image processing method according to claim 3, which is characterized in that the preset matching rule includes default anchor point Numerical value M, described according to preset matching rule, the determining matched target anchor point of prediction block includes: in the anchor point data

The friendship of the prediction block and the anchor point data and ratio are obtained, by the friendship and than arranging according to descending sequence Sequence determines in the sequence and hands over and be the matched target anchor point of the prediction block than corresponding anchor point data for first M.

5. image processing method according to claim 1-4, which is characterized in that described at least one is pre- to described It surveys frame and carries out substep Screening Treatment, obtaining target prediction result includes:

Substep Screening Treatment is carried out at least one described prediction block based on two step non-maxima suppression methods, obtains the target Prediction result.

6. image processing method according to claim 5, which is characterized in that described to be based on two step non-maxima suppression methods Substep Screening Treatment is carried out at least one described prediction block, obtaining the target prediction result includes:

The first Overlapping parameters and coordinate of the prediction block are obtained, the prediction that first Overlapping parameters are greater than first threshold is deleted Frame, using the mean value of the coordinate of the prediction block of the deletion and the coordinate of the prediction block of reservation as the coordinate of new prediction block；

The second Overlapping parameters of the new prediction block are obtained, the prediction that second Overlapping parameters are less than second threshold is deleted Frame.

7. image processing method according to claim 8, which is characterized in that the method also includes:

Obtain the class probability of the prediction block；

8. a kind of electronic equipment characterized by comprising obtain module, prediction module and screening module, in which:

The acquisition module, for obtaining destination image data to be processed；

The prediction module, for being based on default neural network model, using N number of network output layer as prediction interval to the target Image data carries out multi-scale prediction, obtains the first prediction result, first prediction result includes at least one prediction block, institute Stating N is the integer greater than 1；

The screening module obtains target prediction result for carrying out substep Screening Treatment at least one described prediction block.

9. a kind of electronic equipment, which is characterized in that including processor and memory, the memory is for storing one or more A program, one or more of programs are configured to be executed by the processor, and described program includes for executing such as right It is required that the described in any item methods of 1-7.

10. a kind of computer readable storage medium, which is characterized in that the computer readable storage medium is for storing electron number According to the computer program of exchange, wherein the computer program executes computer as claim 1-7 is described in any item Method.