Summary of the invention
The embodiment of the invention provides a kind of method and systems for detecting unmanned plane silhouette target, at least to solve existing skill
Art there is technical issues that processing accuracy and efficiency cannot when carrying out object detection and recognition to unmanned plane image.
According to an aspect of an embodiment of the present invention, a kind of method for detecting unmanned plane image is provided, which is characterized in that
Whether the difference value of the reference frame and present frame that judge image is more than threshold value, wherein the reference frame is the phase of the present frame
Adjacent previous frame, the image are made of multiple targets;The reference frame is extracted if the difference value is more than threshold value and described is worked as
The respective feature of previous frame;The feature of the reference frame is transmitted to the present frame by light stream network model;It will be described current
Frame is combined according to different default weights as Enhanced feature from the feature passed over from the reference frame, wherein institute
Stating weight is changeless space weight in feature channel;It detects the Enhanced feature and obtains the target detection of the image
With the result of identification and the result of semantic segmentation.
Further, it includes: to the reference that whether the difference value of the reference frame and present frame that judge target, which is more than threshold value,
Frame and the present frame carry out inter-frame difference operation and obtain difference value;Judge whether the difference value is more than threshold value;If described
Difference value is less than threshold value and the testing result of the reference frame is then transmitted to the present frame.
Further, extracting the reference frame and the respective feature of the present frame includes: to be mentioned by feature extraction network
Take the reference frame and the respective feature of the present frame.
Further, it detects the Enhanced feature and obtains the result of the object detection and recognition and the result of semantic segmentation
It include: to detect the Enhanced feature by high-precision detection network to obtain the detection and recognition result and semantic segmentation of the image
As a result, wherein the high-precision detection network is the algorithm that a kind of pair of still image is detected and identified.
Further, it detects the Enhanced feature and obtains the object detection and recognition result and semantic segmentation of the image
As a result include: after inhibited by before and after frames, high confidence level tracking and spatial position correction to the target detection of the image with
Identification and semantic segmentation result carry out the promotion of precision, wherein the before and after frames inhibition is by the detection on whole section of image
As a result for statistical analysis, it sorts to all detection windows according to confidence level, score height and the low window of score is selected, by score
Low window subtracts particular value, for making the window for detecting right and wrong space out;The high confidence level tracking is to utilize
Track algorithm chooses the high target of confidence level as tracking starting point, from the tracking starting point forward backward whole in testing result
A enterprising line trace of image generates pursuit path, causes confidence level to be lower than if tracking the variation of target during tracking
When preset threshold, then stop tracking;Then the high target of confidence level is chosen from remaining target as the tracking starting point, weight
Aforesaid operations are carried out again;If the window is consistent with the window occurred on pursuit path before, directly skip, most
Afterwards, the result of object detection and recognition and the result of semantic segmentation are corrected using the pursuit path;The space bit
It sets to rectify and is exactly based on tracking result the testing result around each position is compared, the position as IOU > 0.5 is mesh
Target final position, IOU are the parts of two regions overlapping divided by the Set-dissection in two regions.
Further, the reference frame is using first frame as the reference frame, then with having synthesized the of Enhanced feature
Two frames replace the reference frame, and so on, or give a time interval tin, first frame is used as reference frame, t firstin
The reference frame of t moment is exactly t-t after momentinThe frame at moment.
Further, the feature of the reference frame is transmitted to the present frame by light stream network model includes: by light
Flow field My→x=F (x, y) carries out estimation by the light stream network model, wherein × it is reference frame, y is present frame;Root
The feature of the reference frame is transmitted to present frame according to light stream propagation function, wherein the light stream propagation function is fx→y=W
(fx, My→x)=W (fx, F (x, y)), wherein W () is a bilinear propagation function, is each led to for being applied to feature
On all positions in road, fx→yIndicate the feature that frame y is traveled to from frame x.
According to another aspect of an embodiment of the present invention, a kind of system for detecting unmanned plane silhouette target is additionally provided, comprising:
Judgment module, for judging whether reference frame and the difference value of present frame of image are more than threshold value, wherein the reference frame is institute
State the adjacent previous frame of present frame;Extraction module, for extracting the reference frame and described if the difference value is more than threshold value
The respective feature of present frame;Transfer module, for the feature of the reference frame to be transmitted to described work as by light stream network model
Previous frame;Feature enhancing module, for presetting the present frame according to different from the feature passed over from the reference frame
Weight is combined as Enhanced feature, wherein the weight is changeless space weight in feature channel;Detect mould
Block, for detect the Enhanced feature obtain the image object detection and recognition result and semantic segmentation result.
Other side according to an embodiment of the present invention has been also provided to a kind of storage medium, has protected on the storage medium
Computer program stored executes the upper method when described program is run.
Other side according to an embodiment of the present invention has been also provided to a kind of processor, has held when described program is run
The above-mentioned method of row.
It in embodiments of the present invention, whether is more than threshold value using the difference value of the reference frame and present frame that judge target,
In, the reference frame is the adjacent previous frame of the present frame;If the difference value be more than threshold value if extract the reference frame and
The respective feature of present frame;The feature of the reference frame is transmitted to the present frame by light stream network model;By institute
Present frame is stated to be combined according to different default weights as Enhanced feature from the feature passed over from the reference frame,
In, the weight is changeless space weight in feature channel;It detects the Enhanced feature and obtains the target detection
With the mode of the result of the result and semantic segmentation of identification, realizes and take into account efficient while guaranteeing precision to complex environment
Lower unmanned aerial vehicle remote sensing images carry out object detection and recognition, and fine profile can be carried out to the target detected and is described, is being mentioned
While rising efficiency to motion blur, block, video defocuses, metamorphosis diversity, illumination variation diversity etc. are no longer sensitive,
With preferable robustness, generalization and real-time, and then solves the prior art and target detection is being carried out to unmanned plane image
There is technical issues that processing accuracy and efficiency cannot when with identification.
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 invention
Attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is only
The embodiment of a part of the invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill people
The model that the present invention protects all should belong in member's every other embodiment obtained without making creative work
It encloses.
It should be noted that term " first " in description and claims of this specification and above-mentioned attached drawing, "
It two " is etc. to be used to distinguish similar objects, without being used to describe a particular order or precedence order.It should be understood that using in this way
Data be interchangeable under appropriate circumstances, so as to the embodiment of the present invention described herein can in addition to illustrating herein or
Sequence other than those of description is implemented.In addition, term " including " and " having " and their any deformation, it is intended that cover
Cover it is non-exclusive include, for example, the process, method, system, product or equipment for containing a series of steps or units are not necessarily limited to
Step or unit those of is clearly listed, but may include be not clearly listed or for these process, methods, product
Or other step or units that equipment is intrinsic.
According to embodiments of the present invention, a kind of embodiment of the method for detecting unmanned plane silhouette target is provided, needs to illustrate
It is that step shown in the flowchart of the accompanying drawings can execute in a computer system such as a set of computer executable instructions,
Also, although logical order is shown in flow charts, and it in some cases, can be to be different from sequence execution herein
Shown or described step.
Fig. 1 is a kind of detection unmanned plane image method according to an embodiment of the present invention, as shown in Figure 1, this method includes such as
Lower step:
Whether the difference value of step S102, the reference frame and present frame that judge image are more than threshold value, wherein reference frame is to work as
The adjacent previous frame of previous frame;
Step S104 extracts reference frame and the respective feature of present frame if difference value is more than threshold value;
The feature of reference frame is transmitted to present frame by light stream network model by step S106;
Present frame is combined into from the feature passed over from reference frame according to different default weights by step S108
For Enhanced feature, wherein weight is changeless space weight in feature channel;
Step S1010, detection Enhanced feature obtain silhouette target detection and the result of identification and the result of semantic segmentation.
The above method can apply to unmanned plane in targets in ocean detection and identification, mine management, ground traffic control management, sea
It is permeated with oil detection, unmanned aerial vehicle remote sensing images Objects extraction, automatic Pilot, medical image processing in administrative area, passes through the above method
The problems such as simply and efficiently carrying out the acquisition of image information, and reducing existing de- frame, poor quality in video, also
It is to reduce the fuzzy equal adverse effect caused by testing result.It can by reference to whether the difference value of frame and present frame is more than threshold value
So that unmanned plane is when being detected, only detection changes big image frame, so that reducing operand accelerates arithmetic speed, and uses
The transmitting that light stream network carries out feature can solve time in video to a certain extent, contextual information makes testing result
At adverse effect, can be taken into account while guaranteeing precision efficient under complex environment unmanned aerial vehicle remote sensing images carry out target
Detection and identification, the above method can also carry out fine profile to the target detected and describe;While raising efficiency pair
Motion blur, block, video defocuses, metamorphosis diversity, illumination variation diversity etc. are no longer sensitive, there is preferable robust
Property, generalization and real-time, to solve the prior art, when being detected and being identified to target image, there are processing accuracies
The technical issues of cannot being taken into account with efficiency.
When above-mentioned unmanned plane carries out the detection of image valid frame, which includes that the valid frame based on difference extracts (spy
Sign extract) and a kind of specific reference frame choose two parts, the purpose of the module is the efficiency in order to promote whole network, makes it
It can reach the requirement of real-time detection in non-top hardware configuration, the selection for particular reference frame can at one
In the embodiment of choosing, i.e., reference frame can be using first frame as reference frame, then with the second frame for having synthesized Enhanced feature
Replace reference frame, and so on, or can be a given time interval tin, first frame is used as reference frame, t firstinMoment
The reference frame of t moment is exactly t-t afterwardsinThe frame at moment.
For the extracting method of the valid frame based on difference, in a kind of optional embodiment, i.e., to reference frame with work as
Previous frame carries out inter-frame difference operation and obtains difference value;Judge whether difference value is more than threshold value;If difference value is less than threshold value
The testing result of reference frame is transmitted to present frame.
Feature extraction can be passed through in an optional embodiment by extracting reference frame and the respective feature of present frame
Network extracts reference frame and the respective feature of present frame.It can be ResNet101+FPN that this feature, which extracts network,.
After the feature for extracting reference frame and present frame in through the above steps, worked as using feature (information) enhancing of reference frame
The feature of previous frame, and object detection and recognition is carried out using advanced high-precision detection network, while the tool of object can be obtained
Body profile, also referred to as semantic segmentation, so that detection reaches higher precision.
Below by one optionally embodiment the detection of above-mentioned unmanned plane image valid frame is illustrated:
The effective frame detector of unmanned plane image is to compare reference frame and present frame, judges whether there is biggish difference
It is different.It can choose or carry out the last time the target detection of Enhanced feature with Fixed Time Interval when choosing reference frame
Frame with identification is as reference frame.Unmanned plane image is input in the effective frame detector of unmanned plane image, is detected in valid frame
First frame is judged whether it is in device, the object detection and recognition module that Enhanced feature is directly sent to if it is first frame carries out
Processing, obtained result directly export.It is effective that it is input to unmanned plane image together using first frame as reference frame x and subsequent frame y
(x-y) is calculated in frame detector, whether is subjected to by the difference of preset two interframe of threshold decision, if two interframe
Difference value is less than preset threshold, and the object detection and recognition module that present frame does not need incoming Enhanced feature is handled, only
The testing result of reference frame need to be passed directly to present frame;If the difference value of two interframe is more than preset threshold, present frame is passed
Enter the object detection and recognition module of Enhanced feature.The treatment process of the effective frame detector of unmanned plane image is as shown in Figure 2.Fig. 3
The necessity for carrying out valid frame detection is illustrated using the example of ocean, Fig. 3 is the image frame at certain moment, and Fig. 4 is subsequent time
Image frame, Fig. 5 is certain moment adjacent next reference frame.It is clear that Fig. 3 and Fig. 4 is without too big variation, ship in image
The mobile distance of body can almost be ignored, and many meters will be wasted by this moment carrying out object detection and recognition to Fig. 3 and Fig. 4 respectively again
It calculates resource and Fig. 5 is passed to the mesh of Enhanced feature so the object detection results of Fig. 3 are just traveled to Fig. 4 by the embodiment of the present invention
Mark detection is detected with identification.
In a kind of optional embodiment, detection Enhanced feature obtains the result and semantic segmentation of object detection and recognition
Result can detect network detection by high-precision and carry out, wherein high-precision detection network is that a still image carries out
The algorithm of detection and identification, but if be applied directly in video will due to do not have using in video time, on
Hereafter etc. information cause detection effect bad.And light stream network is in order to avoid the shadow caused by detection such as motion blur in video
It rings, by characteristic aggregation of the before and after frames on motion path into the feature of present frame, to enhance the feature of present frame, improves mesh
The precision of mark detection and identification.Above-mentioned high-precision detection network is two stage object detection and recognition method, first stage
It is called region and suggests network (Region Proposal Network, RPN, region proposes that network or region suggest network), uses
To extract the bounding box of candidate target;Second stage using in candidate frame point of interest Chi Hualai extract feature then into
Row classification and bounding box return, while generating a binary mask for each point of interest.Mask is by the space layout of an object
It is encoded, unlike class label or frame, high-precision detection network can be aligned to make by the pixel of convolution
Space structure is extracted with mask.
High-precision object detection and recognition network is the structure of Mask R-CNN, is carrying out the same of object detection and recognition
When there are a parallel branch, which can carry out semantic segmentation to target while carrying out object detection and recognition, this
The whole series are also referred to as example segmentation.
In an optional embodiment, the feature of reference frame can be transmitted to by present frame by light stream network model
It include: by optical flow field My→x=F (x, y) carries out estimation by light stream network model, wherein x is reference frame, and y is current
Frame;The feature of reference frame is transmitted to present frame according to light stream propagation function, wherein light stream propagation function is fx→y=W (fx,
My→x)=W (fx, F (x, y)), wherein W () is a bilinear propagation function, for being applied to each channel of feature
On all positions, fx→yIndicate the feature that frame y is traveled to from frame x.
As shown in fig. 6, the feature of the information enhancement present frame using adjacent preceding reference frame, then enhanced feature is inputted
Detection network to the end carries out object detection and recognition, uses feature extraction network to extract respectively in present frame and reference frame first
Respective feature;In order to enhance the feature of present frame, adjacent preceding reference frame and current is estimated using a kind of specific light flow network
Then the feature of adjacent preceding reference frame is traveled to present frame according to light stream campaign by the movement of frame;Later by adjacent preceding reference frame
Feature and the feature of present frame pass through preset weighting network together and be combined;The feature that finally combination is obtained passes
It is sent in a kind of advanced high-precision detection network and object detection and recognition and semantic segmentation is carried out to present frame.
The above process is illustrated below by an optional embodiment:
The feature for enhancing present frame by light stream network, for reference frame x and present frame y, optical flow field My→x=F (x, y) is logical
It is network-evaluated to cross a kind of specific light stream;The feature of reference frame is transmitted to present frame, can be defined according to light stream propagation function
For fx→y=W (fx, My→x)=W (fx, F (x, y)), wherein W () is a bilinear propagation function, for being applied to feature
On all positions in each channel, fx→yIndicate the feature that frame y is traveled to from frame x.The spy that present frame will be passed over from reference frame
Sign is combined, so that illumination, posture, visual angle, non-rigid deformation etc. be avoided to influence.The present embodiment is in difference when combination
The spatial position space weight taking different weights, and all feature channels shares is allowed to set, after combination feature is
f′y=w1fx→y+w2fy, w1And w2Indicate the significance level on each spatial position.Then the feature after combination is sent to below
Detector in.
After detection Enhanced feature obtains the result of object detection and recognition and the result of semantic segmentation, a kind of optional
It, can be by before and after frames inhibition, high confidence level tracking and spatial position correction to the knot of object detection and recognition in embodiment
The result of fruit and semantic segmentation further promotes precision, wherein before and after frames inhibition is by the detection knot on whole section of image
Fruit is for statistical analysis, sorts to all detection windows according to confidence level, selects score height and the low window of score, and score is low
Window subtract particular value, for make detect right and wrong window space out;High confidence level tracking is calculated using tracking
Method chooses the high target of confidence level as tracking starting point in testing result, and from tracking, starting point is enterprising in entire image backward forward
Line trace generates pursuit path, if the variation for tracking target during tracking leads to confidence level lower than preset threshold,
Then stop tracking;Then the high target of confidence level is chosen from remaining target as tracking starting point, repeats aforesaid operations;
If window is consistent with the window occurred on pursuit path before, directly skip, finally, using pursuit path to detection
Recognition result and meaning of one's words segmentation result are corrected;It rectifys and is exactly based on tracking result to the inspection around each position in spatial position
Result is surveyed to compare, the position as IOU > 0.5 is the final position of target, IOU be the overlapping of two regions part divided by
The Set-dissection in two regions.
Above-mentioned before and after frames inhibit, high confidence level tracking and spatial position can mutual any combination, repaired by result
Just, can further reduce motion blur, block, video defocuses, metamorphosis diversity, illumination variation diversity etc. are to knot
The influence of fruit.
Due to the problem of there may be erroneous detections in a small amount of frame of image, the information of single frames can not distinguish these erroneous detections,
Most of testing result has balloon score in each frame of video, so erroneous detection can be considered as exceptional value by the present embodiment, then
Inhibit their confidence level.It is for statistical analysis to the testing result on whole section of image, to all detection windows according to confidence level
Sequence, selects that score is higher and the lower window of score, is likely to erroneous detection to the lower window of score, it is specific to subtract some
Value improves the precision of target detection so that the confidence level of detection right and wrong spaces out.
The problem of reducing missing inspection in such a way that before and after frames feature enhances, but missing inspection effect continuous for multiframe is not
Good, so the present embodiment carries out high confidence level tracking processing to testing result, the present embodiment chooses the target of detection highest scoring
It as the starting point of tracking, is tracked in entire image video backward forward based on this starting point, generates pursuit path;
Then highest scoring is selected track from remaining target, it should be noted that if this window is in pursuit path before
In occurred, then directly skipping, next target is selected to be tracked;Algorithm iteration executes, and uses score threshold as eventually
Only condition.Target recall rate can be improved in obtained pursuit path, and certain amendment is carried out to testing result.Finally, using with
The target that track result detects each position around it compares, and when IOU > 0.5 is considered as the final position of target.
After the combination of a variety of modified result modes, finally obtained ratio of precision only has in the result that object detection and recognition obtains
Certain promotion can reduce the calculation amount of object detection and recognition, to reduce calculating cost under the premise of guaranteeing precision;
Simultaneously can reduce in image due to motion blur, block, metamorphosis diversity, illumination variation diversity the problems such as bring
Precision influences, and has sufficiently excavated the potential information (such as temporal information and front and back frame information) in image.
Above-mentioned entire identification process is illustrated with an optional embodiment below with reference to Fig. 7:
Using the feature of reference frame frame information enhancing present frame, target detection is carried out using advanced high-precision detection network
With identification, simultaneously as the embodiment of the present invention, which detects network by advanced high-precision, can obtain the specific profile of target;Fusion
The modified result (before and after frames inhibit, high confidence level tracks and the multiple combinations of spatial position correction) of method has sufficiently excavated image
In useful information, and then object detection and recognition result is modified, through the invention can be in the premise for guaranteeing precision
The lower calculation amount for reducing object detection and recognition, to reduce calculating cost;Simultaneously can reduce in image due to motion blur,
Block, metamorphosis diversity, illumination variation diversity the problems such as bring precision influence, sufficiently excavated potential in image
Information (such as temporal information and front and back frame information).
Calculus of differences is carried out to two frames by inter-frame difference operation, the corresponding pixel of different frame subtracts each other, and judges gray scale difference
Square value, i.e., calculating (x-y)2, wherein x and y respectively indicate reference frame and present frame, by given threshold to calculated result into
The object detection and recognition module that present frame is passed to Enhanced feature is carried out mesh if present frame is more than given threshold by row judgement
Mark detection and identification;Otherwise, the testing result of reference frame is transmitted to present frame.Wherein selection there are two types of reference frames, one is
The first frame of the object detection and recognition of Enhanced feature will be carried out first as reference frame, then with the target for carrying out Enhanced feature
Second frame of detection and identification replaces with reference frame, and so on;Another kind is to give a time interval tin, first frame is first
As reference frame, tinThe reference frame of t moment is exactly t-t after momentinThe frame at moment.
In the video of unmanned plane shooting, there are many approximate fixed scenes, such as the remote sensing shadow shot across the sea
Picture, may only have sea in tens seconds, without any object, and at this time carrying out object detection and recognition frame by frame again will waste
Very big computing resource, the embodiment of the present invention lose interest in sea, it is of interest that various objects present on sea.
So the embodiment of the present invention, which chooses effective frame using the effective frame detector of unmanned plane image, carries out object detection and recognition, it is right
Testing result is transmitted by adjacent preceding reference frame in non-effective frame.The purpose of the effective frame detector of unmanned plane image is exactly to accelerate
The speed of image processing.Image has difference and little between temporal locality, that is, two frames of front and back, so if according to biography
The method of system detected frame by frame using convolutional neural networks will the inefficiency of detection and seem to be not necessarily to.
The difference operation principle of two interframe used in the embodiment of the present invention is simple, and calculation amount is small, can quickly detect the fortune in scene
Moving-target.
Using the feature of the information enhancement present frame of adjacent preceding reference frame, then the inspection by the input of enhanced feature to the end
Survey grid network carries out object detection and recognition, specifically, uses feature extraction network to extract respectively in present frame and reference frame first each
From feature;In order to enhance the feature of present frame, adjacent preceding reference frame and present frame are estimated using a kind of specific light flow network
Movement, the feature of adjacent preceding reference frame is then traveled into present frame according to light stream campaign, light stream network is exactly one given
Reference frame x and present frame y, optical flow field My→x=F (x, y) is network-evaluated by a kind of specific light stream;The feature of reference frame is passed
It is multicast to present frame, f can be defined as according to light stream propagation functionx→y=W (fx, My→x)=W (fx, F (x, y)), wherein W () is
One bilinear propagation function, for being applied to all positions in each channel of feature, fx→yIt indicates to travel to frame y from frame x
Feature;The feature of adjacent preceding reference frame and the feature of present frame pass through preset weighting network together and are combined;It will
It combines in the detection network of obtained feature transmission to the end and object detection and recognition is carried out to present frame, simultaneously as this hair
The advanced high-precision that bright embodiment uses detects the characteristics of network, and the embodiment of the present invention can obtain the specific wheel of target simultaneously
It is wide.
The method of object detection and recognition can be divided into two classes: by positioning, identify that the stage as recurrence task rolls up entirely
Product method and will test positioning, target identification is divided into two stage method, the method precision in stage is relatively high, but speed ratio
It is relatively slow, by the effective frame detector of unmanned plane image that the embodiment of the present invention designs, the speed of two stage object detection and recognition
Degree greatly promotes.Object in video will receive motion blur, block, video defocuses, metamorphosis diversity, illumination variation are more
The influence of sample etc., so the information of adjacent previous frame is transmitted to currently by the embodiment of the present invention using a kind of specific light stream network
Frame carries out feature enhancing, then puts into enhanced feature and carries out Objective extraction in the network of object detection and recognition.
The modified result of fusion method uses the fusion of various ways, including before and after frames inhibit, high confidence level tracking, space
Aligning.Before and after frames inhibit by for statistical analysis to the testing result on whole section of image, to all detection windows according to
Confidence level sequence, selects that score is higher and the lower window of score, subtracts some specific value to the lower window of score, thus
So that the window of detection right and wrong spaces out.High confidence level tracking is to choose to set in testing result using track algorithm
Starting point of the highest target of reliability as tracking generates tracking from this starting point forward backward in the enterprising line trace of entire image
Track, during tracking if when the variation of tracking target leads to confidence level lower than a certain threshold value (such as 0.1), stop with
Track;Then starting point of the highest target of confidence level as tracking is chosen from remaining target, repeats aforesaid operations;If
Certain window occurred on track before in subsequent tracking, directly skipped.Finally, using these pursuit paths to result into
Row correction.Spatial position correction is compared using the target that tracking result detects each position around it, IOU > 0.5
When be considered as target final position.After the combination of a variety of modified result modes, finally obtained ratio of precision is only in mesh
The result that mark detection is obtained with identification has certain promotion.
The embodiment of the invention also provides a kind of detection unmanned plane silhouette target system, which can pass through judgment module
82, extraction module 84, transmitting mould 86, feature enhancing module 88, detection module 810 realize its function.It should be noted that this hair
A kind of map generation system of bright embodiment can be used for executing a kind of detection unmanned plane image provided by the embodiment of the present invention
The method of goal approach, a kind of detection unmanned plane silhouette target of the embodiment of the present invention can also be mentioned through the embodiment of the present invention
A kind of detection unmanned plane silhouette target system supplied executes.Fig. 8 is a kind of detection unmanned plane shadow according to an embodiment of the present invention
As the schematic diagram of the system of target.As shown in figure 8, a kind of system for detecting unmanned plane silhouette target includes: a kind of to detect nobody
Machine silhouette target system characterized by comprising
Judgment module 82, for judging whether reference frame and the difference value of present frame of target are more than threshold value, wherein reference
Frame is the adjacent previous frame of present frame;
Extraction module 84, for extracting reference frame and the respective feature of present frame if difference value is more than threshold value;
Transfer module 86, for the feature of reference frame to be transmitted to present frame by light stream network model;
Feature enhancing module 88, for by present frame from the feature passed over from reference frame according to different default weights
Being combined becomes Enhanced feature, wherein weight is changeless space weight in feature channel;
Detection module 810 obtains the result of object detection and recognition and the result of semantic segmentation for detecting Enhanced feature.
It is above-mentioned it is a kind of detection unmanned plane silhouette target system embodiment be and it is a kind of detect unmanned plane silhouette target method
It is corresponding, so being repeated no more for beneficial effect.
The embodiment of the invention provides a kind of storage medium, storage medium includes the program of storage, wherein is run in program
When control storage medium where equipment execute the above method.
The embodiment of the invention provides a kind of processor, processor includes the program of processing, wherein runs time control in program
Equipment executes the above method where processor processed.
The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.
In the above embodiment of the invention, it all emphasizes particularly on different fields to the description of each embodiment, does not have in some embodiment
The part of detailed description, reference can be made to the related descriptions of other embodiments.
In several embodiments provided herein, it should be understood that disclosed technology contents can pass through others
Mode is realized.Wherein, the apparatus embodiments described above are merely exemplary, such as the division of the unit, Ke Yiwei
A kind of logical function partition, there may be another division manner in actual implementation, for example, multiple units or components can combine or
Person is desirably integrated into another system, or some features can be ignored or not executed.Another point, shown or discussed is mutual
Between coupling, direct-coupling or communication connection can be through some interfaces, the INDIRECT COUPLING or communication link of unit or module
It connects, can be electrical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
On unit.It can some or all of the units may be selected to achieve the purpose of the solution of this embodiment according to the actual needs.
It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product
When, it can store in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially
The all or part of the part that contributes to existing technology or the technical solution can be in the form of software products in other words
It embodies, which is stored in a storage medium, including some instructions are used so that a computer
Equipment (can for personal computer, server or network equipment etc.) execute each embodiment the method for the present invention whole or
Part steps.And storage medium above-mentioned includes: that USB flash disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited
Reservoir (RAM, Random Access Memory), mobile hard disk, magnetic or disk etc. be various to can store program code
Medium.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered
It is considered as protection scope of the present invention.