CN107423760A - Based on pre-segmentation and the deep learning object detection method returned - Google Patents

Based on pre-segmentation and the deep learning object detection method returned Download PDF

Info

Publication number
CN107423760A
CN107423760A CN201710598875.XA CN201710598875A CN107423760A CN 107423760 A CN107423760 A CN 107423760A CN 201710598875 A CN201710598875 A CN 201710598875A CN 107423760 A CN107423760 A CN 107423760A
Authority
CN
China
Prior art keywords
mrow
frame
msub
acquiescence
msubsup
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
CN201710598875.XA
Other languages
Chinese (zh)
Inventor
孙伟
潘蓉
卞磊
王鹏
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.)
Xidian University
Original Assignee
Xidian 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 Xidian University filed Critical Xidian University
Priority to CN201710598875.XA priority Critical patent/CN107423760A/en
Publication of CN107423760A publication Critical patent/CN107423760A/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
    • G06F18/21—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
    • 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
    • G06F18/2163—Partitioning the feature space
    • 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
    • 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
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • 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
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of object detection method based on pre-segmentation and the deep learning returned, mainly solve the problems, such as that existing object detection method is long to small target deteection low precision and detection time.Its implementation is：1) area-of-interest of image to be detected is extracted using Quadtree Partition algorithm；2) feature extraction is carried out to area-of-interest using basic convolutional layer and auxiliary convolutional layer, obtains the characteristic pattern of multiple yardsticks；3) positional information of acquiescence frame is calculated on the characteristic pattern of multiple yardsticks, is detected using convolution filter on the characteristic pattern of multiple yardsticks, obtains multiple prediction frames and multiple category scores；4) bezel locations and classification information of final target are obtained to multiple prediction frames and multiple category scores using non-maxima suppression.The present invention can fast and accurately be detected to the Small object in image, be detected in real time available for the target in unmanned plane.

Description

Based on pre-segmentation and the deep learning object detection method returned
Technical field
The invention belongs to Image Information Processing field, specifically a kind of deep learning object detection method, can be used for Accurate positioning in real time and classification to target.
Background technology
Target detection is a challenging problem in computer vision field, and its core missions is in static images Or certain Target Recognition Algorithms and search strategy are used in video, obtain specific objective position in an image or a video and class Not.The method of target detection is broadly divided into the algorithm of target detection of feature based and machine learning and based on deep learning at present Detection method.The method of wherein feature based and machine learning is by carrying out regional choice, feature extraction, grader to target The processes such as classification realize target detection.Regional choice is to carry out traversal to entire image by sliding window and select to there may be mesh Target frame, but time complexity is too high, redundancy window are made a slip of the tongue more, directly affects the speed and performance of feature extraction and classification. The feature commonly used in feature extraction has Haar wavelet characters, HOG features, SIFT feature and composite character etc., due to the light of image According to condition, the diversity of form of background and target etc. is higher to the robustness requirement of feature, and the feature quality of extraction is directly Influence the precision of target classification.Traditional grader mainly includes support vector machines and iterator Adaboost.Due to being pin To the identification mission of some feature, model generalization ability, it is difficult to be identified in actual applications to targeting accuracy.Opened from 2014 Begin, the algorithm of target detection based on deep learning achieves great breakthrough, overcomes lacking in traditional algorithm of target detection Point.The algorithm of target detection based on deep learning of main flow is broadly divided into two classes at present：Deep learning mesh based on candidate region Mark detection algorithm and the deep learning algorithm of target detection based on recurrence.The representative of algorithm of target detection based on candidate region is The R-CNN algorithms that R Girshick are proposed, the detection framework combination candidate region of the algorithm and convolutional neural networks CNN are divided Class.SPP-NET is generated by the speed-raising of R-CNN successive optimizations, Fast R-CNN and Faster R-CNN, the precision of target detection and Speed all improves a lot, but due to being divided into positioning and two steps of classification and positioning consumption when such method carries out target detection Shi Taichang, therefore target detection can not be carried out in real time.Deep learning algorithm of target detection based on homing method is representational There are YOLO and SSD, this kind of algorithm is mainly position and the classification for predicting target directly from image to be detected by the Return Law, this Kind method causes target detection speed to greatly speed up, and can reach the requirement of real-time target detection, but to the size of input picture There is strict demand and target location positioning is poor, Small object that can not be in detection image.YOLO and SSD300 requires defeated respectively It is 448*448 and 300*300 to enter picture size, reduces image to be detected to specific dimensions, can lose image detail, lead to not Detect Small object.
The content of the invention
It is an object of the invention to for above-mentioned existing technical problem, propose a kind of based on pre-segmentation and the depth returned The object detection method of study, to preserve image detail, improve the Real time detection performance to Small object.
The technical thought of the present invention is to obtain area-of-interest by carrying out quaternary tree pre-segmentation in the input image；Pass through The Analysis On Multi-scale Features figure of multiple dimensioned convolutional layer extraction area-of-interest；Target classification and prediction frame are predicted by convolution filter Position；Final target classification and target location coordinate is obtained by non-maxima suppression.
According to above-mentioned thinking, implementation of the invention includes as follows：
(1) established according to QuadTree algorithm and convolutional neural networks VGG-16 based on pre-segmentation and the deep learning net returned Network model；
(2) network model built is trained on training set of images；
(2a) uses image set PASCAL VOC2007 and PASCAL VOC2012 training dataset to be used as training set Image set PASCAL VOC2007 test data set is as test set；
The acquiescence frame on characteristic pattern generated in the mark frame and network model of (2b) to marking image in training set Matched；
The target loss function L (x, l, c, g) of (2c) tectonic network model；
Wherein, x is characterized the acquiescence frame on figure, and l is prediction block, and g is mark frame, and c is characterized the acquiescence side on figure Category score set of the frame in each classification, LconfAcquiescence frame on (x, c) expression characteristic pattern is on category score set c Softmax Classification Loss functions, Lloc(x, l, g) represents positioning loss function, and N represents the acquiescence side matched with mark frame Frame number, parameter alpha are arranged to 1 by cross validation；
(2d) minimizes loss function using gradient descent method and the weight parameter in network is successively reversely adjusted simultaneously, obtains To the network model trained；
(3) original image to be detected is input in the network model trained, obtains the target in image to be detected Classification and position coordinates.
The invention has the advantages that：
1) present invention due to image to be detected carry out area-of-interest pre-segmentation, avoid because picture is excessive and without legal The problem of position Small object；
2) due to only carrying out feature extraction to area-of-interest in the present invention, rather than feature extraction is carried out to whole image, Amount of calculation when reducing feature extraction and calculate the time.
3) feature of the present invention due to extracting area-of-interest using convolutional layer, its feature have displacement, rotation and scaling Consistency, the problem of avoiding engineer's poor robustness, it is more suitable for target detection.
4) present invention is predicted by convolution filter to characteristic pattern, is obtained the other confidence level of a series of target class and is obtained Divide the position coordinates with target, improve computational efficiency.
Brief description of the drawings
Fig. 1 is the implementation process figure of the present invention；
Fig. 2 is the network structure built in the present invention；
Image to be detected that Fig. 3 is used when being and being tested in the present invention；
Fig. 4 is the area-of-interest figure extracted in the present invention using QuadTree algorithm；
Fig. 5 is the simulation result figure for carrying out target detection to image to be detected with the present invention.
Embodiment
Reference picture 1, step is as follows for of the invention realizing：
Step 1, establish based on pre-segmentation and the deep learning network model returned.
The target detection network for being currently based on deep learning is divided into two major classes：One kind is the deep learning based on candidate region Target detection network, such as R-CNN, Fast R-CNN and Faster R-CNN；Another kind of is the deep learning mesh based on recurrence Mark detection network, such as YOLO and SSD, the present invention are proposed based on pre-segmentation and the deep learning object detection method returned.Mesh The method of preceding extraction area-of-interest includes：Area-of-interest exacting method based on threshold value, based on the interested of edge extracting Method for extracting region, the area-of-interest exacting method based on Quadtree Partition, the region of interesting extraction based on region growing Method etc., the region of interesting extraction method structure pre-segmentation Internet of Quadtree Partition is used in of the invention.
Reference picture 2, this step are implemented as follows
(1a) utilizes QuadTree algorithm structure area-of-interest pre-segmentation Internet；
(1a1) sets segmentation threshold in QuadTree algorithm as M, and maximum fractionation number is Q=1024, by image to be detected According to four sub-regions are horizontally and vertically divided into, wherein 0 ＜ M ＜ 255；
(1a2) calculates the average gray value of every sub-regions after segmentation, and subregion of the average gray value more than M is continued Four sub-regions are divided into, stop segmentation when the average gray value of subregion is less than M or segmentation times reach Q, and remember Record its positional information；
(1a3) finds the minimum sub-district positioned at image to be detected upper left corner and the lower right corner according to the positional information of subregion The position of area-of-interest in the coordinate in domain, as image to be detected；
(1b) establishes target detection Internet according to convolutional neural networks VGG-16；
Convolutional neural networks currently used for target identification have AlexNet, VGG-16, GoogLeNet, ResNet etc., this Target detection network is established using convolutional neural networks VGG-16 in invention, implementation step is as follows：
(1b1) uses the stage stage1-stage5 in convolutional neural networks VGG-16 as target detection Internet Basic convolutional layer, and full articulamentum fc6, fc7 therein are replaced with into two convolutional layers conv6, conv7, while add four The new auxiliary convolutional layer of convolutional layer conv8, conv9, conv10, conv11 as target detection Internet, four newly added The size of convolutional layer is respectively 10 × 10,5 × 5,3 × 3,1 × 1；
(1b2) forms the detection layers of target detection network using a series of convolution filter；
(1b3) is pressed down using the non-maximum in the target detection network R-CNN based on candidate region and convolutional neural networks Preparative layer forms the output layer of target detection Internet.
Step 2, the network model built is trained on training set of images.
The method being trained at present to deep learning network is broadly divided into two classes：From the unsupervised learning of lower rising and oneself Downward supervised learning is pushed up, is trained in the present invention using top-down supervised learning method, realizes that step is as follows：
(2a) selects the training set of images for training；
The image set for being usually used in target detection network training has：Imagenet image sets, PASCAL VOC image sets, COCO Image set etc., the present invention use image set PASCAL VOC2007 and PASCAL VOC2012 training dataset as training Collection, test set is used as by the use of image set PASCAL VOC2007 test data set；
The acquiescence frame on characteristic pattern generated in the mark frame and network model of (2b) to marking image in training set Matched；
(2b1) calculates size and the position of the acquiescence frame on characteristic pattern：
The ratio of width to height of setting acquiescence frame has 5 kinds of different values, respectively a={ 1,2,3,1/2,1/3 }, calculates k-th The ratio of width to height is a in characteristic patternτAcquiescence frame widthAnd height
Wherein aτFor τ kind the ratio of width to height, 0≤τ≤5, sminRepresent the length of side of acquiescence frame and the minimum ratio of input picture Value, smaxThe length of side of acquiescence frame and the maximum ratio of input picture are represented, k ∈ [1, E], E represent the characteristic pattern in network model Number；
The centre coordinate (xcen, ycen) of acquiescence frame on k-th of characteristic pattern isWherein | fk| it is the size of k-th of characteristic pattern, u, v represent the coordinate at characteristic pattern midpoint, u, v ∈ [0, | fk|]；
(2b2) is obtained marking frame and is given tacit consent to the Jaccard between frame according to the size and centre coordinate of acquiescence frame Overlap coefficient, it is positive sample Pos to select acquiescence frame of the Jaccard overlap coefficients value more than 0.5, and other are negative sample Neg；
(2b21) calculate the acquiescence frame x upper left corner coordinate (xleft, yleft) and the lower right corner coordinate (xrigh, yrigh)：
(2b22) calculates coordinate (xmin, ymin) and the lower right corner in the upper left corner of the acquiescence frame with marking frame intersection Coordinate (xmax, ymax)：
Xmin=max (xleft, xgleft),
Ymin=max (yleft, ygleft),
Xmax=max (xrigh, xgrigh),
Ymax=max (yrigh, ygrigh)；
Wherein (xgleft, ygleft) and (xgrigh, ygrigh) represent to mark respectively the coordinate in the frame g upper left corner and The coordinate in the lower right corner；
(2b23) calculates acquiescence frame x with marking the area inter (x, g) of frame g intersections:
Inter (x, g)=(max (ymax-ymin, 0)) * (max (xmax-xmin, 0))；
(2b24) calculates acquiescence frame x and marks the Jaccard overlap coefficients between frame g：
It is positive sample Pos to select acquiescence frame of J (x, the g) coefficient value more than 0.5, and other are negative sample Neg, completes mark The matching of frame and acquiescence frame.
The target loss function L (x, l, c, g) of (2c) tectonic network model:
(2c1) is predicted using convolution filter on characteristic pattern, obtains giving tacit consent to classification of the frame on all categories Score set c and prediction frame are relative to position offset (Δ x, Δ y, the Δ w, Δ h), wherein (Δ x, Δ y) for giving tacit consent to frame Offset of the prediction frame centre coordinate relative to acquiescence frame centre coordinate is represented, Δ w represents that prediction frame is wide relative to silent Recognize the wide offset of frame, Δ h represents prediction frame height relative to the high offset of acquiescence frame；
The classification score set c of (2c2) according to the acquiescence frame on characteristic pattern on all categories, calculate softmax points Class loss function Lconf(x,c)：
Wherein, whenRepresent that i-th of acquiescence frame matches with j-th of mark frame that classification is p,Represent I-th of acquiescence frame mismatches with j-th of mark frame that classification is p, and 0≤i≤N, N represent the acquiescence matched with mark frame Frame number, 1≤p≤H, H are total categorical measure, and 0≤j≤T, T are the quantity of mark frame,Represent i-th in positive sample The average on all categories of individual acquiescence frame,Represent i-th in negative sample2Individual acquiescence frame is in all categories On average, 0≤i2≤N2, N2Represent and the unmatched acquiescence frame number of mark frame；
(2c3) calculates positioning loss function Lloc(x,l,g)：
Wherein (cx, cy) is by (centre coordinate of the acquiescence frame x after Δ x, Δ y) compensation, w, h are by (Δ w, Δ H) the wide and high of frame is given tacit consent to after compensating,I-th of prediction frame that offset is m is represented,Represent the jth that offset is m Individual prediction frame；
(2c4) is according to Classification Loss function Lconf(x, c) and positioning loss function Lloc(x, l, g), obtain target loss letter Number L (x, l, c, g)：
Wherein, x is characterized the acquiescence frame on figure, and l is prediction block, and g is mark frame, and c is characterized the acquiescence side on figure Category score set of the frame in each classification, LconfAcquiescence frame on (x, c) expression characteristic pattern is on category score set c Softmax Classification Loss functions, Lloc(x, l, g) represents positioning loss function, and N represents the acquiescence side matched with mark frame Frame number, parameter alpha are arranged to 1 by cross validation；
(2d) minimizes loss function using gradient descent method, while the weight parameter in network is successively reversely adjusted, The network model trained.
Step 3, original image to be detected is input in the network model trained, obtains the mesh in image to be detected Mark classification and position coordinates.
(3a) carries out Quadtree Partition in pre-segmentation layer to input picture, extracts area-of-interest；
(3b) carries out feature extraction using basic convolutional layer and auxiliary convolutional layer to area-of-interest, obtains multiple yardsticks Characteristic pattern；
(3c) calculates the position coordinates of acquiescence frame on the characteristic pattern of multiple yardsticks
(3d) is predicted on the characteristic pattern of multiple yardsticks using convolution filter, obtains the mesh in multiple prediction frames Mark category score and predict the position offset of the relative acquiescence frame of frame；
(3e) gives tacit consent to frame relatively using non-maxima suppression to the target classification in multiple prediction frames and prediction frame Position offset suppressed, obtain the position of the target classification and the relative acquiescence frame of prediction frame in final prediction frame Offset is put, and according to position offset (Δ xfinal, Δ yfinal, Δ wfinal, the Δ of the relative acquiescence frame of prediction frame Hfinal) and acquiescence frame position coordinatesObtain the position coordinates of prediction frame
The effect of the present invention can be further illustrated by following experiment.
1. experimental subjects
Four image to be detected a, b, c, d shown in test data set and Fig. 3 that experimental subjects is PASCAL VOC2007；
2. experimental procedure
(2.1) respectively using Fast-RCNN network models, Faster-RCNN network models, YOLO network models, The training of SSD300 network models and network model of the invention in image set PASCAL VOC2007 and PASCAL VOC2012 Training pattern on collection；
(2.2) respectively use (2.1) in train Fast-RCNN network models, Faster-RCNN network models, The network model of YOLO network models, SSD300 network models and the present invention is carried out in PASCAL VOC2007 test data sets Test, accuracy of detection and the detection speed for obtaining network model are as shown in table 1；
(2.3) model trained using the present invention carries out target detection successively on four pictures shown in Fig. 3, wherein The result of area-of-interest is extracted to image to be detected as shown in figure 4, final testing result is as shown in figure 5, wherein：
Fig. 5 a carry out the simulation result of target detection with the present invention to Fig. 3 a, and target classification is cat, and target location is frame；
Fig. 5 b carry out the simulation result of target detection with the present invention to Fig. 3 b, and target classification is ship, and target location is frame；
Fig. 5 c carry out the simulation result of target detection with the present invention to Fig. 3 c, and target classification is aircraft, and target location is side Frame；Fig. 5 d carry out the simulation result of target detection with the present invention to Fig. 3 d, and target classification is cat, and target location is frame.
Accurate positioning of the network model to Small object, the classification of the present invention is can be seen that by Fig. 5 b and Fig. 5 c testing result Accurately.
3. experimental data counts：
Respectively using train Fast-RCNN network models, Faster-RCNN network models, YOLO network models, SSD300 network models and the network model of the present invention are tested in PASCAL VOC2007 test data sets, obtained inspection Survey precision and detection speed is as shown in table 1：
Table 1
Algorithm model Training dataset Accuracy of detection (%) Detection speed (frame/second)
Fast-RCNN 07++12 68.4 3
Faster-RCNN 07++12 70.4 5
YOLO 07++12 57.9 47
SSD300 07++12 72.4 59
Context of methods 07++12 74.9 45
As it can be seen from table 1 accuracy of detection and detection speed ratio that the network model of the present invention is tested on test set Fast-RCNN network models, the accuracy of detection of Faster-RCNN network models and detection speed are all significantly increased, with SSD300 network models, YOLO network models are compared, and network model of the invention can detect while detection speed is ensured Precision improves.The rate request detected in real time is greater than that 25 frames are per second, and it is per second that detection speed of the invention reaches 45 frames, meets real When testing requirements.

Claims (5)

1. based on pre-segmentation and the deep learning object detection method returned, including：
(1) established according to QuadTree algorithm and convolutional neural networks VGG-16 based on pre-segmentation and the deep learning network mould returned Type；
(2) network model built is trained on training set of images；
(2a) uses image set PASCAL VOC2007 and PASCAL VOC2012 training dataset to use image as training set Collect PASCAL VOC2007 test data set as test set；
The acquiescence frame on characteristic pattern generated in the mark frame and network model of (2b) to marking image in training set is carried out Matching；
The target loss function L (x, l, c, g) of (2c) tectonic network model；
<mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>l</mi> <mo>,</mo> <mi>c</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <mrow> <mo>(</mo> <msub> <mi>L</mi> <mrow> <mi>c</mi> <mi>o</mi> <mi>n</mi> <mi>f</mi> </mrow> </msub> <mo>(</mo> <mrow> <mi>x</mi> <mo>,</mo> <mi>c</mi> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>&amp;alpha;L</mi> <mrow> <mi>l</mi> <mi>o</mi> <mi>c</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>l</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> </mrow>
Wherein, x is characterized the acquiescence frame on figure, and l is prediction block, and for mark frame, the acquiescence frame that c is characterized on figure exists g Category score set in each classification, LconfAcquiescence frame on (x, c) expression characteristic pattern is on category score set c Softmax Classification Loss functions, Lloc(x, l, g) represents positioning loss function, and N represents the acquiescence frame matched with mark frame Number, parameter alpha are arranged to 1 by cross validation；
(2d) minimizes loss function using gradient descent method and the weight parameter in network is successively reversely adjusted simultaneously, is instructed The network model perfected；
(3) original image to be detected is input in the network model trained, obtains the target classification in image to be detected And position coordinates.
2. according to the method for claim 1, wherein step (1) is built according to QuadTree algorithm and convolutional neural networks VGG-16 Be based on pre-segmentation and the deep learning network model returned, carries out in accordance with the following steps：
(1a) utilizes QuadTree algorithm structure area-of-interest pre-segmentation Internet：
(1a1) sets segmentation threshold in QuadTree algorithm as M, and maximum fractionation number is Q, by image to be detected according to level side It is divided into four sub-regions to vertical direction；
(1a2) calculates the average gray value of every sub-regions after segmentation, and subregion of the average gray value more than M is continued to split For four sub-regions, stop segmentation when the average gray value of subregion is less than M or segmentation times reach Q, and record it Positional information；
(1a3) is found positioned at the minimum subregion in image to be detected upper left corner and the lower right corner according to the positional information of subregion The position of area-of-interest in coordinate, as image to be detected；
(1b) establishes target detection Internet according to convolutional neural networks VGG-16：
(1b1) uses bases of the stage stage1-stage5 as target detection Internet in convolutional neural networks VGG-16 Convolutional layer, and full articulamentum fc6, fc7 therein are replaced with into two convolutional layers, while four new convolutional layers are added as target Detect the auxiliary convolutional layer of Internet；
(1b2) forms the detection layers of target detection network using a series of convolution filter；
(1b3) uses the non-maxima suppression layer in the target detection network R-CNN based on candidate region and convolutional neural networks Form the output layer of target detection Internet.
3. according to the method for claim 1, to the mark frame and net of mark image in training set wherein in step (2b) Acquiescence frame on the characteristic pattern generated in network model is matched, and is carried out in accordance with the following steps：
(2b1) calculates size and the position of the acquiescence frame on characteristic pattern：
Setting the ratio of width to height for giving tacit consent to frame on each characteristic pattern has 5 kinds of different ratios, respectively a={ 1,2,3,1/2,1/ 3 }, it is a to calculate the ratio of width to height in k-th of characteristic patternτAcquiescence frame widthAnd height
<mfenced open = "" close = ""> <mtable> <mtr> <mtd> <mrow> <msubsup> <mi>w</mi> <mi>k</mi> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msubsup> <mo>=</mo> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>min</mi> </msub> <mo>+</mo> <mfrac> <mrow> <msub> <mi>s</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>s</mi> <mi>min</mi> </msub> </mrow> <mrow> <mi>m</mi> <mo>-</mo> <mn>1</mn> </mrow> </mfrac> <mo>(</mo> <mrow> <mi>k</mi> <mo>-</mo> <mn>1</mn> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <msqrt> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msqrt> <mo>,</mo> </mrow> </mtd> <mtd> <mrow> <msubsup> <mi>h</mi> <mi>k</mi> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msubsup> <mo>=</mo> <mfrac> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>min</mi> </msub> <mo>+</mo> <mfrac> <mrow> <msub> <mi>s</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>s</mi> <mi>min</mi> </msub> </mrow> <mrow> <mi>m</mi> <mo>-</mo> <mn>1</mn> </mrow> </mfrac> <mo>(</mo> <mrow> <mi>k</mi> <mo>-</mo> <mn>1</mn> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <msqrt> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msqrt> </mfrac> </mrow> </mtd> </mtr> </mtable> </mfenced>
Wherein aτFor τ kind the ratio of width to height, 1≤τ≤5, sminRepresent the length of side of acquiescence frame and the side of input picture that the ratio of width to height is 1 Long minimum ratio, smaxRepresent the maximum ratio of the length of side for the acquiescence frame that the ratio of width to height is 1 and the length of side of input picture, k ∈ [1, E], E represent the characteristic pattern number in network model；
The centre coordinate (xcen, ycen) for calculating the acquiescence frame on k-th of characteristic pattern isWherein (u, V) coordinate at expression characteristic pattern midpoint, and u, v ∈ [0, | fk|], | fk| it is the size of k-th of characteristic pattern；
(2b2) calculates acquiescence frame x and mark according to the size of acquiescence frame and the position coordinates of centre coordinate and mark frame Jaccard overlap coefficients between frame g：
(2b21) calculate the acquiescence frame x upper left corner coordinate (xleft, yleft) and the lower right corner coordinate (xrigh, yrigh)：
<mrow> <mi>x</mi> <mi>l</mi> <mi>e</mi> <mi>f</mi> <mi>t</mi> <mo>=</mo> <mi>x</mi> <mi>c</mi> <mi>e</mi> <mi>n</mi> <mo>-</mo> <msubsup> <mi>w</mi> <mi>k</mi> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msubsup> <mo>/</mo> <mn>2</mn> <mo>,</mo> </mrow>
<mrow> <mi>y</mi> <mi>l</mi> <mi>e</mi> <mi>f</mi> <mi>t</mi> <mo>=</mo> <mi>y</mi> <mi>c</mi> <mi>e</mi> <mi>n</mi> <mo>-</mo> <msubsup> <mi>h</mi> <mi>k</mi> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msubsup> <mo>/</mo> <mn>2</mn> <mo>,</mo> </mrow>
<mrow> <mi>x</mi> <mi>r</mi> <mi>i</mi> <mi>g</mi> <mi>h</mi> <mo>=</mo> <mi>x</mi> <mi>c</mi> <mi>e</mi> <mi>n</mi> <mo>+</mo> <msubsup> <mi>w</mi> <mi>k</mi> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msubsup> <mo>/</mo> <mn>2</mn> <mo>,</mo> </mrow>
<mrow> <mi>y</mi> <mi>r</mi> <mi>i</mi> <mi>g</mi> <mi>h</mi> <mo>=</mo> <mi>y</mi> <mi>c</mi> <mi>e</mi> <mi>n</mi> <mo>+</mo> <msubsup> <mi>h</mi> <mi>k</mi> <msub> <mi>a</mi> <mi>&amp;tau;</mi> </msub> </msubsup> <mo>/</mo> <mn>2</mn> <mo>;</mo> </mrow>
(2b22) calculates coordinate (x min, y min) and the lower right corner in the upper left corner of the acquiescence frame with marking frame intersection Coordinate (x max, y max)：
X min=max (xleft, xgleft),
Y min=max (yleft, ygleft),
X max=max (xrigh, xgrigh),
Y max=max (yrigh, ygrigh)；
Wherein (xgleft, ygleft) and (xgrigh, ygrigh) represent to mark coordinate and the bottom right in the frame g upper left corner respectively The coordinate at angle；
(2b23) calculates acquiescence frame x with marking the area inter (x, g) of frame g intersections:
Inter (x, g)=(max (y max-y min, 0)) * (max (x max-x min, 0))
(2b24) calculates acquiescence frame x and marks the Jaccard overlap coefficients between frame g：
<mrow> <mi>J</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <mi>int</mi> <mi>e</mi> <mi>r</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> </mrow> <mrow> <mo>(</mo> <mi>y</mi> <mi>r</mi> <mi>i</mi> <mi>g</mi> <mi>h</mi> <mo>-</mo> <mi>y</mi> <mi>l</mi> <mi>e</mi> <mi>f</mi> <mi>t</mi> <mo>)</mo> <mo>*</mo> <mo>(</mo> <mi>x</mi> <mi>r</mi> <mi>i</mi> <mi>g</mi> <mi>h</mi> <mo>-</mo> <mi>x</mi> <mi>l</mi> <mi>e</mi> <mi>f</mi> <mi>t</mi> <mo>)</mo> <mo>+</mo> <mo>(</mo> <mi>y</mi> <mi>g</mi> <mi>r</mi> <mi>i</mi> <mi>g</mi> <mi>h</mi> <mo>-</mo> <mi>y</mi> <mi>g</mi> <mi>l</mi> <mi>e</mi> <mi>f</mi> <mi>t</mi> <mo>)</mo> <mo>*</mo> <mo>(</mo> <mi>x</mi> <mi>g</mi> <mi>r</mi> <mi>i</mi> <mi>g</mi> <mi>h</mi> <mo>-</mo> <mi>x</mi> <mi>g</mi> <mi>l</mi> <mi>e</mi> <mi>f</mi> <mi>t</mi> <mo>)</mo> <mo>-</mo> <mi>int</mi> <mi>e</mi> <mi>r</mi> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> </mfrac> </mrow>
It is positive sample Pos to select acquiescence frame of J (x, the g) coefficient value more than 0.5, and other are negative sample Neg, completes mark frame Matching with giving tacit consent to frame.
4. according to the method for claim 1, wherein in step (2c) tectonic network model target loss function L (x, l, C, g), carry out as follows：
(2c1) is predicted using convolution filter on characteristic pattern, obtains giving tacit consent to classification score of the frame on all categories Set c and prediction frame are relative to position offset (Δ x, Δ y, the Δ w, Δ h), wherein (Δ x, Δ y) are represented for giving tacit consent to frame Offset of the frame centre coordinate relative to acquiescence frame centre coordinate is predicted, wherein Δ w represents that prediction frame is wide relative to silent Recognize the wide offset of frame, wherein Δ h represents prediction frame height relative to the high offset of acquiescence frame；
The classification score set c of (2c2) according to the acquiescence frame on characteristic pattern on all categories, calculate softmax classification damages Lose function Lconf(x,c)：
<mrow> <msub> <mi>L</mi> <mrow> <mi>c</mi> <mi>o</mi> <mi>n</mi> <mi>f</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>c</mi> <mo>)</mo> </mrow> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>&amp;Element;</mo> <mi>P</mi> <mi>o</mi> <mi>s</mi> </mrow> <mi>N</mi> </munderover> <msubsup> <mi>x</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> <mi>p</mi> </msubsup> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mrow> <mo>(</mo> <msubsup> <mover> <mi>c</mi> <mo>^</mo> </mover> <mi>i</mi> <mrow> <mi>P</mi> <mi>o</mi> <mi>s</mi> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>-</mo> <munder> <mo>&amp;Sigma;</mo> <mrow> <msub> <mi>i</mi> <mn>2</mn> </msub> <mo>&amp;Element;</mo> <mi>N</mi> <mi>e</mi> <mi>g</mi> </mrow> </munder> <mi>log</mi> <mrow> <mo>(</mo> <msubsup> <mover> <mi>c</mi> <mo>^</mo> </mover> <msub> <mi>i</mi> <mn>2</mn> </msub> <mrow> <mi>N</mi> <mi>e</mi> <mi>g</mi> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>,</mo> </mrow>
Wherein, whenRepresent that i-th of acquiescence frame matches with j-th of mark frame that classification is p,Represent i-th Individual acquiescence frame mismatches with j-th of mark frame that classification is p, and 0≤i≤N, N represent the acquiescence side matched with mark frame Frame number, 1≤p≤H, H are total categorical measure, and 0≤j≤T, T are the quantity of mark frame,Represent i-th in positive sample Give tacit consent to the average on all categories of frame,Represent i-th in negative sample2Individual acquiescence frame is on all categories Average, 0≤i2≤N2, N2Represent and the unmatched acquiescence frame number of mark frame；
(2c3) calculates positioning loss function Lloc(x,l,g)：
<mrow> <msub> <mi>L</mi> <mrow> <mi>l</mi> <mi>o</mi> <mi>c</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>l</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>&amp;Element;</mo> <mi>P</mi> <mi>o</mi> <mi>s</mi> </mrow> <mi>N</mi> </munderover> <munder> <mo>&amp;Sigma;</mo> <mrow> <mi>m</mi> <mo>&amp;Element;</mo> <mo>{</mo> <mi>c</mi> <mi>x</mi> <mo>,</mo> <mi>c</mi> <mi>y</mi> <mo>,</mo> <mi>w</mi> <mo>,</mo> <mi>h</mi> <mo>}</mo> </mrow> </munder> <msubsup> <mi>x</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> <mi>p</mi> </msubsup> <msub> <mi>smooth</mi> <mrow> <mi>L</mi> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <msubsup> <mi>l</mi> <mi>i</mi> <mi>m</mi> </msubsup> <mo>-</mo> <msubsup> <mover> <mi>g</mi> <mo>^</mo> </mover> <mi>j</mi> <mi>m</mi> </msubsup> <mo>)</mo> </mrow> </mrow>
Wherein (cx, cy) is by (centre coordinate of the acquiescence frame x after Δ x, Δ y) compensation, w, h are by (Δ w, Δ h) are mended Give tacit consent to the wide and high of frame after repaying,I-th of prediction frame that offset is m is represented,Expression offset is pre- j-th of m Survey frame；
(2c4) is according to Classification Loss function Lconf(x, c) and positioning loss function Lloc(x, l, g), obtain target loss function L (x,l,c,g)：
<mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>l</mi> <mo>,</mo> <mi>c</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <mrow> <mo>(</mo> <msub> <mi>L</mi> <mrow> <mi>c</mi> <mi>o</mi> <mi>n</mi> <mi>f</mi> </mrow> </msub> <mo>(</mo> <mrow> <mi>x</mi> <mo>,</mo> <mi>c</mi> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>&amp;alpha;L</mi> <mrow> <mi>l</mi> <mi>o</mi> <mi>c</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>l</mi> <mo>,</mo> <mi>g</mi> <mo>)</mo> </mrow> <mo>.</mo> </mrow>
5. original image to be detected wherein according to the method for claim 1, is input to the net trained in step (3) In network model, target classification and the position coordinates in image to be detected are obtained, is carried out as follows：
(3a) carries out Quadtree Partition in pre-segmentation layer to input picture, extracts area-of-interest；
(3b) carries out feature extraction using basic convolutional layer and auxiliary convolutional layer to area-of-interest, obtains the feature of multiple yardsticks Figure；
(3c) calculates the position coordinates of acquiescence frame on the characteristic pattern of multiple yardsticks；
(3d) is predicted on the characteristic pattern of multiple yardsticks using convolution filter, obtains the target class in multiple prediction frames The position offset of other score and the relative acquiescence frame of prediction frame；
(3e) is using non-maxima suppression to the target classification in multiple prediction frames and the position of the relative acquiescence frame of prediction frame Put offset to be suppressed, obtain the target classification in final prediction frame and predict that the position of the relative acquiescence frame of frame is inclined Shifting amount, and prediction frame is obtained according to the position offset of the relative acquiescence frame of prediction frame and the position coordinates of acquiescence frame Position coordinates.
CN201710598875.XA 2017-07-21 2017-07-21 Based on pre-segmentation and the deep learning object detection method returned Pending CN107423760A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710598875.XA CN107423760A (en) 2017-07-21 2017-07-21 Based on pre-segmentation and the deep learning object detection method returned

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710598875.XA CN107423760A (en) 2017-07-21 2017-07-21 Based on pre-segmentation and the deep learning object detection method returned

Publications (1)

Publication Number Publication Date
CN107423760A true CN107423760A (en) 2017-12-01

Family

ID=60430914

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710598875.XA Pending CN107423760A (en) 2017-07-21 2017-07-21 Based on pre-segmentation and the deep learning object detection method returned

Country Status (1)

Country Link
CN (1) CN107423760A (en)

Cited By (61)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108009525A (en) * 2017-12-25 2018-05-08 北京航空航天大学 A kind of specific objective recognition methods over the ground of the unmanned plane based on convolutional neural networks
CN108171752A (en) * 2017-12-28 2018-06-15 成都阿普奇科技股份有限公司 A kind of sea ship video detection and tracking based on deep learning
CN108256464A (en) * 2018-01-12 2018-07-06 适普远景遥感信息技术（北京）有限公司 High-resolution remote sensing image urban road extracting method based on deep learning
CN108257114A (en) * 2017-12-29 2018-07-06 天津市万贸科技有限公司 A kind of transmission facility defect inspection method based on deep learning
CN108288270A (en) * 2018-02-05 2018-07-17 南京邮电大学 A kind of object detection method based on channel trimming and full convolution deep learning
CN108304787A (en) * 2018-01-17 2018-07-20 河南工业大学 Road target detection method based on convolutional neural networks
CN108334878A (en) * 2018-02-07 2018-07-27 北京影谱科技股份有限公司 Video images detection method and apparatus
CN108337416A (en) * 2018-03-09 2018-07-27 天津港(集团)有限公司 It is a kind of for the automatic lifting identification device of perimeter protection and recognition methods
CN108376235A (en) * 2018-01-15 2018-08-07 深圳市易成自动驾驶技术有限公司 Image detecting method, device and computer readable storage medium
CN108460382A (en) * 2018-03-26 2018-08-28 西安电子科技大学 Remote sensing image Ship Detection based on deep learning single step detector
CN108510012A (en) * 2018-05-04 2018-09-07 四川大学 A kind of target rapid detection method based on Analysis On Multi-scale Features figure
CN108498089A (en) * 2018-05-08 2018-09-07 北京邮电大学 A kind of noninvasive continuous BP measurement method based on deep neural network
CN108573238A (en) * 2018-04-23 2018-09-25 济南浪潮高新科技投资发展有限公司 A kind of vehicle checking method based on dual network structure
CN108595544A (en) * 2018-04-09 2018-09-28 深源恒际科技有限公司 A kind of document picture classification method
CN108629767A (en) * 2018-04-28 2018-10-09 Oppo广东移动通信有限公司 A kind of method, device and mobile terminal of scene detection
CN108647655A (en) * 2018-05-16 2018-10-12 北京工业大学 Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks
CN108765392A (en) * 2018-05-20 2018-11-06 复旦大学 A kind of digestive endoscope lesion detection and recognition methods based on sliding window
CN108846826A (en) * 2018-04-24 2018-11-20 深圳大学 Object detecting method, device, image processing equipment and storage medium
CN108898628A (en) * 2018-06-21 2018-11-27 北京纵目安驰智能科技有限公司 Three-dimensional vehicle object's pose estimation method, system, terminal and storage medium based on monocular
CN108960175A (en) * 2018-07-12 2018-12-07 天津艾思科尔科技有限公司 A kind of licence plate recognition method based on deep learning
CN109002841A (en) * 2018-06-27 2018-12-14 淮阴工学院 A kind of building element extracting method based on Faster-RCNN model
CN109117794A (en) * 2018-08-16 2019-01-01 广东工业大学 A kind of moving target behavior tracking method, apparatus, equipment and readable storage medium storing program for executing
CN109147254A (en) * 2018-07-18 2019-01-04 武汉大学 A kind of video outdoor fire disaster smog real-time detection method based on convolutional neural networks
CN109344821A (en) * 2018-08-30 2019-02-15 西安电子科技大学 Small target detecting method based on Fusion Features and deep learning
CN109409365A (en) * 2018-10-25 2019-03-01 江苏德劭信息科技有限公司 It is a kind of that method is identified and positioned to fruit-picking based on depth targets detection
CN109446888A (en) * 2018-09-10 2019-03-08 唯思科技（北京）有限公司 A kind of elongated class article detection method based on convolutional neural networks
CN109597087A (en) * 2018-11-15 2019-04-09 天津大学 A kind of 3D object detection method based on point cloud data
CN109635666A (en) * 2018-11-16 2019-04-16 南京航空航天大学 A kind of image object rapid detection method based on deep learning
CN109685008A (en) * 2018-12-25 2019-04-26 云南大学 A kind of real-time video object detection method
CN109685528A (en) * 2017-12-18 2019-04-26 北京京东尚科信息技术有限公司 System and method based on deep learning detection counterfeit product
CN109684803A (en) * 2018-12-19 2019-04-26 西安电子科技大学 Man-machine verification method based on gesture sliding
CN109711326A (en) * 2018-12-25 2019-05-03 云南大学 A kind of video object detection method based on shallow-layer residual error network
CN109977783A (en) * 2019-02-28 2019-07-05 浙江新再灵科技股份有限公司 Method based on the independent boarding detection of vertical ladder scene perambulator
CN109977943A (en) * 2019-02-14 2019-07-05 平安科技（深圳）有限公司 A kind of images steganalysis method, system and storage medium based on YOLO
CN110110586A (en) * 2019-03-18 2019-08-09 北京理工雷科电子信息技术有限公司 The method and device of remote sensing airport Airplane detection based on deep learning
CN110210472A (en) * 2018-02-28 2019-09-06 佛山科学技术学院 A kind of method for checking object based on depth network
CN110363122A (en) * 2019-07-03 2019-10-22 昆明理工大学 A kind of cross-domain object detection method based on multilayer feature alignment
CN110543801A (en) * 2018-05-29 2019-12-06 北京林业大学 Pine pest detection method, system and device based on neural network and unmanned aerial vehicle aerial image
CN110688925A (en) * 2019-09-19 2020-01-14 国网山东省电力公司电力科学研究院 Cascade target identification method and system based on deep learning
CN110749324A (en) * 2019-10-28 2020-02-04 深圳市赛为智能股份有限公司 Unmanned aerial vehicle rescue positioning method and device, computer equipment and storage medium
CN110781793A (en) * 2019-10-21 2020-02-11 合肥成方信息技术有限公司 Artificial intelligence real-time image recognition method based on quadtree algorithm
CN110796127A (en) * 2020-01-06 2020-02-14 四川通信科研规划设计有限责任公司 Embryo prokaryotic detection system based on occlusion sensing, storage medium and terminal
CN110826575A (en) * 2019-12-13 2020-02-21 哈尔滨工程大学 Underwater target identification method based on machine learning
WO2020038205A1 (en) * 2018-08-24 2020-02-27 腾讯科技（深圳）有限公司 Target detection method and apparatus, computer-readable storage medium, and computer device
CN110929774A (en) * 2019-11-18 2020-03-27 腾讯科技（深圳）有限公司 Method for classifying target objects in image, method and device for training model
CN110942140A (en) * 2019-11-29 2020-03-31 任科扬 Artificial neural network difference and iteration data processing method and device
WO2020093624A1 (en) * 2018-11-07 2020-05-14 五邑大学 Antenna downward inclination angle measurement method based on multi-scale detection algorithm
WO2020107886A1 (en) * 2018-11-29 2020-06-04 Huawei Technologies Co., Ltd. Loading apparatus and method for convolution with stride or dilation of 2
CN111340768A (en) * 2020-02-21 2020-06-26 之江实验室 Multi-center effect compensation method based on PET/CT intelligent diagnosis system
CN111382787A (en) * 2020-03-06 2020-07-07 芯薇(上海)智能科技有限公司 Target detection method based on deep learning
CN111476167A (en) * 2020-04-09 2020-07-31 北京中科千寻科技有限公司 student-T distribution assistance-based one-stage direction remote sensing image target detection method
CN111611918A (en) * 2020-05-20 2020-09-01 重庆大学 Traffic flow data set acquisition and construction method based on aerial photography data and deep learning
CN111626419A (en) * 2020-07-20 2020-09-04 成都安智杰科技有限公司 Convolutional neural network structure, target detection method and device
CN111723737A (en) * 2020-06-19 2020-09-29 河南科技大学 Target detection method based on multi-scale matching strategy deep feature learning
CN111738070A (en) * 2020-05-14 2020-10-02 华南理工大学 Automatic accurate detection method for multiple small targets
CN111753682A (en) * 2020-06-11 2020-10-09 中建地下空间有限公司 Hoisting area dynamic monitoring method based on target detection algorithm
CN111798416A (en) * 2019-06-20 2020-10-20 太原理工大学 Intelligent glomerulus detection method and system based on pathological image and deep learning
CN112363844A (en) * 2021-01-12 2021-02-12 之江实验室 Convolutional neural network vertical segmentation method for image processing
CN112926681A (en) * 2021-03-29 2021-06-08 复旦大学 Target detection method and device based on deep convolutional neural network
CN113269725A (en) * 2021-04-28 2021-08-17 安徽理工大学 Coal gangue rapid detection method based on imaging technology and deep learning
US20220207825A1 (en) * 2019-09-17 2022-06-30 SZ DJI Technology Co., Ltd. Machine vision-based tree recognition method and device

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101937566A (en) * 2010-09-20 2011-01-05 西安电子科技大学 SAR image segmentation method combining background information and maximum posterior marginal probability standard
CN103077536A (en) * 2012-12-31 2013-05-01 华中科技大学 Space-time mutative scale moving target detection method
CN103218819A (en) * 2013-04-20 2013-07-24 复旦大学 Automatic selection method for optimal homogenous region of ultrasound image and based on quad-tree decomposition
CN106447658A (en) * 2016-09-26 2017-02-22 西北工业大学 Significant target detection method based on FCN (fully convolutional network) and CNN (convolutional neural network)
CN106600560A (en) * 2016-12-22 2017-04-26 福州大学 Image defogging method for automobile data recorder
US20170169315A1 (en) * 2015-12-15 2017-06-15 Sighthound, Inc. Deeply learned convolutional neural networks (cnns) for object localization and classification

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101937566A (en) * 2010-09-20 2011-01-05 西安电子科技大学 SAR image segmentation method combining background information and maximum posterior marginal probability standard
CN103077536A (en) * 2012-12-31 2013-05-01 华中科技大学 Space-time mutative scale moving target detection method
CN103218819A (en) * 2013-04-20 2013-07-24 复旦大学 Automatic selection method for optimal homogenous region of ultrasound image and based on quad-tree decomposition
US20170169315A1 (en) * 2015-12-15 2017-06-15 Sighthound, Inc. Deeply learned convolutional neural networks (cnns) for object localization and classification
CN106447658A (en) * 2016-09-26 2017-02-22 西北工业大学 Significant target detection method based on FCN (fully convolutional network) and CNN (convolutional neural network)
CN106600560A (en) * 2016-12-22 2017-04-26 福州大学 Image defogging method for automobile data recorder

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WEI LIU等；: ""SSD: Single Shot MultiBox Detector"", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 *

Cited By (89)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685528A (en) * 2017-12-18 2019-04-26 北京京东尚科信息技术有限公司 System and method based on deep learning detection counterfeit product
CN108009525A (en) * 2017-12-25 2018-05-08 北京航空航天大学 A kind of specific objective recognition methods over the ground of the unmanned plane based on convolutional neural networks
CN108009525B (en) * 2017-12-25 2018-10-12 北京航空航天大学 A kind of specific objective recognition methods over the ground of the unmanned plane based on convolutional neural networks
CN108171752A (en) * 2017-12-28 2018-06-15 成都阿普奇科技股份有限公司 A kind of sea ship video detection and tracking based on deep learning
CN108257114A (en) * 2017-12-29 2018-07-06 天津市万贸科技有限公司 A kind of transmission facility defect inspection method based on deep learning
CN108256464A (en) * 2018-01-12 2018-07-06 适普远景遥感信息技术（北京）有限公司 High-resolution remote sensing image urban road extracting method based on deep learning
CN108256464B (en) * 2018-01-12 2020-08-11 适普远景遥感信息技术（北京）有限公司 High-resolution remote sensing image urban road extraction method based on deep learning
CN108376235A (en) * 2018-01-15 2018-08-07 深圳市易成自动驾驶技术有限公司 Image detecting method, device and computer readable storage medium
CN108304787A (en) * 2018-01-17 2018-07-20 河南工业大学 Road target detection method based on convolutional neural networks
CN108288270A (en) * 2018-02-05 2018-07-17 南京邮电大学 A kind of object detection method based on channel trimming and full convolution deep learning
CN108288270B (en) * 2018-02-05 2022-06-03 南京邮电大学 Target detection method based on channel pruning and full convolution deep learning
CN108334878A (en) * 2018-02-07 2018-07-27 北京影谱科技股份有限公司 Video images detection method and apparatus
CN108334878B (en) * 2018-02-07 2021-01-05 北京影谱科技股份有限公司 Video image detection method, device and equipment and readable storage medium
CN110210472A (en) * 2018-02-28 2019-09-06 佛山科学技术学院 A kind of method for checking object based on depth network
CN108337416A (en) * 2018-03-09 2018-07-27 天津港(集团)有限公司 It is a kind of for the automatic lifting identification device of perimeter protection and recognition methods
CN108460382A (en) * 2018-03-26 2018-08-28 西安电子科技大学 Remote sensing image Ship Detection based on deep learning single step detector
CN108460382B (en) * 2018-03-26 2021-04-06 西安电子科技大学 Optical remote sensing image ship detection method based on deep learning single-step detector
CN108595544A (en) * 2018-04-09 2018-09-28 深源恒际科技有限公司 A kind of document picture classification method
CN108573238A (en) * 2018-04-23 2018-09-25 济南浪潮高新科技投资发展有限公司 A kind of vehicle checking method based on dual network structure
CN108846826A (en) * 2018-04-24 2018-11-20 深圳大学 Object detecting method, device, image processing equipment and storage medium
CN108629767A (en) * 2018-04-28 2018-10-09 Oppo广东移动通信有限公司 A kind of method, device and mobile terminal of scene detection
CN108510012B (en) * 2018-05-04 2022-04-01 四川大学 Target rapid detection method based on multi-scale feature map
CN108510012A (en) * 2018-05-04 2018-09-07 四川大学 A kind of target rapid detection method based on Analysis On Multi-scale Features figure
CN108498089A (en) * 2018-05-08 2018-09-07 北京邮电大学 A kind of noninvasive continuous BP measurement method based on deep neural network
CN108498089B (en) * 2018-05-08 2022-03-25 北京邮电大学 Noninvasive continuous blood pressure measuring method based on deep neural network
CN108647655A (en) * 2018-05-16 2018-10-12 北京工业大学 Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks
CN108647655B (en) * 2018-05-16 2022-07-12 北京工业大学 Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network
CN108765392A (en) * 2018-05-20 2018-11-06 复旦大学 A kind of digestive endoscope lesion detection and recognition methods based on sliding window
CN108765392B (en) * 2018-05-20 2022-03-18 复旦大学 Digestive tract endoscope lesion detection and identification method based on sliding window
CN110543801A (en) * 2018-05-29 2019-12-06 北京林业大学 Pine pest detection method, system and device based on neural network and unmanned aerial vehicle aerial image
CN108898628A (en) * 2018-06-21 2018-11-27 北京纵目安驰智能科技有限公司 Three-dimensional vehicle object's pose estimation method, system, terminal and storage medium based on monocular
CN109002841A (en) * 2018-06-27 2018-12-14 淮阴工学院 A kind of building element extracting method based on Faster-RCNN model
CN109002841B (en) * 2018-06-27 2021-11-12 淮阴工学院 Building component extraction method based on fast-RCNN model
CN108960175A (en) * 2018-07-12 2018-12-07 天津艾思科尔科技有限公司 A kind of licence plate recognition method based on deep learning
CN109147254B (en) * 2018-07-18 2021-05-18 武汉大学 Video field fire smoke real-time detection method based on convolutional neural network
CN109147254A (en) * 2018-07-18 2019-01-04 武汉大学 A kind of video outdoor fire disaster smog real-time detection method based on convolutional neural networks
CN109117794A (en) * 2018-08-16 2019-01-01 广东工业大学 A kind of moving target behavior tracking method, apparatus, equipment and readable storage medium storing program for executing
US11710293B2 (en) 2018-08-24 2023-07-25 Tencent Technology (Shenzhen) Company Limited Target detection method and apparatus, computer-readable storage medium, and computer device
WO2020038205A1 (en) * 2018-08-24 2020-02-27 腾讯科技（深圳）有限公司 Target detection method and apparatus, computer-readable storage medium, and computer device
CN109344821A (en) * 2018-08-30 2019-02-15 西安电子科技大学 Small target detecting method based on Fusion Features and deep learning
CN109446888A (en) * 2018-09-10 2019-03-08 唯思科技（北京）有限公司 A kind of elongated class article detection method based on convolutional neural networks
CN109409365A (en) * 2018-10-25 2019-03-01 江苏德劭信息科技有限公司 It is a kind of that method is identified and positioned to fruit-picking based on depth targets detection
US11145089B2 (en) 2018-11-07 2021-10-12 Wuyi University Method for measuring antenna downtilt based on multi-scale detection algorithm
WO2020093624A1 (en) * 2018-11-07 2020-05-14 五邑大学 Antenna downward inclination angle measurement method based on multi-scale detection algorithm
CN109597087B (en) * 2018-11-15 2022-07-01 天津大学 Point cloud data-based 3D target detection method
CN109597087A (en) * 2018-11-15 2019-04-09 天津大学 A kind of 3D object detection method based on point cloud data
CN109635666A (en) * 2018-11-16 2019-04-16 南京航空航天大学 A kind of image object rapid detection method based on deep learning
CN109635666B (en) * 2018-11-16 2023-04-18 南京航空航天大学 Image target rapid detection method based on deep learning
WO2020107886A1 (en) * 2018-11-29 2020-06-04 Huawei Technologies Co., Ltd. Loading apparatus and method for convolution with stride or dilation of 2
CN109684803B (en) * 2018-12-19 2021-04-20 西安电子科技大学 Man-machine verification method based on gesture sliding
CN109684803A (en) * 2018-12-19 2019-04-26 西安电子科技大学 Man-machine verification method based on gesture sliding
CN109685008A (en) * 2018-12-25 2019-04-26 云南大学 A kind of real-time video object detection method
CN109711326A (en) * 2018-12-25 2019-05-03 云南大学 A kind of video object detection method based on shallow-layer residual error network
CN109977943B (en) * 2019-02-14 2024-05-07 平安科技（深圳）有限公司 Image target recognition method, system and storage medium based on YOLO
CN109977943A (en) * 2019-02-14 2019-07-05 平安科技（深圳）有限公司 A kind of images steganalysis method, system and storage medium based on YOLO
CN109977783A (en) * 2019-02-28 2019-07-05 浙江新再灵科技股份有限公司 Method based on the independent boarding detection of vertical ladder scene perambulator
CN110110586A (en) * 2019-03-18 2019-08-09 北京理工雷科电子信息技术有限公司 The method and device of remote sensing airport Airplane detection based on deep learning
CN111798416B (en) * 2019-06-20 2023-04-18 太原理工大学 Intelligent glomerulus detection method and system based on pathological image and deep learning
CN111798416A (en) * 2019-06-20 2020-10-20 太原理工大学 Intelligent glomerulus detection method and system based on pathological image and deep learning
CN110363122A (en) * 2019-07-03 2019-10-22 昆明理工大学 A kind of cross-domain object detection method based on multilayer feature alignment
CN110363122B (en) * 2019-07-03 2022-10-11 昆明理工大学 Cross-domain target detection method based on multi-layer feature alignment
US20220207825A1 (en) * 2019-09-17 2022-06-30 SZ DJI Technology Co., Ltd. Machine vision-based tree recognition method and device
CN110688925A (en) * 2019-09-19 2020-01-14 国网山东省电力公司电力科学研究院 Cascade target identification method and system based on deep learning
CN110688925B (en) * 2019-09-19 2022-11-15 国网智能科技股份有限公司 Cascade target identification method and system based on deep learning
CN110781793A (en) * 2019-10-21 2020-02-11 合肥成方信息技术有限公司 Artificial intelligence real-time image recognition method based on quadtree algorithm
CN110749324A (en) * 2019-10-28 2020-02-04 深圳市赛为智能股份有限公司 Unmanned aerial vehicle rescue positioning method and device, computer equipment and storage medium
CN110929774A (en) * 2019-11-18 2020-03-27 腾讯科技（深圳）有限公司 Method for classifying target objects in image, method and device for training model
CN110929774B (en) * 2019-11-18 2023-11-14 腾讯科技（深圳）有限公司 Classification method, model training method and device for target objects in image
CN110942140A (en) * 2019-11-29 2020-03-31 任科扬 Artificial neural network difference and iteration data processing method and device
CN110942140B (en) * 2019-11-29 2022-11-08 任科扬 Artificial neural network difference and iteration data processing method and device
CN110826575A (en) * 2019-12-13 2020-02-21 哈尔滨工程大学 Underwater target identification method based on machine learning
CN110796127A (en) * 2020-01-06 2020-02-14 四川通信科研规划设计有限责任公司 Embryo prokaryotic detection system based on occlusion sensing, storage medium and terminal
CN111340768B (en) * 2020-02-21 2021-03-09 之江实验室 Multi-center effect compensation method based on PET/CT intelligent diagnosis system
US11715562B2 (en) 2020-02-21 2023-08-01 Zhejiang Lab Method for multi-center effect compensation based on PET/CT intelligent diagnosis system
CN111340768A (en) * 2020-02-21 2020-06-26 之江实验室 Multi-center effect compensation method based on PET/CT intelligent diagnosis system
CN111382787A (en) * 2020-03-06 2020-07-07 芯薇(上海)智能科技有限公司 Target detection method based on deep learning
CN111476167B (en) * 2020-04-09 2024-03-22 北京中科千寻科技有限公司 One-stage direction remote sensing image target detection method based on student-T distribution assistance
CN111476167A (en) * 2020-04-09 2020-07-31 北京中科千寻科技有限公司 student-T distribution assistance-based one-stage direction remote sensing image target detection method
CN111738070A (en) * 2020-05-14 2020-10-02 华南理工大学 Automatic accurate detection method for multiple small targets
CN111611918B (en) * 2020-05-20 2023-07-21 重庆大学 Traffic flow data set acquisition and construction method based on aerial data and deep learning
CN111611918A (en) * 2020-05-20 2020-09-01 重庆大学 Traffic flow data set acquisition and construction method based on aerial photography data and deep learning
CN111753682A (en) * 2020-06-11 2020-10-09 中建地下空间有限公司 Hoisting area dynamic monitoring method based on target detection algorithm
CN111723737A (en) * 2020-06-19 2020-09-29 河南科技大学 Target detection method based on multi-scale matching strategy deep feature learning
CN111723737B (en) * 2020-06-19 2023-11-17 河南科技大学 Target detection method based on multi-scale matching strategy deep feature learning
CN111626419A (en) * 2020-07-20 2020-09-04 成都安智杰科技有限公司 Convolutional neural network structure, target detection method and device
CN112363844A (en) * 2021-01-12 2021-02-12 之江实验室 Convolutional neural network vertical segmentation method for image processing
CN112363844B (en) * 2021-01-12 2021-04-09 之江实验室 Convolutional neural network vertical segmentation method for image processing
CN112926681A (en) * 2021-03-29 2021-06-08 复旦大学 Target detection method and device based on deep convolutional neural network
CN113269725A (en) * 2021-04-28 2021-08-17 安徽理工大学 Coal gangue rapid detection method based on imaging technology and deep learning

Similar Documents

Publication Publication Date Title
CN107423760A (en) Based on pre-segmentation and the deep learning object detection method returned
CN105608456B (en) A kind of multi-direction Method for text detection based on full convolutional network
CN109948425B (en) Pedestrian searching method and device for structure-aware self-attention and online instance aggregation matching
CN112861720B (en) Remote sensing image small sample target detection method based on prototype convolutional neural network
CN110276269B (en) Remote sensing image target detection method based on attention mechanism
CN107609525B (en) Remote sensing image target detection method for constructing convolutional neural network based on pruning strategy
CN106778835B (en) Remote sensing image airport target identification method fusing scene information and depth features
CN106875381B (en) Mobile phone shell defect detection method based on deep learning
CN108427912B (en) Optical remote sensing image target detection method based on dense target feature learning
CN110929607B (en) Remote sensing identification method and system for urban building construction progress
CN108596055B (en) Airport target detection method of high-resolution remote sensing image under complex background
CN106228125B (en) Method for detecting lane lines based on integrated study cascade classifier
CN111753828B (en) Natural scene horizontal character detection method based on deep convolutional neural network
CN103049763B (en) Context-constraint-based target identification method
CN110569738B (en) Natural scene text detection method, equipment and medium based on densely connected network
CN110175613A (en) Street view image semantic segmentation method based on Analysis On Multi-scale Features and codec models
CN111445488B (en) Method for automatically identifying and dividing salt body by weak supervision learning
CN106682569A (en) Fast traffic signboard recognition method based on convolution neural network
CN107437100A (en) A kind of picture position Forecasting Methodology based on the association study of cross-module state
CN107784288A (en) A kind of iteration positioning formula method for detecting human face based on deep neural network
CN109858327B (en) Character segmentation method based on deep learning
CN109377511B (en) Moving target tracking method based on sample combination and depth detection network
CN111709397A (en) Unmanned aerial vehicle variable-size target detection method based on multi-head self-attention mechanism
CN110472572A (en) The quick identification and classification method of naval target under a kind of complex environment
CN113221956B (en) Target identification method and device based on improved multi-scale depth model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20171201

WD01 Invention patent application deemed withdrawn after publication