CN110503095A - Alignment quality evaluation method, localization method and the equipment of target detection model - Google Patents
Alignment quality evaluation method, localization method and the equipment of target detection model Download PDFInfo
- Publication number
- CN110503095A CN110503095A CN201910794302.3A CN201910794302A CN110503095A CN 110503095 A CN110503095 A CN 110503095A CN 201910794302 A CN201910794302 A CN 201910794302A CN 110503095 A CN110503095 A CN 110503095A
- Authority
- CN
- China
- Prior art keywords
- friendship
- frame
- correction value
- relative position
- target detection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides alignment quality evaluation method, localization method and the equipment of a kind of target detection model, by calculating the prediction block and the corresponding true friendship of frame and the catercorner length of ratio, centre distance and minimum vertex-covering frame, relative position parameter is calculated using the catercorner length of centre distance and minimum vertex-covering frame, and it is handed over and is compared with relative position parameters revision, it can be on the basis of reflection compares object intersection area, the distance between reflection comparison object relationship, it can more accurately reflect crossing condition, promote positioning accuracy.Target detection and localization is further extended, can effectively promote the sensitivity and accuracy rate of detection positioning for the scene of more detection heavy dense targets distributions.
Description
Technical field
The invention belongs to technical field of computer vision more particularly to a kind of alignment quality evaluation sides of target detection model
Method, localization method and equipment.
Background technique
It hands over and than IoU (Intersection over Union), also referred to as Jaccard index, is for comparing two timess
The most frequently used measurement of similitude between shape of anticipating.IoU by the shape attribute of comparison other (such as the width of two bounding boxes,
Height and position) it is encoded to area attribute, then calculate the normalization measurement for paying close attention to its area (or volume).This area
Attribute influences IoU measurement by target scale size.Just because of this characteristic, computer vision field is for assessing
The all properties measurement of the tasks such as Target Segmentation, target following and target detection all relies on the measurement.
Target detection is different from other Computer Vision Tasks, only focuses in two tasks of classification and positioning.From target
From the point of view of detection field, the following promotion that will more focus on target positioning aspect.And from the subjective visual evaluation of people for, I
To object positioning requirement it is very stringent.
Existing target detection evaluation index is substantially measured friendship and compares, but hand over and cannot than index IoU
Actually reflect that human eye positions the measurement for returning accuracy to target well, the IoU of standard is that 0.5 this threshold value is considered marking
It is quasi- excessively loose, and too high IoU threshold value can make algorithm study encounter because of the ambiguity tagging or error label of data set
Bottleneck.The problem of choosing in addition to threshold value, there is also some fatal defects for IoU measurement, if two objects are not overlapped, IoU value
It will be zero, and can not positional relationship between reversed two objects.Simultaneously.More precisely, two objects are multiple and different
It is overlapped on direction, and crosspoint level is identical, IoU will be essentially equal.Therefore, the value of IoU function does not reflect two objects
Between how to overlap.It is therefore proposed that with human eye subjective feeling evaluation be more consistent IoU measurement very it is necessary to.
Summary of the invention
In view of the deficiencies of the prior art, it is an object of the present invention to provide candidates in a kind of computer vision model training to confine
The method of position overcomes the defect handed over and cannot reflect comparison object's position relationship than IoU by amendment friendship and the value than IoU, from
And promote precision of the computer vision in target positioning training.
The technical solution that the present invention solves the problems, such as is:
On the one hand, a kind of alignment quality evaluation method of target detection model is provided, comprising:
Target to be detected is positioned using target detection model, obtains corresponding prediction block；
Calculate the prediction block and the corresponding true friendship of frame and the diagonal line length of ratio, centre distance and minimum vertex-covering frame
Degree；
Calculate the relative position parameter of the centre distance and the corresponding minimum vertex-covering frame catercorner length；
Using the relative position parameters revision friendship accordingly and compares, handed over and compare correction value；
The alignment quality of the setting target detection model is evaluated according to the friendship and than correction value.
In some embodiments, it is opposite with the corresponding minimum vertex-covering frame catercorner length to calculate the centre distance
Location parameter, comprising:
The quotient for calculating the centre distance and the minimum vertex-covering frame catercorner length obtains the relative position parameter.
In some embodiments, using the relative position parameters revision friendship accordingly and compare, handed over and ratio is repaired
Positive value, comprising:
Calculate it is described friendship and compare the difference with the relative position parameter, obtain it is described hand over and compare correction value.
In some embodiments, the positioning matter of the setting target detection model is evaluated according to the friendship and than correction value
Amount, comprising:
By the friendship and the prediction block than correction value greater than the first setting value is judged as that positioning result is correct.
On the other hand, also a kind of object localization method of the present invention, comprising:
The multiple prediction blocks for obtaining multiple targets to be detected respectively by target detection and localization operation and corresponding classification are generally
Rate value；
Calculate the catercorner length of the friendship between the prediction block and ratio, centre distance and minimum vertex-covering frame；
Calculate the relative position parameter of the centre distance and the corresponding minimum vertex-covering frame catercorner length；
Using the relative position parameters revision friendship accordingly and compares, handed over and compare correction value；
Using the friendship and than correction value as the optimum prediction frame of each target to be detected of standard screening.
In some embodiments, it is opposite with the corresponding minimum vertex-covering frame catercorner length to calculate the centre distance
Location parameter, comprising:
The quotient for calculating the centre distance and the minimum vertex-covering frame catercorner length obtains the relative position parameter.
In some embodiments, using the relative position parameters revision friendship accordingly and compare, handed over and ratio is repaired
Positive value, comprising:
Calculate it is described friendship and compare the difference with the relative position parameter, obtain it is described hand over and compare correction value.
In some embodiments, using the friendship and than correction value as the optimum prediction of target to be detected described in standard screening
Frame, comprising:
Using the friendship and the optimum prediction frame is filtered out as standard with non-maxima suppression iteration than correction value.
On the other hand, the application also provides a kind of computer equipment, including memory, processor and storage are on a memory
And the computer program that can be run on a processor, when the processor executes described program the step of the realization above method.
On the other hand, the application also provides a kind of computer readable storage medium, is stored thereon with computer program, the journey
The step of above method is realized when sequence is executed by processor.
The beneficial effects of the present invention are:
The alignment quality evaluation method of target detection model of the present invention, by using the candidate frame and described true
The catercorner length amendment friendship of the minimum vertex-covering frame of the centre distance of frame and the candidate frame and the true frame is simultaneously
Than, it can be on the basis of reflection compares object intersection area, further reflection compares the distance between object relationship, Neng Gougeng
Accurately reflect crossing condition, promotes positioning accuracy.Object localization method of the present invention, by using the friendship and than correcting
Value is used as parameter, substitutes friendship in the prior art and compares, when carrying out screening optimum prediction frame by maximum inhibition, Neng Gouti
Rise susceptibility.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.In the accompanying drawings:
Fig. 1 is the center aligned condition position view of candidate frame and true frame in an example of the invention；
Fig. 2 is position view under candidate frame off-centring state in Fig. 1；
Fig. 3 is in an example of the invention comprising two frame lower right corner aligned position schematic diagrames under state；
Fig. 4 is to connect in an of the invention example comprising long side on the right side of two frames under state and axis aligned position schematic diagram；
Fig. 5 is in an example of the invention comprising two frame center aligned position schematic diagrames under state；
Fig. 6 is in another example of the present invention comprising two frame lower right corner aligned position schematic diagrames under state；
Fig. 7 is two frame long side overlapping positions schematic diagrames under part coincidence status in an example of the invention；
Fig. 8 is two frames diagonally overlapping positions schematic diagram under the coincidence status of part in an of the invention example；
Fig. 9 is not two frames diagonally connecting position schematic diagram under coincidence status in an of the invention example；
Figure 10 is the not alignment of two frame bottoms and long side connecting position schematic diagram under coincidence status in an of the invention example；
Figure 11 is that long side does not connect and axisymmetric position schematic diagram on the right side of two frames under coincidence status in an of the invention example；
Figure 12 is that broadside does not connect and axisymmetric position schematic diagram on the right side of two frames under coincidence status in an of the invention example；
Figure 13 is the flow diagram of the alignment quality evaluation method of target detection model in one embodiment of the invention；
Figure 14 is the flow diagram of object localization method in one embodiment of the invention.
Specific embodiment
Understand in order to make the object, technical scheme and advantages of the embodiment of the invention clearer, with reference to the accompanying drawing to this hair
Bright embodiment is described in further details.Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but simultaneously
It is not as a limitation of the invention.
In the prior art, in the evaluation procedure of object detection field target alignment quality, mainly by calculating network mould
Friendship and the ratio of prediction block and true frame corresponding with true value that type operation generates, and prediction block is reflected by the size of friendship and ratio
Alignment quality, when handing over and than closer to 1, the alignment quality of prediction block is better.
Area is handed over and is the default assessment measurement for being currently used for the positioning of object detection field target than IoU, it is for assert
With the real property and false positive in one group of prediction of differentiation.Use hand over and than IoU as evaluation index when, it is necessary to selection accurately
Metric threshold.For example, in PASCAL VOC challenge, well-known detection accuracy measured value (i.e. mean value mean accuracy mAP)
It is to be calculated based on fixed IoU threshold value 0.5.But arbitrarily selection IoU threshold value can not reflect determining for algorithms of different completely
Position performance, any positioning accuracy higher than the threshold value are all treated substantially equally.Therefore many that subjective sense visually is positioned to target
Quite unreasonable high mAP can be obtained on VOC data set by the very big algorithm of difference.Although later VOC data set will
IoU threshold value is promoted to 0.75, but essential problem is not still addressed.
In order to keep target detection performance measurement less sensitive to the selection of IoU threshold value, the evaluation of MS COCO benchmark test collection
Average mAP is calculated in multiple IoU threshold values, such as takes a threshold value, each threshold value at interval of 0.05 between 0.5 to 0.95
A mAP index is calculated, finally 10 values are taken average as final AP.Although so operation enables to existing first gradient
The AP value of algorithm is both less than 0.5, seems so that very high room for promotion has been located to targeting accuracy in algorithm, but due to advising greatly
Modulus is according to a large amount of error labels present in collection and a large amount of ambiguity taggings that can not be solved, so that algorithm of target detection develops
Bottleneck is reached.
In the prior art, a kind of GIoU (extensive wide of Generalized Intersection over Union is further related to
Justice, which is handed over, simultaneously to be compared), it by introduce minimum vertex-covering rectangle by the concept of IoU extend it is extensive to nonoverlapping situation come solve IoU without
Method pays close attention to the weakness of not intersection area, and its calculation formula is GIoU=IoU- (C-U)/C, and wherein IoU is to hand over and compare, and C is
The area of minimum vertex-covering rectangle, U are the area of two frames simultaneously.It hands over and cannot reflect how two objects are overlapped than IoU, at two
Object frame does not have that distance (value range is 0 to 1) also cannot be distinguished in the case where intersection, and GIoU can be concerned about IoU institute
The size for the not intersection being not concerned with, but for frame return quality evaluation be not still it is exhaustive, only
Area attribute is paid close attention to, and the four dimensions that frame returns out are (x, y, w, h) or (x1,y1,x2,y2), it is clear that only with area category
Property indicates it is a kind of Proxy Signature Scheme, and the expression of location information is damaged.It concentrates in large-scale data, can still reach quickly
To the precision upper limit.
Therefore, on the basis of existing technology, as shown in figure 13, the present invention provides a kind of positioning of target detection model
Quality evaluating method, comprising:
Step 101: target to be detected being positioned using target detection model, obtains corresponding prediction block；
Step 102: calculating prediction block and the corresponding true friendship of frame and the diagonal line of ratio, centre distance and minimum vertex-covering frame
Length；
Step 103: calculating the relative position ginseng of the centre distance and the corresponding minimum vertex-covering frame catercorner length
Number；
Step 104: being handed over and compared accordingly using relative position parameters revision, handed over and compare correction value；
Step 105: according to the alignment quality for handing over and evaluating than correction value the setting target detection model.
In above-mentioned steps 101, which can be the journey obtained according to various existing object detection methods
Sequence module or functional module.The target to be detected can be specified content to be detected (such as face or animal in image
Face), can only have a content in an image, i.e. the specified content, alternatively, may include the specified content in an image
Non-targeted content with other.It can detecte the position of target to be detected in a figure using the target detection model, it can be with
The form of prediction block shows the target position of the model inspection, and the locating accuracy of different target detection models may
It is different, so can be evaluated the positioning performance of the target detection model using subsequent step.
In above-mentioned steps 102, the friendship and than that can be calculated by existing calculation, i.e. IoU.Centre distance can
To refer to the distance between the geometric center of two frames, for example, this two in the case where prediction block and true frame are rectangle frame
Centre distance between a frame can with the diagonal line intersection point of the true frame of the diagonal line intersection point and rectangle of the prediction block of rectangle it
Between distance.The minimum rectangle frame of minimum vertex-covering frame coverage prediction frame and true frame, in the case, catercorner length can be with
For the cornerwise length of any bar.
In the present embodiment, it hands over and on the basis of compare calculating, is further advanced by prediction block and corresponding really frame
The amendment of the catercorner length of centre distance and minimum vertex-covering frame is handed over and is compared, can be Chong Die with true frame area in reflection prediction block
On the basis of relationship, further reflects the relativeness of prediction block and true frame center, improve the sensitivity of threshold ratings
Degree.
In some embodiments, above-mentioned steps 103, that is, calculate centre distance and corresponding minimum vertex-covering frame diagonal line length
The relative position parameter of degree, more specifically, it may include step:
1031, the quotient of centre distance and minimum vertex-covering frame catercorner length is calculated, relative position parameter is obtained.
In the present embodiment, in order to further reflect the relative positional relationship of prediction block and true frame in friendship and ratio,
The susceptibility of promotion threshold value evaluation；Wherein, it is poor to represent physical location for centre distance, and minimum vertex-covering frame catercorner length represents prediction
Maximum distance within the scope of frame and true frame is joined the quotient of centre distance and minimum vertex-covering frame catercorner length as relative position
Number, can more effectively reflect the relative positional relationship of prediction block Yu true frame.
Further, relative position parameter is bigger, then it represents that prediction block is remoter with the relative position of true frame, registration
It is poorer；Relative position parameter is bigger, then it represents that prediction block is closer with the relative position of true frame, and registration is better；Therefore, phase
It is negatively correlated to the value and registration of location parameter.
In other embodiments, the relative position parameter can be calculated using other modes, for example, can centered on distance
Increase a coefficient with the quotient of minimum vertex-covering frame catercorner length；Alternatively, can be by centre distance and minimum vertex-covering frame diagonal line
It is poor that length is made, then with resulting difference divided by centre distance, or divided by minimum vertex-covering frame catercorner length.
In some embodiments, above-mentioned steps 104, that is, handed over and compared accordingly using relative position parameters revision, handed over
And than correction value, more specifically, it may include step:
1041, the difference of friendship and ratio and relative position parameter is calculated, is handed over and compares correction value.Correction value can be to hand over and compare
With the absolute value of the difference of relative position parameter.
In the present embodiment, the degree handed over and overlapped than being able to reflect prediction block with true frame area, but cannot be into one
Step reflection relative positional relationship；Registration due to handing over and than reflecting prediction block and true frame by the overlapping degree of area, when
Hand over and than it is bigger when, indicate overlapping part it is bigger, then registration is better；When hand over and than it is smaller when, expression overlapping part it is smaller, then
Registration is poorer；Therefore, it hands over and the value of ratio is positively correlated with registration.
On this basis, calculate hand over and than and relative position parameter difference as hand over and than correction value, can unify with again
Right correlation, the registration of accurate evaluation prediction block and true frame, and then the alignment quality of evaluation and foreca frame.Specifically,
Can according to hand over and than correction value whether setting threshold range (can rule of thumb or the precise requirements of positioning come it is true
The quality of alignment quality is judged in calmly)；Alternatively, can compare, different target detection model is corresponding to be handed over and sentences than correction value
Disconnected model determines the relatively fine or not situation of quality.
In the present embodiment, friendship can be directly used and than the alignment quality of correction value evaluation and foreca frame, that is, hands over and ratio is repaired
Positive value is bigger, then prediction block alignment quality is better.
Illustratively, it hands over and than the calculation method of correction value can include:
S1 calculates the centre distance d of prediction block and true frame；
S2, generates the minimum vertex-covering frame of prediction block and true frame, and obtains the catercorner length c of minimum vertex-covering frame；
The quotient d/c of S3, centre distance d and catercorner length c will be handed over as relative position parameter and ratio subtracted relative position
Parameter is handed over and than correction value CIoU, specific friendship and than correction value CIoU=IoU-d/c.
Illustratively, as shown in Figure 1, when being aligned for prediction block and true frame center, wherein a length of w of true frame1,
Width is h1, a length of w of prediction block2, width h2.W in figure1>w2, h1<h2, according to handing over and definition than IoU, then the IoU of the two frames
Value is (w2×h1)/(w1×h1+w2×h2-w2×h1).Further according to the calculating definition of GIoU, the GIoU value of the two frames is (w2×
h1)/(w1×h1+w2×h2-w2×h1)-(w1×h2-w1×h1-w2×h2+w2×h1)/w1×h2。
Similarly in Fig. 2, on the basis of Fig. 1, prediction block deviates the center of true frame, IoU value and GIoU value with
Fig. 1 is identical.However, the coincidence situation of two frames of Fig. 1 and Fig. 2 is not obviously identical, therefore only use IoU or GIoU as frame
The measurement of coincidence degree is clearly to have shortcoming.Reviewing friendship and than correction value CIoU, two frame central points are overlapped in Fig. 1,
Distance is 0, therefore a d/c value is that two frame central point distances are w in 0, CIoU=IoU, Fig. 2 behind CIoU1-w2/ 2, it is minimum
Covering rectangle diagonal line length is sqrt (w1×w1+h2×h2), therefore its CIoU=IoU- (w1-w2/2)/sqrt(w1×w1+h2×
h2), it is clear that the numerical value is less than the case where center Fig. 1 is aligned, therefore can identify the coincidence effect and do not have that Fig. 1 is good, this just with
Actual conditions are consistent.
Further, it is not special case in order to illustrate the case where Fig. 1 and Fig. 2, continues discussing two frame perfect registrations, packet herein
Contain, intersect, is non-intersecting, the various situations such as infinity.
As shown in Fig. 3, Fig. 4 and Fig. 5, when two frames are inclusion relation, it is assumed that one of them big frame a length of 100, width are
60；A length of the 50 of another small frame, width 30.
In Fig. 3, when small frame is included in the lower right corner of big frame, IoU=0.25, GIoU=0.25, CIoU=0.
In Fig. 4, when small frame connects in the side long side of big frame and when axial symmetry, IoU=0.25, GIoU=0.25, CIoU
=0.1214.
In Fig. 5, when small frame is aligned with big frame center, IoU=0.25, GIoU=0.25, CIoU=0.25.
Comparison diagram 3, Fig. 4 and the parameter of Fig. 5 as it can be seen that CIoU compares IoU and GIoU, can further reaction frame and frame it
Between cross reference, be embodied in the degree that center between two baskets is aligned.When two frame center degree of registration are higher, CIoU value
It is higher.
It is as also shown in fig. 6, it is assumed that one of them big frame a length of 100, width 60；A length of the 25 of another small frame, width 15,
When small frame is included in the lower right corner of big frame, IoU=0.0625, GIoU=0.0625, CIoU=-0.3125.
As shown in Figure 7 and Figure 8, it is assumed that two frames length is 100, and wide is 60, is partially overlapped.
In Fig. 7, when two frame intersection width are 15, IoU=0.1429, GIoU=0.1429, CIoU=-0.1675.
In Fig. 8, two frames diagonally 1/4 area be overlapped when, IoU=0.1429, GIoU=-0.0794, CIoU=-
0.1905。
As shown in Fig. 9, Figure 10, Figure 11 and Figure 12, it is assumed that big frame a length of 100, width 60；Small frame a length of 50, width 30；Two
Frame is not overlapped.
In Fig. 9, when big frame and small frame diagonally connect, IoU=0, GIoU=-0.4444, CIoU=-0.5.
In Figure 10, when big frame and small frame bottom alignment and long side connect when, IoU=0, GIoU=-0.3333, CIoU=-
0.3826。
In Figure 11, when big frame and small frame side long side connect and when axial symmetry, IoU=0, GIoU=-0.3333, CIoU
=-0.3345.
In Figure 12, when big frame and small frame side broadside connect and when axial symmetry, IoU=0, GIoU=-0.3333, CIoU
=-0.4642.
When two frames are completely coincident, their IoU=GIoU=CIoU=1；When two frames are at a distance of infinity, IoU
=0, GIoU=CIoU=-1.
In conclusion CIoU inherits the scale invariability of GIoU；CIoU is also regarded as under IoU as GIoU
Boundary is less than or equal to IoU；Recurrence loss loss=1-CIoU can be used when calculating loss, even if two comparison frames are non-intersecting
There is also gradient passbacks for time；The codomain of CIoU is identical as GIoU, but focuses more on and return frame in difficulty, this point is just used for reference
The thought of Classification Loss Focal Loss.
In further embodiments, the evaluation for the prediction block completed for detection positioning, can calculate according to true frame
It hands over and than correction value, will hand over and the prediction block for being greater than the first setting value than correction value is judged as that positioning result is correct.For example,
First setting value is 0.35.
In further embodiments, can simultaneously using hand over and ratio and friendship and than correction value as judgment criteria, namely when in advance
When surveying the friendship of frame and true frame and be greater than the first setting value than correction value, and handing over and compare greater than the second setting value, prediction block is judged
It is correct for positioning result.For example, the first setting value is 0.35, the second setting value is 0.5.
In the present embodiment, due to handing over and including friendship and ratio and relative position parameter, practical application than the factor of correction value
Will appear in the process different friendships and than and relative position parameter be calculated it is identical friendship and than the situation of correction value, lead to not
Effectively distinguished, for example, CIoU=IoU-d/c=0.5-0.05=0.45 in one example, in another example, CIoU
=IoU-d/c=0.55-0.1=0.45；Therefore, on the basis of limiting friendship and comparing, friendship is further limited and than correction value, energy
The susceptibility of enough promotion threshold value evaluations.
Further, it can also be further calculated for target detection and localization result and specify classification object in single image
Setting accuracy specifies the mean accuracy of classification object in multiple image, the mean accuracy of multiple classification objects in multiple image
Mean value, for reflecting the alignment quality of target detection model, step includes: to calculate in single width target detection image for specified class
The other correct prediction block quantity of object detection result, and ratio is sought with specified classification object sum, obtain the specified classification object
The accurate rate of positioning result.
Based on multiple target detection images, calculate separately in each target detection image for the specified classification object
The accurate rate of body is simultaneously averaged, and the mean accuracy of specified classification object positioning is obtained.
Based on multiple target detection images, corresponding mean accuracy is calculated separately for multiple classification objects and asks equal
Value, obtains mean accuracy mean value.
Wherein, the value of accurate rate, mean accuracy and mean accuracy mean value is higher, then it represents that the positioning of target detection model
Quality is better.
On the other hand, based on invention structure identical with the alignment quality evaluation method of target detection model shown in Figure 13
Think, the embodiment of the invention also provides a kind of object localization methods, and overlaps will not be repeated.As shown in figure 14, some implementations
The object localization method of example, it may include:
Step 201: obtaining the multiple prediction blocks and correspondence of multiple targets to be detected respectively by target detection and localization operation
Class probability value；
Step 202: calculating the friendship between prediction block and the catercorner length of ratio, centre distance and minimum vertex-covering frame；
Step 203: calculating the relative position parameter of centre distance and corresponding minimum vertex-covering frame catercorner length；
Step 204: using the relative position parameters revision friendship accordingly and comparing, handed over and compare correction value；
Step 205: to hand over and than correction value as the optimum prediction frame of each target to be detected of standard screening.
In the present embodiment, target detection and localization operation can use the target of the trained formation of deep learning neural network
Detection model can export multiple probability for multiple targets to be detected respectively and reach finger by the operation of target detection model
Determine the prediction block of threshold value.
In the actual application process, multiple distribution possibility in target positions to be detected of same type are sparse may also be intensive.
Specifically, target detection model exports multiple prediction blocks again for multiple targets to be detected under intensive scene respectively, it is each to be detected
The prediction block of target is mutually overlapping, chooses optimal frame as screening criteria only with IoU in this case, what is judged is effective
Property can reduce, can not efficiently distinguish different detection targets, showing susceptibility reduces or the feelings that excessively merge of output valve
Condition.In order to effectively improve under intensive scene, for the fan-out capability of each detection target optimum prediction frame, the present embodiment is using warp
The catercorner length of two frame centre distances and minimum vertex-covering frame it is modified friendship and than correction value as standard of comparison, be able to ascend spirit
Sensitivity.
In some embodiments, above-mentioned steps 203, that is, calculate the centre distance and the corresponding minimum vertex-covering frame
The relative position parameter of catercorner length, more specifically, it may include step: 2031, calculate the centre distance and the minimum
The quotient of covering frame catercorner length obtains the relative position parameter.
In some embodiments, above-mentioned steps 204, that is, simultaneously using the corresponding friendship of the relative position parameters revision
Than, handed over and than correction value, more specifically, it may include step: 2041, calculate the friendship and than with the relative position parameter
Difference, obtain it is described friendship and compare correction value.
Specifically, calculating the friendship between each prediction block and than correction value, when being handed between two prediction blocks and bigger than correction value
When the first given threshold, retains the higher prediction block of class probability value, cast out the lesser prediction block of class probability value；When two
Between prediction block hand over and than correction value less than the first given threshold when, then two prediction blocks retain.It is repeated according to above-mentioned steps
It calculates, finally exports corresponding optimum prediction frame for each detection target.Wherein, the first given threshold is according to practical application scene
It determines, for example, the first given threshold can be 0.5.
In some embodiments, above-mentioned steps 205, that is, using the friendship and than correction value to be to be detected described in standard screening
The optimum prediction frame of target, more specifically, it may include step: 2051, using the friendship and than correction value as the non-maximum of standard
(Nms, Non-maximum suppression) iteration is inhibited to filter out the optimum prediction frame.
In the present embodiment, it in order to reduce calculation amount, can be filtered out using non-maximum restraining Nms iteration described optimal pre-
Survey frame；It is non-very big in the present embodiment compared to traditional non-maximum restraining Nms using handing over and eliminate extra frame as standard than IoU
Inhibit Nms using handing over and being used as evaluation criterion than correction value CIoU, is able to ascend under multiple detection heavy dense targets distribution scenes
Detect location sensitivity and accuracy.
Illustratively, multiple prediction blocks image obtained through target detection and localization operation are according to its corresponding class probability
Be worth it is ascending arranged, such as: A, B, C, D, E, F, G；
Since the maximum prediction block G of class probability value, the friendship of prediction block A, B, C, D, E, F and prediction block G are judged respectively
And whether it is greater than the second given threshold than correction value CIoU, the prediction block side greater than the second given threshold is cast out, and is set less than second
The prediction block for determining threshold value retains, and the wherein friendship of prediction block B, D and prediction block G is obtained after calculating and is greater than second than correction value CIoU
Given threshold then gives up prediction block B, D, retains G.
Further, from remaining prediction block A, C, E, F, the maximum prediction block E of selection sort probability value judges respectively
Whether the friendship of prediction block A, C, F and prediction block E are simultaneously greater than the second given threshold than correction value CIoU, greater than the second given threshold
Prediction block side is cast out, and the prediction block less than the second given threshold retains, and the friendship of wherein prediction block A and prediction block E are obtained after calculating
And be greater than the second given threshold than correction value CIoU, then give up prediction block A, retention forecasting frame E.It repeats to change according to above-mentioned steps
In generation, eventually finds institute prediction block G, E, C with a grain of salt.
Then prediction block G, E, C is respectively the optimum prediction frame for being directed to three detection targets in image.
In further embodiments, target to be detected can be carried out using deep neural network training objective detection model
Positioning.Wherein, using the process of deep neural network training objective detection model, it may include following steps:
S301 generates multiple candidate frames using slip window sampling or selective search method；
S302 calculates candidate frame and the corresponding true friendship of frame and the diagonal line length of ratio, centre distance and minimum vertex-covering frame
Degree；
S303 calculates the quotient of the centre distance and the corresponding minimum vertex-covering frame catercorner length as relative position
Parameter, with friendship and ratio subtracts relative position parameter, is handed over and compares correction value；
S304, will hand over and the candidate frame for being greater than third threshold value than correction value is classified as object frame, carry out for input model
Training；It will hand over and the candidate frame for being less than third threshold value than correction value be classified as background frame.
On the other hand, the application also provides a kind of computer equipment, including memory, processor and storage are on a memory
And the computer program that can be run on a processor, when the processor executes described program the step of the realization above method.
On the other hand, the application also provides a kind of computer readable storage medium, is stored thereon with computer program, the journey
The step of above method is realized when sequence is executed by processor.
In the description of this specification, reference term " one embodiment ", " specific embodiment ", " some implementations
Example ", " such as ", the description of " example ", " specific example " or " some examples " etc. mean it is described in conjunction with this embodiment or example
Particular features, structures, materials, or characteristics are included at least one embodiment or example of the invention.In the present specification,
Schematic expression of the above terms may not refer to the same embodiment or example.Moreover, the specific features of description, knot
Structure, material or feature can be combined in any suitable manner in any one or more of the embodiments or examples.Each embodiment
Involved in the step of sequence be used to schematically illustrate implementation of the invention, sequence of steps therein is not construed as limiting, can be as needed
It appropriately adjusts.
It should be understood by those skilled in the art that, the embodiment of the present invention can provide as method, system or computer program
Product.Therefore, complete hardware embodiment, complete software embodiment or reality combining software and hardware aspects can be used in the present invention
Apply the form of example.Moreover, it wherein includes the computer of computer usable program code that the present invention, which can be used in one or more,
The computer program implemented in usable storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.) produces
The form of product.
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that every one stream in flowchart and/or the block diagram can be realized by computer program instructions
The combination of process and/or box in journey and/or box and flowchart and/or the block diagram.It can provide these computer programs
Instruct the processor of general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices to produce
A raw machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute for real
The device for the function of being specified in present one or more flows of the flowchart and/or one or more blocks of the block diagram.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that it includes referring to that instruction stored in the computer readable memory, which generates,
Enable the manufacture of device, the command device realize in one box of one or more flows of the flowchart and/or block diagram or
The function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that counting
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, thus in computer or
The instruction executed on other programmable devices is provided for realizing in one or more flows of the flowchart and/or block diagram one
The step of function of being specified in a box or multiple boxes.
Particular embodiments described above has carried out further in detail the purpose of the present invention, technical scheme and beneficial effects
Describe in detail it is bright, it should be understood that the above is only a specific embodiment of the present invention, the guarantor being not intended to limit the present invention
Range is protected, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should be included in this
Within the protection scope of invention.
Claims (10)
1. a kind of alignment quality evaluation method of target detection model characterized by comprising
Target to be detected is positioned using target detection model, obtains corresponding prediction block；
Calculate the prediction block and the corresponding true friendship of frame and the catercorner length of ratio, centre distance and minimum vertex-covering frame；
Calculate the relative position parameter of the centre distance and the corresponding minimum vertex-covering frame catercorner length；
Using the relative position parameters revision friendship accordingly and compares, handed over and compare correction value；
The alignment quality of the setting target detection model is evaluated according to the friendship and than correction value.
2. the alignment quality evaluation method of target detection model according to claim 1, which is characterized in that calculate in described
The relative position parameter of heart distance and the corresponding minimum vertex-covering frame catercorner length, comprising:
The quotient for calculating the centre distance and the minimum vertex-covering frame catercorner length obtains the relative position parameter.
3. the alignment quality evaluation method of target detection model according to claim 2, which is characterized in that utilize the phase
To the friendship accordingly of position parameters revision and compares, handed over and compares correction value, comprising:
Calculate it is described friendship and compare the difference with the relative position parameter, obtain it is described hand over and compare correction value.
4. the alignment quality evaluation method of target detection model according to claim 3, which is characterized in that according to the friendship
And the alignment quality of the setting target detection model is evaluated than correction value, comprising:
By the friendship and the prediction block than correction value greater than the first setting value is judged as that positioning result is correct.
5. a kind of object localization method characterized by comprising
Obtain the multiple prediction blocks and corresponding class probability value of multiple targets to be detected respectively by target detection and localization operation；
Calculate the catercorner length of the friendship between the prediction block and ratio, centre distance and minimum vertex-covering frame；
Calculate the relative position parameter of the centre distance and the corresponding minimum vertex-covering frame catercorner length；
Using the relative position parameters revision friendship accordingly and compares, handed over and compare correction value；
Using the friendship and than correction value as the optimum prediction frame of each target to be detected of standard screening.
6. object localization method according to claim 5, which is characterized in that calculate the centre distance and corresponding described
The relative position parameter of minimum vertex-covering frame catercorner length, comprising:
The quotient for calculating the centre distance and the minimum vertex-covering frame catercorner length obtains the relative position parameter.
7. object localization method according to claim 6, which is characterized in that corresponding using the relative position parameters revision
The friendship and compare, handed over and compare correction value, comprising:
Calculate it is described friendship and compare the difference with the relative position parameter, obtain it is described hand over and compare correction value.
8. object localization method according to claim 7, which is characterized in that using the friendship and than correction value as standard screening
The optimum prediction frame of the target to be detected, comprising:
Using the friendship and the optimum prediction frame is filtered out as standard with non-maxima suppression iteration than correction value.
9. a kind of computer equipment including memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, which is characterized in that the processor realizes any one of claim 1 to 8 the method when executing described program
Step.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor
The step of any one of claim 1 to 8 the method is realized when execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910794302.3A CN110503095B (en) | 2019-08-27 | 2019-08-27 | Positioning quality evaluation method, positioning method and device of target detection model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910794302.3A CN110503095B (en) | 2019-08-27 | 2019-08-27 | Positioning quality evaluation method, positioning method and device of target detection model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110503095A true CN110503095A (en) | 2019-11-26 |
| CN110503095B CN110503095B (en) | 2022-06-03 |
Family
ID=68589590
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910794302.3A Active CN110503095B (en) | 2019-08-27 | 2019-08-27 | Positioning quality evaluation method, positioning method and device of target detection model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503095B (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111738072A (en) * | 2020-05-15 | 2020-10-02 | 北京百度网讯科技有限公司 | Training method and device of target detection model and electronic equipment |
| CN111797993A (en) * | 2020-06-16 | 2020-10-20 | 东软睿驰汽车技术（沈阳）有限公司 | Evaluation method and device for deep learning model, electronic equipment and storage medium |
| CN111814850A (en) * | 2020-06-22 | 2020-10-23 | 浙江大华技术股份有限公司 | Defect detection model training method, defect detection method and related device |
| CN112001247A (en) * | 2020-07-17 | 2020-11-27 | 浙江大华技术股份有限公司 | Multi-target detection method, equipment and storage device |
| CN112001453A (en) * | 2020-08-31 | 2020-11-27 | 北京易华录信息技术股份有限公司 | Method and device for calculating accuracy of video event detection algorithm |
| CN112002131A (en) * | 2020-07-16 | 2020-11-27 | 深圳云游四海信息科技有限公司 | In-road parking behavior detection method and device |
| CN112200217A (en) * | 2020-09-09 | 2021-01-08 | 天津津航技术物理研究所 | Identification algorithm evaluation method and system based on infrared image big data |
| CN112287898A (en) * | 2020-11-26 | 2021-01-29 | 深源恒际科技有限公司 | Method and system for evaluating text detection quality of image |
| CN113095301A (en) * | 2021-05-21 | 2021-07-09 | 南京甄视智能科技有限公司 | Road occupation operation monitoring method, system and server |
| CN113408342A (en) * | 2021-05-11 | 2021-09-17 | 深圳大学 | Target detection method for determining intersection ratio threshold based on features |
| CN115035186A (en) * | 2021-12-03 | 2022-09-09 | 荣耀终端有限公司 | Target object marking method and terminal equipment |
| CN115908790A (en) * | 2022-12-28 | 2023-04-04 | 北京斯年智驾科技有限公司 | Target detection center point offset detection method and device and electronic equipment |
| CN116309696A (en) * | 2022-12-23 | 2023-06-23 | 苏州驾驶宝智能科技有限公司 | Multi-category multi-target tracking method and device based on improved generalized cross-over ratio |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
| CN107862437A (en) * | 2017-10-16 | 2018-03-30 | 中国人民公安大学 | The public domain crowd massing method for early warning and system assessed based on risk probability |
| CN108197628A (en) * | 2017-12-07 | 2018-06-22 | 维森软件技术（上海）有限公司 | The joint judgment method of characteristics of image based on deep neural network |
| CN109145756A (en) * | 2018-07-24 | 2019-01-04 | 湖南万为智能机器人技术有限公司 | Object detection method based on machine vision and deep learning |
| CN110070074A (en) * | 2019-05-07 | 2019-07-30 | 安徽工业大学 | A method of building pedestrian detection model |
| CN110084829A (en) * | 2019-03-12 | 2019-08-02 | 上海阅面网络科技有限公司 | Method for tracking target, device, electronic equipment and computer readable storage medium |
-
2019
- 2019-08-27 CN CN201910794302.3A patent/CN110503095B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
| CN107862437A (en) * | 2017-10-16 | 2018-03-30 | 中国人民公安大学 | The public domain crowd massing method for early warning and system assessed based on risk probability |
| CN108197628A (en) * | 2017-12-07 | 2018-06-22 | 维森软件技术（上海）有限公司 | The joint judgment method of characteristics of image based on deep neural network |
| CN109145756A (en) * | 2018-07-24 | 2019-01-04 | 湖南万为智能机器人技术有限公司 | Object detection method based on machine vision and deep learning |
| CN110084829A (en) * | 2019-03-12 | 2019-08-02 | 上海阅面网络科技有限公司 | Method for tracking target, device, electronic equipment and computer readable storage medium |
| CN110070074A (en) * | 2019-05-07 | 2019-07-30 | 安徽工业大学 | A method of building pedestrian detection model |
Non-Patent Citations (4)
| Title |
|---|
| ANIRBANMITRA等: "The region of interest localization for glaucoma analysis from retinal fundus image using deep learning", 《COMPUTER METHODS AND PROGRAMS IN BIOMEDICINE》 * |
| HAMID REZATOFIGHI等: "Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression", 《ARXIV》 * |
| JIAHUI YU等: "UnitBox: An Advanced Object Detection Network", 《ARXIV》 * |
| 耿磊等: "基于集成学习与位置信息约束的前方车辆检测", 《计算机工程与科学》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111738072A (en) * | 2020-05-15 | 2020-10-02 | 北京百度网讯科技有限公司 | Training method and device of target detection model and electronic equipment |
| CN111797993A (en) * | 2020-06-16 | 2020-10-20 | 东软睿驰汽车技术（沈阳）有限公司 | Evaluation method and device for deep learning model, electronic equipment and storage medium |
| CN111797993B (en) * | 2020-06-16 | 2024-02-27 | 东软睿驰汽车技术（沈阳）有限公司 | Evaluation method and device of deep learning model, electronic equipment and storage medium |
| CN111814850A (en) * | 2020-06-22 | 2020-10-23 | 浙江大华技术股份有限公司 | Defect detection model training method, defect detection method and related device |
| CN112002131A (en) * | 2020-07-16 | 2020-11-27 | 深圳云游四海信息科技有限公司 | In-road parking behavior detection method and device |
| CN112001247A (en) * | 2020-07-17 | 2020-11-27 | 浙江大华技术股份有限公司 | Multi-target detection method, equipment and storage device |
| CN112001453A (en) * | 2020-08-31 | 2020-11-27 | 北京易华录信息技术股份有限公司 | Method and device for calculating accuracy of video event detection algorithm |
| CN112001453B (en) * | 2020-08-31 | 2024-03-08 | 北京易华录信息技术股份有限公司 | Method and device for calculating accuracy of video event detection algorithm |
| CN112200217B (en) * | 2020-09-09 | 2023-06-09 | 天津津航技术物理研究所 | Identification algorithm evaluation method and system based on infrared image big data |
| CN112200217A (en) * | 2020-09-09 | 2021-01-08 | 天津津航技术物理研究所 | Identification algorithm evaluation method and system based on infrared image big data |
| CN112287898A (en) * | 2020-11-26 | 2021-01-29 | 深源恒际科技有限公司 | Method and system for evaluating text detection quality of image |
| CN113408342A (en) * | 2021-05-11 | 2021-09-17 | 深圳大学 | Target detection method for determining intersection ratio threshold based on features |
| CN113095301A (en) * | 2021-05-21 | 2021-07-09 | 南京甄视智能科技有限公司 | Road occupation operation monitoring method, system and server |
| CN115035186A (en) * | 2021-12-03 | 2022-09-09 | 荣耀终端有限公司 | Target object marking method and terminal equipment |
| CN116309696A (en) * | 2022-12-23 | 2023-06-23 | 苏州驾驶宝智能科技有限公司 | Multi-category multi-target tracking method and device based on improved generalized cross-over ratio |
| CN116309696B (en) * | 2022-12-23 | 2023-12-01 | 苏州驾驶宝智能科技有限公司 | Multi-category multi-target tracking method and device based on improved generalized cross-over ratio |
| CN115908790A (en) * | 2022-12-28 | 2023-04-04 | 北京斯年智驾科技有限公司 | Target detection center point offset detection method and device and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110503095B (en) | 2022-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110503095A (en) | Alignment quality evaluation method, localization method and the equipment of target detection model | |
| KR102210715B1 (en) | Method, apparatus and device for determining lane lines in road | |
| CN110287932B (en) | Road blocking information extraction method based on deep learning image semantic segmentation | |
| CN105913415B (en) | A kind of image sub-pixel edge extracting method with extensive adaptability | |
| CN108470354A (en) | Video target tracking method, device and realization device | |
| CN109115812A (en) | A kind of weld seam egative film defect identification method and system | |
| KR102021945B1 (en) | Determination of localised quality measurements from a volumetric image record | |
| US10445908B1 (en) | Computer handling of polygons | |
| CN105243356B (en) | A kind of method and device that establishing pedestrian detection model and pedestrian detection method | |
| CN107218894A (en) | A kind of subpixel accuracy thickness of detector detection method of fast and stable | |
| CN115272652A (en) | Dense object image detection method based on multiple regression and adaptive focus loss | |
| CN109117746A (en) | Hand detection method and machine readable storage medium | |
| CN115797736A (en) | Method, device, equipment and medium for training target detection model and target detection | |
| AVENDAÑO | Identification and quantification of concrete cracks using image analysis and machine learning | |
| Xu et al. | An APF-ACO algorithm for automatic defect detection on vehicle paint | |
| CN104915951A (en) | Stippled DPM two-dimensional code area positioning method | |
| CN108171696A (en) | A kind of placenta detection method and device | |
| US20220314113A1 (en) | Generating point cloud completion network and processing point cloud data | |
| CN104637049A (en) | Automatic detection method for optical fiber coiling quality | |
| US20240087105A1 (en) | Systems and Methods for Paint Defect Detection Using Machine Learning | |
| CN110017998A (en) | Vehicle checking method, device and equipment | |
| CN116109701A (en) | Object grabbing method based on passive dual-purpose high light reflection | |
| Wu et al. | Enhanced roadway geometry data collection using an effective video log image-processing algorithm | |
| CN113538351A (en) | External insulation equipment defect degree evaluation method fusing multi-parameter electric signals | |
| Shabankareh et al. | Improvement of edge-tracking methods using genetic algorithm and neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |