CN109034258A - Weakly supervised object detection method based on certain objects pixel gradient figure - Google Patents

Weakly supervised object detection method based on certain objects pixel gradient figure Download PDF

Info

Publication number
CN109034258A
CN109034258A CN201810877293.XA CN201810877293A CN109034258A CN 109034258 A CN109034258 A CN 109034258A CN 201810877293 A CN201810877293 A CN 201810877293A CN 109034258 A CN109034258 A CN 109034258A
Authority
CN
China
Prior art keywords
certain objects
indicate
classification
pixel gradient
image
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
Application number
CN201810877293.XA
Other languages
Chinese (zh)
Inventor
纪荣嵘
沈云航
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xiamen University
Original Assignee
Xiamen University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xiamen University filed Critical Xiamen University
Priority to CN201810877293.XA priority Critical patent/CN109034258A/en
Publication of CN109034258A publication Critical patent/CN109034258A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/469—Contour-based spatial representations, e.g. vector-coding
    • G06V10/473—Contour-based spatial representations, e.g. vector-coding using gradient analysis
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

Based on the Weakly supervised object detection method of certain objects pixel gradient figure, it is related to computer vision technique.It is proposed certain objects pixel gradient figure.In the training process, the certain objects pixel gradient figure of image is extracted.Based on certain objects pixel gradient figure, model can estimate roughly shape and the position of target object；Corresponding image is covered using accumulative certain objects pixel gradient figure, then finely tunes model with the data set covered, it is therefore an objective to allow model that can find the component of more objects；It is proposed a kind of average-maximum value pond neural net layer, this network layer can largely help Weakly supervised target detection.Algorithm does not improve the complexity of network model, without using additional supervision message yet.Largely the experimental results showed that, achieve excellent Weakly supervised target detection and positioning performance.

Description

Weakly supervised object detection method based on certain objects pixel gradient figure
Technical field
The present invention relates to computer vision techniques, more particularly to a kind of based on the Weakly supervised of certain objects pixel gradient figure Object detection method.
Background technique
Target detection is a basic research topic in computer vision field, the main detection different comprising two classes Task: object instance detection and target category detection.The target of first kind Detection task is to identify and position in input picture Know specific some or multiple objects, such as a specific automobile.This generic task be regarded as certain objects sample set and The difference of matching problem in input picture to be detected between certain objects, sample set and the target in input picture to be detected is main Variation from image-forming condition.Second class Detection task concern classification simultaneously positions all possible individuals that predefined classification covers, Such as vehicle detection, pedestrian detection.Compared with object instance Detection task, target category detection has more challenge.Because true The vision difference of many different classes of objects is very little in the world, and the difference between same type objects is not only by slice The influence of part variation, is also influenced by physical object attribute change, for example, in biologically flower be it is extremely various, between individual Color, texture and shape be ever-changing.In real scene target usually only occupy entire scene fraction and May be blocked by other objects or scene in occur visually similar background structure, the appearance of these situations it is also right Object detection task constitutes greatly challenge.
In short, object detection task can be divided into two crucial subtasks: target classification and target positioning.Target classification Task is responsible for judging in input picture whether the object of thoughts category of interest occurs, and a series of label for exporting mixed fractions shows to feel A possibility that object of category of interest appears in input picture.The task of target positioning is responsible for determining class interested in input picture The location and range of other object export bounding box or the object center or the Close edges of object etc. of object, usually rectangular Bounding box is most common selection.
Target detection is the indispensable premise of a large amount of high-level vision tasks, including activity or event recognition, scene content understand Deng.And target detection is also applied to many actual tasks, such as intelligent video monitoring, content-based image retrieval, machine Device people navigation and augmented reality etc..Target detection is of great significance to computer vision field and practical application, several in the past Large quantities of researchers are motivated to pay close attention to simultaneously input research in 10 years.And with powerful machine Learning Theory and signature analysis The development of technology, the relevant research activities of nearly more than ten years target detection project are growing on and on, and have newest research achievement every year It delivers and announces with practical application.Nevertheless, the Detection accuracy of current method is still lower and may not apply to practical logical Detection task.Therefore, target detection is also solved perfectly far away, is still important the research topic of challenge.
The fine supervision message that usual training objective detection needs largely manually to mark: target category label and target position Label.Target category label usually indicates that 1 represents there are corresponding target in figure, and 0 indicates to scheme with only including 0 and 1 vector In be not present corresponding target.And target position label is usually indicated with the form of square enclosure box.Usually only need four Coordinate is assured that a bounding box.This fine target position label usually requires to pay a large amount of manpower and material resources to obtain It takes.Mark deviation can be also introduced in the process of manually mark bounding box and then influences training result.In fact, only having target The data of class label are easier to obtain or mark, for example user uploads image in network, it will usually add to image Attach Title or description.We can obtain the data of a large amount of Weakly supervised label information from internet.Therefore, one naturally Idea is exactly to only use the data of only target category label to carry out training objective detector, this also exactly asking of being studied of the present invention Topic.
Summary of the invention
The purpose of the present invention is to provide the Weakly supervised object detection methods based on certain objects pixel gradient figure.
The present invention includes model training and model reasoning two parts.
The model training the following steps are included:
1) training is finely adjusted to object classifiers using Weakly supervised data set, uses average-maximum pond in training process Change layer；
2) certain objects pixel gradient figure of each classification of input picture based on classification score, certain objects pixel are extracted Gradient map has reacted response of the pixel to certain objects, therefore model can be used the pixel gradient figure of certain objects and estimate roughly Count shape and the position of target object；
3) the certain objects pixel gradient figure for each classification of input picture that adds up, and corresponding image is covered, it obtains new Image；
4) to every image of training dataset, step 2)~3 are repeated), until the classification score to certain objects is less than Threshold value；
5) corresponding image is covered with accumulative certain objects pixel gradient figure, and obtains new Weakly supervised training dataset
6) step 1)~5 are repeated), until new model loss no longer declines.
The model reasoning the following steps are included:
7) input picture obtains the classification score of each classification into model；
8) certain objects pixel gradient figure of each classification of input picture based on classification score is extracted；
9) the certain objects pixel gradient figure for each classification of every image that adds up, and with accumulative certain objects pixel gradient Figure covers corresponding image, obtains new image；
10) step 7)~9 are repeated), until the classification score to certain objects is less than threshold value；
11) component that each object of cumulative pixel gradient figure is extracted using Gauss model, then surrounds square using minimum Shape obtains testing result.
In step 1), the performance of Weakly supervised target detection is improved using average-maximum pond layer, propagated forward and reversely Propagation formula is as follows:
Wherein, the input of average-maximum value pond layer isOutput isWherein, h and w are respectively indicated The height and width of characteristic pattern uIndicate real number range.If input has multiple channels, such asIt then can be following operation Channel executes one by one, finally obtains a vectorSet w={ uij|uij∈ u, uij> ξ, 1≤i≤h, 1≤j≤w }, Wherein, ξ indicates a threshold value, uijIndicate the element of the i-th row in input u, jth column；If the element in u is both less than threshold xi,
ThenIt is defined as global maximum pond:
It is defined as global average pond:
Wherein, u and w respectively indicates the element in u and w, | w | it indicates the number of element in set w, and ξ=0 is set.
In step 2), the pixel gradient figure of the certain objects by extracting each training image, to estimate object The rough shape of body and position:
Wherein,Indicate the penalty values of k-th classification, ykIndicate the label of k-th classification in y, s μ [0,1]KExpression pair The class prediction of image is answered as a result, skIndicate the prediction of k-th classification in s,WithRespectively indicate the mean chart of training dataset Picture and corresponding classification prediction result,It indicatesThe prediction of middle k-th classification, K indicate the class number of data set；Each Image can obtain K certain objects pixel gradient figure, and therefore, image x is relative to lossGradient be:
Wherein, zlIndicate l layers of characteristic pattern,Indicate the gradient map of k-th classification, z1It is exactly x.
In step 3), corresponding input picture is covered according to each certain objects pixel gradient figure:
Or:
Or:
Wherein, t indicates that the number that iteration ingredient excavates, x indicate original input picture, xtIndicate that the t times iteration obtains Image,Indicate the CPG figure of k-th classification,Indicate the CPG figure of k-th classification when the t times iteration,Indicate the The gradient of k-th classification when t iteration；
Then ingredient is iterated to new images to excavate.
In step 5), input picture is covered using the certain objects pixel gradient figure after thresholding, and obtain newly Data set:
Wherein, T indicates the number of iterative learning,When indicating the T times Iterative minor adjustment, the i-th row of image, jth are arranged Pixel, ξ every time at random fromFrom selection,Indicate data set in all pixels average value, ζ indicate 0 to 255 with Machine integer；The data set table that these images form is shown as DT, with current data set DTModel m before fine tuningT-1, and obtain New model mT。
The present invention is a kind of novel Weakly supervised object detection method based on certain objects pixel gradient figure.Many institute's weeks Know, target detection has urgent need to resolve in extremely important status and computer vision field in computer vision field Problem.The main different places of target detection based on Weakly supervised study and the target detection based on supervised learning are to count Fine degree according to collection supervision message is different.Algorithm of target detection based on supervised learning is needed with class label and target The data set of object space label is trained.And the algorithm of target detection based on Weakly supervised study is only used with class label Data set learnt.Because the supervision message amount of class label is the supervision message amount far less than location tags, Only it is known as the algorithm of target detection based on Weakly supervised study with the algorithm of target detection that class label learns.'
The present invention proposes that a kind of algorithm is explored and assists the training of model in conjunction with the information of certain objects pixel gradient figure. Main contents of the invention can summarize following three points:
1. the present invention proposes certain objects pixel gradient figure.In the training process, the present invention extracts the certain objects of image Pixel gradient figure.Based on certain objects pixel gradient figure, model can estimate roughly shape and the position of target object；
2. the present invention covers corresponding image using accumulative certain objects pixel gradient figure, then with the data covered Collection fine tuning model, it is therefore an objective to allow model that can find the component of more objects；
3. the present invention proposes a kind of average-maximum value pond neural net layer, this network layer can be helped largely Weakly supervised target detection.
Algorithm of the invention does not improve the complexity of network model, without using additional supervision message yet.Largely The experimental results showed that method of the invention achieves excellent Weakly supervised target detection and positioning performance.
Detailed description of the invention
Fig. 1 is network structure of the invention.
Fig. 2 is training frame of the invention.
Fig. 3 is input picture, corresponding certain objects pixel gradient figure and bounding box.
Fig. 4 is input picture and corresponding certain objects pixel gradient figure.
Specific embodiment
Following embodiment will the present invention is further illustrated in conjunction with attached drawing.
The main symbol to be used of the definition present invention first.Here it usesIndicate the input figure an of rgb format Picture, y ∈ { 0,1 }KIndicate the class label of correspondence image, whereinIndicate real number range, H and W respectively indicate the height of image And width, K indicate the class number of data set.Each image can obtain K certain objects pixel gradient figure, such as Fig. 1 institute Show.S ∈ [0,1]KIndicate the class prediction result of correspondence image, wherein the class number of K expression data set.Meanwhile it usingWith Respectively indicate training dataset the average image and corresponding classification prediction result.
The present invention uses AlexNet (Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E.Hinton."Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems.2012.)、VGGNet(Simonyan, Karen,and Andrew Zisserman."Very deep convolutionalnetworks for large-scale Image recognition. " ArXiv, 2014.) and GoogLenet (Szegedy, Christian wait " Going deeper with convolutions."Proceedings of the IEEE conference on computer vision and Pattern recognition.2015.) etc. networks as basic model rear end structure.The depth of usual situation drag rear end Deeper, the ability to express of model is also stronger.
The present invention defines a loss:
Wherein, ykIndicate the label of k-th classification in y, skIndicate the prediction of k-th classification in s,It indicatesMiddle k-th The prediction of classification.Therefore, image x is relative to lossGradient be:
Wherein, zlIndicate l layers of characteristic pattern,Indicate the gradient map of k-th classification,Indicate k-th classification Penalty values, z1It is exactly x.
Then these gradients are carried out last certain objects gradient map (the Object-Specific Pixel of accumulative acquisition Gradient map), abbreviation CPG figure:
Wherein, t indicates the number that iteration ingredient excavates.The process that iteration ingredient excavates is as shown in Fig. 2 top half.
Corresponding input picture is covered using each certain objects pixel gradient figure:
Or:
Or:
Wherein, x indicates original input picture, xtIndicate the image that the t times iteration obtains,Indicate k-th classification CPG figure,Indicate the CPG figure of k-th classification when the t times iteration,Indicate the ladder of k-th classification when the t times iteration Degree.
Xt+1Input model obtains classification prediction result, if classification prediction result is higher than certain threshold value, extracts again Certain objects pixel gradient figure；Otherwise certain objects pixel gradient figure is just no longer extracted.Effect such as Fig. 3 institute that iteration ingredient excavates Show.
After obtaining certain objects pixel gradient figure, certain objects pixel gradient figure is filtered with a threshold value:
Wherein, T indicates the number of iterative learning,When indicating the T times Iterative minor adjustment, k-th classification certain objects The element of the i-th row jth column of pixel gradient figure.The process of iterative learning is as shown in the lower half portion Fig. 2.
Input picture is covered using the certain objects pixel gradient figure after thresholding:
Wherein,When indicating the T times Iterative minor adjustment, the pixel of the i-th row of image, jth column, ξ every time at random fromMiddle selection,Indicate that the average value of all pixels in data set, ζ indicate 0 to 255 random integers.These figures As the data set table of composition is shown as DT.With current data set DTModel m before fine tuningT-1, and obtain new model mT.It is newest Model combination is defined as:
Wherein, M is defined as M={ MT-1, mT, max { MT-1, mT, and M0=m0, m0As original disaggregated model.Together When, Q (m) is error of the model m on verifying collection.Above-mentioned fine tuning iterative process iteration always, until Q (mT)=Q (MT-1)。
Meanwhile a kind of average-maximum value pond layer is also added in the training process and examines to improve Weakly supervised target by the present invention The performance of survey.Defining the input of averagely-maximum value pond layer first isOutput isWherein, h and w difference Indicate the height and width of characteristic pattern u,Indicate real number range.If input has multiple channels, such asWherein c is indicated Following operation channel execution one by one then can finally be obtained a vector by the number in channel
Define a new set w={ uij|uij∈ u, uij> ξ, 1≤i≤h, 1≤j≤w }, wherein ξ indicates a threshold Value, uijIndicate the element of the i-th row in input u, jth column.If the element in input u is both less than threshold xi, gatherIt is flat The output of-maximum value pond layer are as follows:
Wherein,It is defined as global maximum pond:
It is defined as global average pond:
Wherein, u and w respectively indicates the element in u and w, | w | indicate the number of element in set w.
The gradient calculation method of average-maximum value pond layer is as follows:
Wherein, uij∈ u, 1≤i≤h, 1≤j≤w, and ξ=0 is set.
In model measurement, it is only necessary to which, test image input model, model output category score, first calculating add up Certain objects pixel gradient figure, then thresholding, finally obtains the result of target detection with minimum area-encasing rectangle frame.Fig. 4 is shown The CPG figure for the parts of images that final mask obtains.

Claims (5)

1. the Weakly supervised object detection method based on certain objects pixel gradient figure, it is characterised in that including model training and model Reasoning two parts；
The model training the following steps are included:
1) training is finely adjusted to object classifiers using Weakly supervised data set, uses average-maximum pond layer in training process；
2) certain objects pixel gradient figure of each classification of input picture based on classification score, certain objects pixel gradient are extracted Figure has reacted response of the pixel to certain objects, therefore model can be used the pixel gradient figures of certain objects and carry out rough estimate mesh Mark shape and the position of object；
3) the certain objects pixel gradient figure for each classification of input picture that adds up, and corresponding image is covered, obtain new image；
4) to every image of training dataset, step 2)~3 are repeated), until the classification score to certain objects is less than threshold value；
5) corresponding image is covered with accumulative certain objects pixel gradient figure, and obtains new Weakly supervised training dataset
6) step 1)~5 are repeated), until new model loss no longer declines；
The model reasoning the following steps are included:
7) input picture obtains the classification score of each classification into model；
8) certain objects pixel gradient figure of each classification of input picture based on classification score is extracted；
9) the certain objects pixel gradient figure for each classification of every image that adds up, and covered with accumulative certain objects pixel gradient figure Corresponding image is covered, new image is obtained；
10) step 7)~9 are repeated), until the classification score to certain objects is less than threshold value；
11) component that each object of cumulative pixel gradient figure is extracted using Gauss model, is then obtained using minimum area-encasing rectangle Obtain testing result.
2. the Weakly supervised object detection method as described in claim 1 based on certain objects pixel gradient figure, it is characterised in that In step 1), the performance of Weakly supervised target detection is improved using average-maximum pond layer, propagated forward and backpropagation formula are such as Under:
Wherein, the input of average-maximum value pond layer isOutput isWherein, h and w respectively indicate feature Scheme the height and width of u,Indicate real number range；If input has multiple channels, such asThen following operation one one A channel executes, and finally obtains a vectorSet w={ uij|uij∈ u, uij> ξ, 1≤i≤h, 1≤j≤w }, wherein ξ Indicate a threshold value, uijIndicate the element of the i-th row in input u, jth column；If the element in u is both less than threshold xi, It is defined as global maximum pond:
It is defined as global average pond:
Wherein, u and w respectively indicates the element in u and w, | w | it indicates the number of element in set w, and ξ=0 is set.
3. the Weakly supervised object detection method as described in claim 1 based on certain objects pixel gradient figure, it is characterised in that In step 2), the pixel gradient figure of the certain objects by extracting each training image estimates the rough shape of target object The position and:
Wherein,Indicate the penalty values of k-th classification, ykIndicate the label of k-th classification in y, s ∈ [0,1]KIndicate corresponding diagram The class prediction of picture is as a result, skIndicate the prediction of k-th classification in s,WithRespectively indicate training dataset the average image and Corresponding classification prediction result,It indicatesThe prediction of middle k-th classification, K indicate the class number of data set；Each image K certain objects pixel gradient figure is obtained, therefore, image x is relative to lossGradient be:
Wherein, zlIndicate l layers of characteristic pattern,Indicate the gradient map of k-th classification, z1It is exactly x.
4. the Weakly supervised object detection method as described in claim 1 based on certain objects pixel gradient figure, it is characterised in that In step 3), corresponding input picture is covered according to each certain objects pixel gradient figure:
Or:
Or:
Wherein, t indicates that the number that iteration ingredient excavates, x indicate original input picture, xtIndicate the figure that the t times iteration obtains Picture,Indicate the CPG figure of k-th classification,Indicate the CPG figure of k-th classification when the t times iteration,It indicates the t times The gradient of k-th classification when iteration；
Then ingredient is iterated to new images to excavate.
5. the Weakly supervised object detection method as described in claim 1 based on certain objects pixel gradient figure, it is characterised in that In step 5), input picture is covered using the certain objects pixel gradient figure after thresholding, and obtain new data set:
Wherein, T indicates the number of iterative learning,When indicating the T times Iterative minor adjustment, the pixel that the i-th row of image, jth arrange, ξ every time at random fromFrom selection,Indicate data set in all pixels average value, ζ indicate 0 to 255 it is random whole Number；The data set table that image forms is shown as DT, with current data set DTModel m before fine tuningT-1, and obtain new model mT。
CN201810877293.XA 2018-08-03 2018-08-03 Weakly supervised object detection method based on certain objects pixel gradient figure Pending CN109034258A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810877293.XA CN109034258A (en) 2018-08-03 2018-08-03 Weakly supervised object detection method based on certain objects pixel gradient figure

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810877293.XA CN109034258A (en) 2018-08-03 2018-08-03 Weakly supervised object detection method based on certain objects pixel gradient figure

Publications (1)

Publication Number Publication Date
CN109034258A true CN109034258A (en) 2018-12-18

Family

ID=64649281

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810877293.XA Pending CN109034258A (en) 2018-08-03 2018-08-03 Weakly supervised object detection method based on certain objects pixel gradient figure

Country Status (1)

Country Link
CN (1) CN109034258A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993742A (en) * 2019-04-04 2019-07-09 哈尔滨工业大学 Bridge Crack method for quickly identifying based on diagonal operator reciprocal
CN110287970A (en) * 2019-06-25 2019-09-27 电子科技大学 A kind of Weakly supervised object positioning method based on CAM and cover
CN111523585A (en) * 2020-04-16 2020-08-11 厦门大学 Weak supervision target detection method based on improved depth residual error network
CN112182268A (en) * 2020-09-27 2021-01-05 北京达佳互联信息技术有限公司 Image classification method and device, electronic equipment and storage medium
CN115439688A (en) * 2022-09-01 2022-12-06 哈尔滨工业大学 Weak supervision object detection method based on surrounding area perception and association

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102654902A (en) * 2012-01-16 2012-09-05 江南大学 Contour vector feature-based embedded real-time image matching method
US20130195365A1 (en) * 2012-02-01 2013-08-01 Sharp Laboratories Of America, Inc. Edge based template matching
CN108062574A (en) * 2017-12-31 2018-05-22 厦门大学 A kind of Weakly supervised object detection method based on particular category space constraint

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102654902A (en) * 2012-01-16 2012-09-05 江南大学 Contour vector feature-based embedded real-time image matching method
US20130195365A1 (en) * 2012-02-01 2013-08-01 Sharp Laboratories Of America, Inc. Edge based template matching
CN108062574A (en) * 2017-12-31 2018-05-22 厦门大学 A kind of Weakly supervised object detection method based on particular category space constraint

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
YUNHANG SHEN 等: "Weakly Supervised Object Detection via Object-Specific Pixel Gradient", 《IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS》 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993742A (en) * 2019-04-04 2019-07-09 哈尔滨工业大学 Bridge Crack method for quickly identifying based on diagonal operator reciprocal
CN109993742B (en) * 2019-04-04 2020-03-17 哈尔滨工业大学 Bridge crack rapid identification method based on diagonal derivative operator
CN110287970A (en) * 2019-06-25 2019-09-27 电子科技大学 A kind of Weakly supervised object positioning method based on CAM and cover
CN110287970B (en) * 2019-06-25 2021-07-27 电子科技大学 Weak supervision object positioning method based on CAM and covering
CN111523585A (en) * 2020-04-16 2020-08-11 厦门大学 Weak supervision target detection method based on improved depth residual error network
CN111523585B (en) * 2020-04-16 2022-05-31 厦门大学 Weak supervision target detection method based on improved depth residual error network
CN112182268A (en) * 2020-09-27 2021-01-05 北京达佳互联信息技术有限公司 Image classification method and device, electronic equipment and storage medium
CN112182268B (en) * 2020-09-27 2024-04-05 北京达佳互联信息技术有限公司 Image classification method, device, electronic equipment and storage medium
CN115439688A (en) * 2022-09-01 2022-12-06 哈尔滨工业大学 Weak supervision object detection method based on surrounding area perception and association

Similar Documents

Publication Publication Date Title
US10592780B2 (en) Neural network training system
CN109034258A (en) Weakly supervised object detection method based on certain objects pixel gradient figure
CN108764085B (en) Crowd counting method based on generation of confrontation network
CN106611420B (en) The SAR image segmentation method constrained based on deconvolution network and sketch map direction
Ghamisi et al. Multilevel image segmentation based on fractional-order Darwinian particle swarm optimization
CN106951830B (en) Image scene multi-object marking method based on prior condition constraint
Văduva et al. Latent Dirichlet allocation for spatial analysis of satellite images
Oliva et al. Scene-centered description from spatial envelope properties
Li et al. Adaptive deep convolutional neural networks for scene-specific object detection
CN101211356A (en) Image inquiry method based on marking area
Shahab et al. How salient is scene text?
CN104715251B (en) A kind of well-marked target detection method based on histogram linear fit
CN114022759A (en) Airspace finite pixel target detection system and method fusing neural network space-time characteristics
CN106897681A (en) A kind of remote sensing images comparative analysis method and system
CN110222572A (en) Tracking, device, electronic equipment and storage medium
Grigorev et al. Depth estimation from single monocular images using deep hybrid network
Yadav et al. An improved deep learning-based optimal object detection system from images
CN111582410B (en) Image recognition model training method, device, computer equipment and storage medium
CN107423771B (en) Two-time-phase remote sensing image change detection method
Marvaniya et al. Small, sparse, but substantial: techniques for segmenting small agricultural fields using sparse ground data
Shaikh et al. A contemporary approach for object recognition based on spatial layout and low level features’ integration
Wang et al. Salient object detection using biogeography-based optimization to combine features
Chen et al. Edge Enhanced GCIFFNet: A Multiclass Semantic Segmentation Network Based on Edge Enhancement and Multiscale Attention Mechanism
CN109492530B (en) Robust visual object tracking method based on depth multi-scale space-time characteristics
Zhai et al. GAN-BiLSTM network for field-road classification on imbalanced GNSS recordings

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20181218