CN109740662A - Image object detection method based on YOLO frame - Google Patents
Image object detection method based on YOLO frame Download PDFInfo
- Publication number
- CN109740662A CN109740662A CN201811621484.6A CN201811621484A CN109740662A CN 109740662 A CN109740662 A CN 109740662A CN 201811621484 A CN201811621484 A CN 201811621484A CN 109740662 A CN109740662 A CN 109740662A
- Authority
- CN
- China
- Prior art keywords
- image
- size
- classification
- classification number
- value
- 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.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 27
- 238000000034 method Methods 0.000 claims abstract description 46
- 238000006243 chemical reaction Methods 0.000 claims abstract description 23
- 238000012545 processing Methods 0.000 claims abstract description 14
- 238000012360 testing method Methods 0.000 claims abstract description 6
- 230000001629 suppression Effects 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 19
- 238000000605 extraction Methods 0.000 claims description 12
- 230000005540 biological transmission Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 230000005611 electricity Effects 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims 1
- 238000005070 sampling Methods 0.000 claims 1
- 238000011161 development Methods 0.000 abstract description 4
- 238000013135 deep learning Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 4
- 239000000203 mixture Substances 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image object detection methods based on YOLO frame.This method is by establishing YOLO frame model, input YOLO frame model obtains multiple prediction tensor values after testing image to be converted into the image of fixed size, each prediction tensor value obtains prediction rectangle frame by conversion formula retrospectively calculate, obtained multiple prediction rectangle frames are passed through into non-maximum suppression algorithm process, obtain most reliable rectangle frame, and the classification that target can be obtained and position are converted it in original image, YOLO frame model is in original YOLO frame foundation, it makes improvements, so that detectability greatly promotes, so that original Small object that can't detect is able to detect that, and the approximate location that and target target be can be accurately identified, it is compared to traditional modeling method simultaneously, reduce the development time, substantially increase the accuracy and speed of detection, so that algorithm Real-time ability can be reached.It is suitble to promote and apply in technical field of data processing.
Description
Technical field
The present invention relates to technical field of data processing, especially a kind of image object detection method based on YOLO frame.
Background technique
The task of target detection is position and the classification found out all interested targets in image, and determine target, this
It is one of the key problem in computer vision.Target detection mainly solve the problems, such as be target itself query, " Who Am I? "
" I where? ", i.e., target be in the picture what and where the problem of.
Small object in image is locating in present daily life as it can be seen that such as general camera shoots remote object
When will generate Small object；When vehicle identifies the small traffic lights in front at a distance in automatic Pilot；Medical microscope is clapped
Also it will appear Small object etc. when taking the photograph cell image.Small object is closely coupled for our daily life, closely bound up, by grinding
Studying carefully Small object can be finer with the life of let us, more convenient.
Small object is specifically defined there are two types of mode, and one is absolute dimension definition, another kind is that relative size defines.Absolutely
The pixel for defining target in i.e. image to size is less than 32*32, that is, is regarded as Small object.Relative size definition is target picture
Element is less than wide and high 1/10th of image, so that it may be considered Small object.In the research of this patent, Small object is all absolute ruler
Very little definition.
In the development history of target detection, what is mainly studied before this is big target, i.e., occupies in an image more
The target of pixel.It is constantly progressive with research, just starts the Small object that gradually begins one's study, and in the small mesh that most begins one's study
Target image be also not using general camera shoot image, but infrared camera shooting image.It is introduced by machine learning
After image, the image of general camera shooting is just begun to use to make a search.
For the research of the small target deteection of general camera shooting, there are mainly two types of study routes to carry out at present, traditional
Machine learning method and the preferable deep learning method of effect.Traditional machine learning method mainly has HOG+SVM, DPM, Haar
+ Adaboost etc., the algorithm above mainly completes detection work using the single features of image and a classifier, this
Method has a fatal defect, and generalization ability is too weak.Simple to understand to be exactly that HOG feature is mainly used for pedestrian detection, Harr is special
Sign is mainly used for Face datection, the model that training obtains is used for other targets, obtained effect can be very poor.This defect is always
Scientists are annoying, until the appearance of deep learning method, just gradually improve this problem.
In the algorithm of target detection of current mainstream, deep learning occupies mainstream, not only due to the effect of its detection is good, and
And the required time is also gradually decreasing.There are two " school ", single phase detections and two for current main deep learning method
Stage detection.Two " schools " respectively have the advantage of oneself, can be selected to be suitble to itself according to the actual conditions of the application of oneself
Method.Two stage method occupies the rivers and mountains great Bi with the advantage of accuracy, and classical method has FastR-CNN, Faster R-
CNN, MaskR-CNN, the main thought of this method are to inherit conventional method, after continuously improve till now achievement.The party
Method used convolutional network to extract feature before this, and generate target may where region, these possible regions are put into point
In class device and recurrence device, identification mission and Detection task are done respectively, finally obtains classification and the exact position of target.Though this method
Right accuracy is high, but is not achieved in real time, therefore the method for single phase is come into being.
Single-stage process occupies the ground of a side with speed advantage, typically has YOLO serial in single-stage process.YOLO
Main thought be a given picture, classification and the position of target are directly obtained by the thought of recurrence.Although this method is fast
Degree is fast, but accuracy is unsatisfactory, especially when detecting Small object.
Above-mentioned method is all the algorithm in general target detection, also seldom for the algorithm of small target deteection, there is utilization
The information on Small object periphery is helped to detect Small object and is predicted in a network using more shallow layer target, the above-mentioned party
Method is not significant to the promotion effect of Small object, and the required time is also huge.
Summary of the invention
Technical problem to be solved by the invention is to provide one kind can detect Small object, and can accurately know
Other target is the approximate location of what and target, and enables the algorithm to the figure based on YOLO frame for reaching real-time ability
As object detection method.
The technical solution adopted by the present invention to solve the technical problems are as follows: should the image object detection side based on YOLO frame
Method, comprising the following steps:
1) YOLO frame model, is established；
The foundation of the YOLO frame model includes the following steps:
A, the image for acquiring transmission line of electricity establishes data set；
B, each image concentrated to data pre-processes, and the pretreatment includes image cropping, scales, overturning,
Displacement, rotation, brightness adjustment, plus noise；
C, feature extraction is carried out to the image by step B processing；The process of the feature extraction is as follows: will pass through step
Each image of B processing inputs darknet53 network respectively and carries out character extraction, and obtaining three sizes is respectively 13*
The characteristic pattern of 13,26*26,52*52；
D, each image is obtained after predicting network processes respectively by feature extraction three obtained characteristic pattern
Three tensor values；Details are provided below: the characteristic pattern of 13*13 being first passed through to the convolutional layer of 5 conversions, passes through 2 later
It predicts convolutional layer, the vector of 3* (the classification number of 4+1+ data set) is finally obtained in each characteristic point, in conjunction with 13*13 spy
The size for levying figure, obtains the tensor value of one (batch_size, 3,13,13,3* (the classification numbers of 4+1+ data set))；It is right later
The characteristic pattern of 13*13 is up-sampled to obtain the characteristic pattern of 26*26, the 26*26 characteristic pattern with this image obtained in step C
It blends, obtains a new characteristic pattern, at convolutional layer and 2 prediction convolutional layers of the new characteristic pattern by 5 conversions
An equal amount of vector is obtained after reason in each characteristic point, in conjunction with the size of 26*26 characteristic pattern, obtains one
The tensor value of (batch_size, 3,26,26,3* (the classification number of 4+1+ data set))；Then the characteristic pattern of 26*26 is done and is adopted
Sample obtains the characteristic pattern of 52*52, blends with the 52*52 characteristic pattern of this image obtained in step C, obtains a new spy
After sign figure, the convolutional layer converted to the new characteristic pattern by 5 and 2 prediction convolutional layer processing in each characteristic point
(batch_size, 3,52,52,3* (4+1+ are obtained in conjunction with the size of 52*52 characteristic pattern to an equal amount of vector
The classification number of data set)) tensor value；Wherein number 3 indicates that the quantity of the anchor on this feature figure, number 4 indicate prediction
The centre coordinate value of obtained prediction result and wide high level, number 1 indicate the confidence level of prediction block, and classification number is indicated in this feature
The probability of classification is predicted on point；
E, the label data of each image is obtained, the label data includes centre coordinate value bx, by, wide high level bw,
Bh and classification, and the label data of acquisition is converted into training data, specific conversion process is as described below: by the number of tags of acquisition
According to centre coordinate value bx, by and wide high level bw, bh substitute into following conversion formula and obtain the centre coordinate value tx of training data,
Ty, wide high level
Tw, th and confidence level, the classification number of training data and the classification number of label data are identical；The conversion formula
Are as follows:
Wherein cx, cy are predicted grid, and pw, ph are predetermined anchor value；
F, penalty values calculate, and the penalty values include coordinate loss and confidence level loss, Classification Loss, specific calculation
It is as follows:
The centre coordinate value of the training data obtained in each characteristic point and wide high level are obtained on this feature point respectively
The centre coordinate value in (batch_size, 3,13,13,3* (the classification number of 4+1+ data set)) tensor value and wide high level that arrive,
Centre coordinate value in (batch_size, 3,26,26,3* (the classification number of 4+1+ data set)) tensor value and wide high level,
Centre coordinate value and wide high level in (batch_size, 3,52,52,3* (the classification number of 4+1+ data set)) tensor value pass through flat
The mode of square error loss carrys out coordinates computed loss；
By the confidence level of the training data obtained in each characteristic point respectively with the (batch_ that is obtained on this feature point
Size, 3,13,13,3* (the classification number of 4+1+ data set)) confidence level in tensor value, (batch_size, 3,26,26,3* (4
The classification number of+1+ data set)) confidence level, (batch_size, the 3,52,52,3* (classification of 4+1+ data set in tensor value
Number)) confidence level in tensor value calculates confidence level in such a way that square error is lost；
By the classification number of the training data obtained in each characteristic point respectively with the (batch_ that is obtained on this feature point
Size, 3,13,13,3* (the classification number of 4+1+ data set)) classification number, (batch_size, 3,26,26,3* (4 in tensor value
The classification number of+1+ data set)) classification number in tensor value, (batch_size, the 3,52,52,3* (classification of 4+1+ data set
Number)) the classification number in tensor value calculates Classification Loss by way of intersecting entropy loss；
G, by the algorithm that reversed gradient is propagated be iterated calculating gradually decrease penalty values until penalty values no longer reduce,
Obtain final YOLO frame model；
2) input YOLO frame model obtains multiple predictions, testing image to be converted into the image of fixed size after
Magnitude, each prediction tensor value includes centre coordinate value tx, ty, wide high level tw, th and confidence level and classification number, by acquisition
Predict centre coordinate value tx, the ty of tensor value, wide high level tw, th substitute into following conversion formula retrospectively calculate and obtain prediction rectangle frame
Centre coordinate value bx, by, width high level bw, bh and confidence level, predict the classification number of frame data and the class of prediction tensor value
It does not count identical；The conversion formula are as follows:
Wherein cx, cy are predicted grid, and pw, ph are predetermined anchor value；
3) obtained multiple prediction rectangle frames, are obtained into most reliable rectangle frame by non-maximum suppression algorithm process, and
Most reliable rectangle frame is transformed into classification and the position that target can be obtained in original image.
Beneficial effects of the present invention: the image object detection method based on YOLO frame is somebody's turn to do by establishing YOLO frame mould
Type, input YOLO frame model obtains multiple prediction tensor values after testing image to be converted into the image of fixed size, often
A prediction tensor value by conversion formula retrospectively calculate obtain prediction rectangle frame, by obtained multiple prediction rectangle frames by it is non-most
Big restrainable algorithms processing, obtains most reliable rectangle frame, and most reliable rectangle frame is transformed into original image, target can be obtained
Classification and position YOLO frame model of the present invention by using deep learning in original YOLO frame foundation, it is right
It is improved, so that detectability greatly promotes, so that original Small object that can't detect is able to detect that, and can be quasi-
Really identification target is the approximate location of what and target, while being compared to traditional modeling method, reduces a large amount of mathematics
Formula calculates and model construction, reduces the development time, substantially increases the accuracy and speed of detection, enable the algorithm to reach
To real-time ability, and YOLO frame model of the present invention is using the YOLO method of single phase, this method and its
His two stage method is compared occupies huge advantage in speed, and can further promote detection essence by cascade method
Degree.
Specific embodiment
The image object detection method based on YOLO frame, comprising the following steps:
1) YOLO frame model, is established；
The foundation of the YOLO frame model includes the following steps:
A, the image for acquiring transmission line of electricity establishes data set；
B, each image concentrated to data pre-processes, and the pretreatment includes image cropping, scales, overturning,
Displacement, rotation, brightness adjustment, plus noise；
C, feature extraction is carried out to the image by step B processing；The process of the feature extraction is as follows: will pass through step
Each image of B processing inputs darknet53 network respectively and carries out character extraction, and obtaining three sizes is respectively 13*
The characteristic pattern of 13,26*26,52*52；
D, each image is obtained after predicting network processes respectively by feature extraction three obtained characteristic pattern
Three tensor values；Details are provided below: the characteristic pattern of 13*13 being first passed through to the convolutional layer of 5 conversions, passes through 2 later
It predicts convolutional layer, the vector of 3* (the classification number of 4+1+ data set) is finally obtained in each characteristic point, in conjunction with 13*13 spy
The size for levying figure, obtains the tensor value of one (batch_size, 3,13,13,3* (the classification numbers of 4+1+ data set))；It is right later
The characteristic pattern of 13*13 is up-sampled to obtain the characteristic pattern of 26*26, the 26*26 characteristic pattern with this image obtained in step C
It blends, obtains a new characteristic pattern, at convolutional layer and 2 prediction convolutional layers of the new characteristic pattern by 5 conversions
An equal amount of vector is obtained after reason in each characteristic point, in conjunction with the size of 26*26 characteristic pattern, obtains one
The tensor value of (batch_size, 3,26,26,3* (the classification number of 4+1+ data set))；Then the characteristic pattern of 26*26 is done and is adopted
Sample obtains the characteristic pattern of 52*52, blends with the 52*52 characteristic pattern of this image obtained in step C, obtains a new spy
After sign figure, the convolutional layer converted to the new characteristic pattern by 5 and 2 prediction convolutional layer processing in each characteristic point
(batch_size, 3,52,52,3* (4+1+ are obtained in conjunction with the size of 52*52 characteristic pattern to an equal amount of vector
The classification number of data set)) tensor value；Wherein number 3 indicates that the quantity of the anchor on this feature figure, number 4 indicate prediction
The centre coordinate value of obtained prediction result and wide high level, number 1 indicate the confidence level of prediction block, and classification number is indicated in this feature
The probability of classification is predicted on point；
E, the label data of each image is obtained, the label data includes centre coordinate value bx, by, wide high level bw,
Bh and classification, and the label data of acquisition is converted into training data, specific conversion process is as described below: by the number of tags of acquisition
According to centre coordinate value bx, by and wide high level bw, bh substitute into following conversion formula and obtain the centre coordinate value tx of training data,
Ty, wide high level
Tw, th and confidence level, the classification number of training data and the classification number of label data are identical；The conversion formula
Are as follows:
Wherein cx, cy are predicted grid, and pw, ph are predetermined anchor value；
F, penalty values calculate, and the penalty values include coordinate loss and confidence level loss, Classification Loss, specific calculation
It is as follows:
The centre coordinate value of the training data obtained in each characteristic point and wide high level are obtained on this feature point respectively
The centre coordinate value in (batch_size, 3,13,13,3* (the classification number of 4+1+ data set)) tensor value and wide high level that arrive,
Centre coordinate value in (batch_size, 3,26,26,3* (the classification number of 4+1+ data set)) tensor value and wide high level,
Centre coordinate value and wide high level in (batch_size, 3,52,52,3* (the classification number of 4+1+ data set)) tensor value pass through flat
The mode of square error loss carrys out coordinates computed loss；
By the confidence level of the training data obtained in each characteristic point respectively with the (batch_ that is obtained on this feature point
Size, 3,13,13,3* (the classification number of 4+1+ data set)) confidence level in tensor value, (batch_size, 3,26,26,3* (4
The classification number of+1+ data set)) confidence level, (batch_size, the 3,52,52,3* (classification of 4+1+ data set in tensor value
Number)) confidence level in tensor value calculates confidence level in such a way that square error is lost；
By the classification number of the training data obtained in each characteristic point respectively with the (batch_ that is obtained on this feature point
Size, 3,13,13,3* (the classification number of 4+1+ data set)) classification number, (batch_size, 3,26,26,3* (4 in tensor value
The classification number of+1+ data set)) classification number in tensor value, (batch_size, the 3,52,52,3* (classification of 4+1+ data set
Number)) the classification number in tensor value calculates Classification Loss by way of intersecting entropy loss；
G, by the algorithm that reversed gradient is propagated be iterated calculating gradually decrease penalty values until penalty values no longer reduce,
Obtain final YOLO frame model；
2) input YOLO frame model obtains multiple predictions, testing image to be converted into the image of fixed size after
Magnitude, the fixed size are ordinarily selected to 416*416 size, and each prediction tensor value includes centre coordinate value tx, ty,
Wide high level tw, th and confidence level and classification number, by centre coordinate value tx, the ty, wide high level tw, th of the prediction tensor value of acquisition
It substitutes into following conversion formula retrospectively calculate and obtains centre coordinate the value bx, by, width high level bw, bh and confidence of prediction rectangle frame
Degree predicts that the classification number of frame data is identical as the prediction classification number of tensor value；The conversion formula are as follows:
Wherein cx, cy are predicted grid, and pw, ph are predetermined anchor value；
3) obtained multiple prediction rectangle frames, are obtained into most reliable rectangle frame by non-maximum suppression algorithm process, and
Most reliable rectangle frame is transformed into classification and the position that target can be obtained in original image.
Beneficial effects of the present invention: the image object detection method based on YOLO frame is somebody's turn to do by establishing YOLO frame mould
Type, input YOLO frame model obtains multiple prediction tensor values after testing image to be converted into the image of fixed size, often
A prediction tensor value by conversion formula retrospectively calculate obtain prediction rectangle frame, by obtained multiple prediction rectangle frames by it is non-most
Big restrainable algorithms processing, obtains most reliable rectangle frame, and most reliable rectangle frame is transformed into original image, target can be obtained
Classification and position YOLO frame model of the present invention by using deep learning in original YOLO frame foundation, it is right
It is improved, so that detectability greatly promotes, so that original Small object that can't detect is able to detect that, and can be quasi-
Really identification target is the approximate location of what and target, while being compared to traditional modeling method, reduces a large amount of mathematics
Formula calculates and model construction, reduces the development time, substantially increases the accuracy and speed of detection, enable the algorithm to reach
To real-time ability, and YOLO frame model of the present invention is using the YOLO method of single phase, this method and its
His two stage method is compared occupies huge advantage in speed, and can further promote detection essence by cascade method
Degree is 85.4% by the Average Accuracy that the experimental verification YOLO frame model detects target, at GTX1080Ti
Processing speed can achieve 31 frame per second, and Small object Average Accuracy is 72.4%.It therefore, should the image mesh based on YOLO frame
Mark detection method can detecte Small object, and can accurately identify the position that and Small object Small object be, and
Processing speed can comparatively fast reach real-time ability.
Claims (1)
1. the image object detection method based on YOLO frame, it is characterised in that the following steps are included:
1) YOLO frame model, is established；
The foundation of the YOLO frame model includes the following steps:
A, the image for acquiring transmission line of electricity establishes data set；
B, each image concentrated to data pre-processes, and the pretreatment includes image cropping, is scaled, and overturns, displacement,
Rotation, brightness adjustment, plus noise；
C, feature extraction is carried out to the image by step B processing；The process of the feature extraction is as follows: will be by step B
Each image of reason inputs darknet53 network respectively and carries out character extraction, and obtaining three sizes is respectively 13*13,
The characteristic pattern of 26*26,52*52；
D, each image is obtained three after predicting network processes respectively by feature extraction three obtained characteristic pattern
Tensor value；Details are provided below: the characteristic pattern of 13*13 being first passed through to the convolutional layer of 5 conversions, passes through 2 predictions later
Convolutional layer finally obtains the vector of 3* (the classification number of 4+1+ data set), in conjunction with 13*13 characteristic pattern in each characteristic point
Size, obtain the tensor value of one (batch_size, 3,13,13,3* (the classification numbers of 4+1+ data set))；Later to 13*
13 characteristic pattern is up-sampled to obtain the characteristic pattern of 26*26, the 26*26 characteristic pattern phase with this image obtained in step C
Fusion, obtains a new characteristic pattern, the convolutional layer convert to the new characteristic pattern by 5 and 2 prediction convolutional layers processing
An equal amount of vector is obtained in each characteristic point afterwards, in conjunction with the size of 26*26 characteristic pattern, obtains (a batch_
Size, 3,26,26,3* (the classification number of 4+1+ data set)) tensor value；Then up-sampling is done to the characteristic pattern of 26*26 to obtain
The characteristic pattern of 52*52 is blended with the 52*52 characteristic pattern of this image obtained in step C, obtains a new characteristic pattern,
The convolutional layer and 2 prediction convolutional layers convert to the new characteristic pattern by 5 obtain same in each characteristic point after handling
The vector of sample size obtains (batch_size, 3,52,52, a 3* (4+1+ data in conjunction with the size of 52*52 characteristic pattern
The classification number of collection)) tensor value；Wherein number 3 indicates that the quantity of the anchor on this feature figure, number 4 indicate that prediction obtains
Prediction result centre coordinate value and wide high level, number 1 indicates the confidence level of prediction block, and classification number indicates on this feature point
Predict the probability of classification；
E, obtain each image label data, the label data include centre coordinate value bx, by, width high level bw, bh with
Classification, and the label data of acquisition is converted into training data, specific conversion process is as described below: by the label data of acquisition
Centre coordinate value bx, by and wide high level bw, bh substitute into following conversion formula and obtain centre coordinate value tx, the ty of training data, wide
High level
Tw, th and confidence level, the classification number of training data and the classification number of label data are identical；The conversion formula are as follows:
Wherein cx, cy are predicted grid, and pw, ph are predetermined anchor value；
F, penalty values calculate, and the penalty values include coordinate loss and confidence level loss, Classification Loss, and specific calculation is such as
Under:
By the centre coordinate value of the training data obtained in each characteristic point and wide high level respectively with obtained on this feature point
Centre coordinate value in (batch_size, 3,13,13,3* (the classification number of 4+1+ data set)) tensor value and wide high level,
Centre coordinate value in (batch_size, 3,26,26,3* (the classification number of 4+1+ data set)) tensor value and wide high level,
Centre coordinate value and wide high level in (batch_size, 3,52,52,3* (the classification number of 4+1+ data set)) tensor value pass through flat
The mode of square error loss carrys out coordinates computed loss；
By the confidence level of the training data obtained in each characteristic point respectively with obtained on this feature point (batch_size,
3,13,13,3* (the classification number of 4+1+ data set)) confidence level in tensor value, (batch_size, 3,26,26,3* (4+1+ number
According to the classification number of collection)) confidence level in tensor value, (batch_size, 3,52,52,3* (the classification number of 4+1+ data set))
Confidence level in magnitude calculates confidence level in such a way that square error is lost；
By the classification number of the training data obtained in each characteristic point respectively with obtained on this feature point (batch_size,
3,13,13,3* (the classification number of 4+1+ data set)) classification number, (batch_size, 3,26,26,3* (4+1+ number in tensor value
According to the classification number of collection)) classification number, (batch_size, 3,52,52,3* (the classification number of 4+1+ data set)) in tensor value
Classification number in magnitude calculates Classification Loss by way of intersecting entropy loss；
G, by the algorithm that reversed gradient is propagated be iterated calculating gradually decrease penalty values until penalty values no longer reduce, obtain
Final YOLO frame model；
2) input YOLO frame model obtains multiple prediction tensors, testing image to be converted into the image of fixed size after
Value, each prediction tensor value includes centre coordinate value tx, ty, wide high level tw, th and confidence level and classification number, by the pre- of acquisition
Centre coordinate value tx, the ty of tensor value are surveyed, wide high level tw, th substitute into following conversion formula retrospectively calculate and obtain prediction rectangle frame
Centre coordinate value bx, by, width high level bw, bh and confidence level predict the classification number of frame data and the classification of prediction tensor value
Number is identical；The conversion formula are as follows:
Wherein cx, cy are predicted grid, and pw, ph are predetermined anchor value；
3) obtained multiple prediction rectangle frames, are obtained into most reliable rectangle frame by non-maximum suppression algorithm process, and will most
Reliable rectangle frame is transformed into the classification that target can be obtained in original image and position.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811621484.6A CN109740662A (en) | 2018-12-28 | 2018-12-28 | Image object detection method based on YOLO frame |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811621484.6A CN109740662A (en) | 2018-12-28 | 2018-12-28 | Image object detection method based on YOLO frame |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109740662A true CN109740662A (en) | 2019-05-10 |
Family
ID=66361722
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811621484.6A Pending CN109740662A (en) | 2018-12-28 | 2018-12-28 | Image object detection method based on YOLO frame |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740662A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490155A (en) * | 2019-08-23 | 2019-11-22 | 电子科技大学 | A kind of no-fly airspace unmanned plane detection method |
| CN110508510A (en) * | 2019-08-27 | 2019-11-29 | 广东工业大学 | A kind of plastic pump defect inspection method, apparatus and system |
| CN111104906A (en) * | 2019-12-19 | 2020-05-05 | 南京工程学院 | Transmission tower bird nest fault detection method based on YOLO |
| CN111337789A (en) * | 2019-10-23 | 2020-06-26 | 西安科技大学 | Method and system for detecting fault electrical element in high-voltage transmission line |
| CN111724355A (en) * | 2020-06-01 | 2020-09-29 | 厦门大学 | Image measuring method for abalone body type parameters |
| CN111738212A (en) * | 2020-07-20 | 2020-10-02 | 平安国际智慧城市科技股份有限公司 | Traffic signal lamp identification method, device, equipment and medium based on artificial intelligence |
| CN111753666A (en) * | 2020-05-21 | 2020-10-09 | 西安科技大学 | Method and system for detecting faults of small targets in power transmission line and storage medium |
| CN111915558A (en) * | 2020-06-30 | 2020-11-10 | 成都思晗科技股份有限公司 | Pin state detection method for high-voltage transmission line |
| CN112668497A (en) * | 2020-12-30 | 2021-04-16 | 南京佑驾科技有限公司 | Vehicle accurate positioning and identification method and system |
| CN114332465A (en) * | 2022-01-05 | 2022-04-12 | 上海秦润数据科技有限公司 | Method for discovering abnormity of natural gas terminal equipment based on computer vision and deep learning |
| CN114648685A (en) * | 2022-03-23 | 2022-06-21 | 成都臻识科技发展有限公司 | Method and system for converting anchor-free algorithm into anchor-based algorithm |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170147905A1 (en) * | 2015-11-25 | 2017-05-25 | Baidu Usa Llc | Systems and methods for end-to-end object detection |
| CN107742093A (en) * | 2017-09-01 | 2018-02-27 | 国网山东省电力公司电力科学研究院 | A kind of infrared image power equipment component real-time detection method, server and system |
| CN108921875A (en) * | 2018-07-09 | 2018-11-30 | 哈尔滨工业大学（深圳） | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane |
| CN109064461A (en) * | 2018-08-06 | 2018-12-21 | 长沙理工大学 | A kind of detection method of surface flaw of steel rail based on deep learning network |
-
2018
- 2018-12-28 CN CN201811621484.6A patent/CN109740662A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170147905A1 (en) * | 2015-11-25 | 2017-05-25 | Baidu Usa Llc | Systems and methods for end-to-end object detection |
| CN107742093A (en) * | 2017-09-01 | 2018-02-27 | 国网山东省电力公司电力科学研究院 | A kind of infrared image power equipment component real-time detection method, server and system |
| CN108921875A (en) * | 2018-07-09 | 2018-11-30 | 哈尔滨工业大学（深圳） | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane |
| CN109064461A (en) * | 2018-08-06 | 2018-12-21 | 长沙理工大学 | A kind of detection method of surface flaw of steel rail based on deep learning network |
Non-Patent Citations (3)
| Title |
|---|
| JOSEPH REDMON: "YOLOv3: An Incremental Improvement", 《ARXIV》 * |
| 木盏: "yolo系列之yolov3【深度解析】", 《CSDN》 * |
| 龚静等: "基于YOLOv2算法的运动车辆目标检测方法研究", 《电子科技》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490155B (en) * | 2019-08-23 | 2022-05-17 | 电子科技大学 | Method for detecting unmanned aerial vehicle in no-fly airspace |
| CN110490155A (en) * | 2019-08-23 | 2019-11-22 | 电子科技大学 | A kind of no-fly airspace unmanned plane detection method |
| CN110508510A (en) * | 2019-08-27 | 2019-11-29 | 广东工业大学 | A kind of plastic pump defect inspection method, apparatus and system |
| CN111337789A (en) * | 2019-10-23 | 2020-06-26 | 西安科技大学 | Method and system for detecting fault electrical element in high-voltage transmission line |
| CN111104906A (en) * | 2019-12-19 | 2020-05-05 | 南京工程学院 | Transmission tower bird nest fault detection method based on YOLO |
| CN111753666B (en) * | 2020-05-21 | 2024-01-23 | 西安科技大学 | Small target fault detection method, detection system and storage medium for power transmission line |
| CN111753666A (en) * | 2020-05-21 | 2020-10-09 | 西安科技大学 | Method and system for detecting faults of small targets in power transmission line and storage medium |
| CN111724355B (en) * | 2020-06-01 | 2022-06-14 | 厦门大学 | Image measuring method for abalone body type parameters |
| CN111724355A (en) * | 2020-06-01 | 2020-09-29 | 厦门大学 | Image measuring method for abalone body type parameters |
| CN111915558A (en) * | 2020-06-30 | 2020-11-10 | 成都思晗科技股份有限公司 | Pin state detection method for high-voltage transmission line |
| CN111915558B (en) * | 2020-06-30 | 2023-12-01 | 成都思晗科技股份有限公司 | Pin state detection method for high-voltage transmission line |
| CN111738212B (en) * | 2020-07-20 | 2020-11-20 | 平安国际智慧城市科技股份有限公司 | Traffic signal lamp identification method, device, equipment and medium based on artificial intelligence |
| CN111738212A (en) * | 2020-07-20 | 2020-10-02 | 平安国际智慧城市科技股份有限公司 | Traffic signal lamp identification method, device, equipment and medium based on artificial intelligence |
| CN112668497A (en) * | 2020-12-30 | 2021-04-16 | 南京佑驾科技有限公司 | Vehicle accurate positioning and identification method and system |
| CN114332465A (en) * | 2022-01-05 | 2022-04-12 | 上海秦润数据科技有限公司 | Method for discovering abnormity of natural gas terminal equipment based on computer vision and deep learning |
| CN114648685A (en) * | 2022-03-23 | 2022-06-21 | 成都臻识科技发展有限公司 | Method and system for converting anchor-free algorithm into anchor-based algorithm |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740662A (en) | Image object detection method based on YOLO frame | |
| CN113065558B (en) | Lightweight small target detection method combined with attention mechanism | |
| WO2018214195A1 (en) | Remote sensing imaging bridge detection method based on convolutional neural network | |
| CN108875600A (en) | A kind of information of vehicles detection and tracking method, apparatus and computer storage medium based on YOLO | |
| CN111160269A (en) | Face key point detection method and device | |
| CN108460403A (en) | The object detection method and system of multi-scale feature fusion in a kind of image | |
| CN105608455B (en) | A kind of license plate sloped correcting method and device | |
| CN108596108B (en) | Aerial remote sensing image change detection method based on triple semantic relation learning | |
| Gong et al. | Object detection based on improved YOLOv3-tiny | |
| CN108960404B (en) | Image-based crowd counting method and device | |
| CN111340123A (en) | Image score label prediction method based on deep convolutional neural network | |
| Tan et al. | Vehicle detection in high resolution satellite remote sensing images based on deep learning | |
| CN109948593A (en) | Based on the MCNN people counting method for combining global density feature | |
| CN105513080B (en) | A kind of infrared image target Salience estimation | |
| CN109635634A (en) | A kind of pedestrian based on stochastic linear interpolation identifies data enhancement methods again | |
| CN103080979A (en) | System and method for synthesizing portrait sketch from photo | |
| CN101923637A (en) | Mobile terminal as well as human face detection method and device thereof | |
| CN104361357A (en) | Photo set classification system and method based on picture content analysis | |
| CN116229248A (en) | Ocean species distribution prediction method, device, equipment and storage medium | |
| Zhang et al. | Research on Surface Defect Detection of Rare‐Earth Magnetic Materials Based on Improved SSD | |
| Li et al. | A self-attention feature fusion model for rice pest detection | |
| Shi et al. | YOLOv5s_2E: Improved YOLOv5s for Aerial Small Target Detection | |
| CN116503750A (en) | Large-range remote sensing image rural block type residential area extraction method and system integrating target detection and visual attention mechanisms | |
| Ouyang et al. | An Anchor-free Detector with Channel-based Prior and Bottom-Enhancement for Underwater Object Detection | |
| Zhang et al. | Design and implementation of object image detection interface system based on PyQt5 and improved SSD algorithm |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190510 |
|
| WD01 | Invention patent application deemed withdrawn after publication |