CN107316058A - Improve the method for target detection performance by improving target classification and positional accuracy - Google Patents
Improve the method for target detection performance by improving target classification and positional accuracy Download PDFInfo
- Publication number
- CN107316058A CN107316058A CN201710450327.2A CN201710450327A CN107316058A CN 107316058 A CN107316058 A CN 107316058A CN 201710450327 A CN201710450327 A CN 201710450327A CN 107316058 A CN107316058 A CN 107316058A
- Authority
- CN
- China
- Prior art keywords
- image
- candidate frame
- target detection
- frame
- convolutional layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
-
- 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
- G06V10/40—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)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of method by improving target classification and positional accuracy improvement target detection performance, its technical characteristics is:Characteristics of image is extracted according to convolutional neural networks framework, and selects M layers of output before convolutional layer to carry out Fusion Features, the characteristic pattern of multiple features is formed;Mesh generation is carried out on convolutional layer M, the target candidate frame of fixed number and size is predicted in each network;Candidate frame is mapped on characteristic pattern and cut, result then will be cut and carry out multiple features connection;By the above results by full articulamentum after, characteristics of image is classified by Softmax sorting algorithms, and online iterative regression positioning is carried out with overlapping area loss function, obtains the result of final target detection.The present invention is reasonable in design, feature is extracted by convolutional neural networks, and multilayer fusion is carried out to characteristics of image, finally characteristics of image is classified using Softmax sorting algorithms, and positioned using overlapping area loss function, obtain good object detection results.
Description
Technical field
It is especially a kind of to be improved by improving target classification and positional accuracy the invention belongs to target detection technique field
The method of target detection performance.
Background technology
The mankind have more than 80% information source in vision in the perception engineering of the material world.And image is at certain
It is with different patterns to mankind's transmission information, and it is as a kind of important is reflected to one kind of objective reality in meaning
Information carrier, with it is directly perceived, abundant in content with being easy to exchange the features such as, be multimedia important composition content, therefore, based on figure
As the various applications for the treatment of technology are just arisen at the historic moment.Images steganalysis and detection technique are exactly wherein most typical application skill
Art.Computer vision research purpose is that the mankind are realized with computer to the perception of objective world, identification and are understood, target detection
(Object Detection) is most common problem in computer vision, and received in theory on computer vision research field
Extensive concern, has broad application prospects.Eyes, which are opened, when machine " opening " sees universal time, it is necessary to judge which be present in its visual field
A little targets, what is respectively, where.The target detection of view-based access control model is image procossing, computer vision, pattern-recognition
Etc. many multi-disciplinary crossing research problems.The purpose of target detection is that target is picked out from the background of differing complexity, and
Indicated in the form of encirclement frame (Bounding Box), so as to the follow-up work such as complete to track, recognize.Therefore, target
Detection is the high-rise background task understood with application, and the quality of its performance will directly affect follow-up target following, action recognition
And the performance of the task on the middle and senior level such as behavior understanding., it is necessary to be handled in real time multiple targets especially in complex scene
When, target, which is automatically extracted and recognized, just seems especially important.Therefore, object detection and recognition is graphical analysis and the base understood
Plinth, furthers investigate object detection and recognition algorithm, very important meaning is suffered from academia and industrial quarters.But for machine
For device, because the dynamic change of complicated identification background and target in itself adds the difficulty of target identification, huge system
The matrix operation of parameter and higher-dimension takes substantial amounts of processing time, and the problem of Target detection and identification also has larger such as recognizes
The degree of accuracy, real-time all await improve.
The main task of target detection be in image sequence target object carry out automatic detection, including judge classification with
Recognize position.Current popular algorithm of target detection, generates 1K-2K candidate frame, then for every first on a pictures
Individual candidate frame extracts feature using CNN convolutional neural networks, and secondly feature is inputted to the SVM classifier or Softmax of each class
Grader judges whether target belongs to such, finally realizes the precisely fixed of target using the position for returning device amendment candidate frame
Position.Traditional algorithm of target detection is had translation, imitative set, rotates using the feature such as SIFT, HOG and LBP by finding in picture
The matching between image is realized etc. the invariant features point under change situation, so as to realize target detection.But extract the quality of feature
The accuracy of classification is directly influenced, due to the Morphological Diversity of target, illumination variation diversity, the factor such as background diversity makes
The feature that a robust must be designed not is that so easily the adaptability of traditional characteristic is not strong.And based on CNN convolutional Neurals
The feature extraction of network has good robustness, and convolutional neural networks are more than one of particular design for identification two-dimensional shapes
Layer perceptron, this network structure has height consistency to translation, proportional zoom, inclination or the deformation of his common form.Carry
The CNN models of feature are taken to be obtained by training in advance, pre-training is complete based on the Computer Vision Recognition challenge matches of ILVCR 2012
Portion's data set is trained, and being then based on the training sets of PASCAL VOC 2007 for pre-training model carries out tuning training, so that
Realize and pass through CNN network extraction picture features.Deep learning is widely used in the depth that target detection comes from Alex et al. propositions
The convolutional Neural net AlexNet network architectures, the framework achieves extraordinary achievement in the matches of ILSVRC 2012, hereafter, volume
Product neutral net is widely used in all kinds of image association areas.The AlexNet of Geoffrey Hinton designs is one 8 layers
CNN frameworks, including 5 convolutional layers and 3 full articulamentums, the error rate of best algorithm at that time is halved, and it demonstrates CNN multiple
Validity under parasitic mode type, and GPU causes training to obtain result in acceptable time range., Christian in 2014
Szegedy proposes GoogleNet frameworks and taken first place in the classification matches of ILSVRC 2014, different from AlexNet
It is:GoogleNet depth (number of plies) is deeper, and width (layer core or neuron number) is wider.The same year, Andrew Zisserman
The VGG-Net frameworks of proposition take first place in the positioning matches of ILSVRC 2014, and unlike AlexNet:VGG-Net
Using more layers, generally there are 16-19 layers.In 2015, the Res-Net frameworks that Kaiming He are proposed were in ILSVRC 2015
Taken first place in classification and positioning match, the model employs 152 layers of deep layer convolutional neural networks.Hinton professor into
Work(, has attracted the concern of a large amount of scholar both at home and abroad;Meanwhile, industrial quarters add deep learning research, Baidu, google,
Facebook sets up deep learning laboratory one after another, by deep learning, carries out image recognition and classification.Although researcher carries
Go out many algorithm of target detection based on deep learning convolutional neural networks, these algorithms also achieve good effect, but
It is still to have many aspects to have much room for improvement, such as picture background is complicated, network inputs size is fixed, candidate frame is excessive, training speed is slow,
Consume the problems such as computer memory, wisp detect inaccurate, complex steps and position not accurate.
The content of the invention
It is to overcome the deficiencies in the prior art the mesh of the present invention, proposes that a kind of reasonable in design, precision is high and stability is strong
Improve the method for target detection performance by improving target classification and positional accuracy.
The present invention solves its technical problem and takes following technical scheme to realize:
It is a kind of to improve the method for target detection performance by improving target classification and positional accuracy, comprise the following steps:
Step 1, characteristics of image is extracted according to convolutional neural networks framework, and select before convolutional layer M layer output progress feature
Fusion, forms the characteristic pattern of multiple features;
Step 2, mesh generation is carried out on convolutional layer M, predict that the target of fixed number and size is waited in each network
Select frame;
Step 3, candidate frame is mapped on characteristic pattern cut, then will cut result and carry out multiple features connection;
Step 4, by the above results by full articulamentum after, characteristics of image is classified by Softmax sorting algorithms,
And online iterative regression positioning is carried out with overlapping area loss function, obtain the result of final target detection.
The specific method of the step 1 comprises the following steps:
(1) the picture with the true encirclement frame of object is input in convolutional neural networks framework first, carried by Caffe
Image is taken to pass through the feature of convolutional neural networks different layers;
(2) the characteristics of image forward convolutional layer exported carries out maximum pondization and operated, and the image that convolutional layer M is exported
Feature carries out deconvolution operation, realizes that output characteristic of the size of output all with middle convolutional layer is in the same size;
(3) the last feature for exporting all convolutional layers is merged, and obtains the characteristic pattern of the multi-feature extraction of image.
The implementation method of the step 2 comprises the following steps:
(1) 6*6 mesh generation is carried out on the characteristic pattern that convolutional layer M is exported;
(2) the candidate frame of object may be included by predicting 4 in the small lattice middle of each single network, this 4 candidate frame tools
There is fixed size and length-width ratio, length-width ratio is respectively 1:1、1:2 and 2:1, only for 1:The candidate frame of 1 length-width ratio, sets 2 kinds
Candidate frame size 0.6 and 0.9;
(3) during network training, we are matched the true encirclement frame and candidate frame of object, by the two
IOU overlapping areas screened more than or equal to 0.7, and delete the candidate frame beyond image boundary;
(4) 100 candidate frames are finally generated on convolutional layer M characteristic pattern.
The implementation method of the step 3 comprises the following steps:
(1), according to 100 candidate frames generated on convolutional layer M characteristic pattern, corresponding multilayer is mapped according to its position
On characteristic pattern, and cut accordingly on multilayer feature figure;
(2) the characteristic pattern square after cutting is done into 1*1 convolution, 3*3 convolution and 5*5 volumes are then carried out respectively to convolution results
Product;
(3) in order to obtain full text information, by multilayer feature figure by maximum pond layer, then by 1*1 convolutional layers and activation
Layer;
(4) the convolution output result of 1*1 convolution, 3*3 convolution, 5*5 convolution and full text information is connected according to tandem
Connect, form the multiple features connection of candidate frame.
The concrete methods of realizing of the step 4 comprises the following steps:
By full articulamentum after, characteristics of image is classified by Softmax sorting algorithms, based on target detection
Data set, has oneself corresponding precision per type objects;
(2) recurrence positioning is carried out to candidate frame by overlapping area loss function so that candidate frame is closer to the true of object
Real encirclement frame, the loss function is candidate frame and the common factor area divided by union area of true encirclement frame;
(3) it is ranked up according to Softmax losses and overlapping area penalty values, positive sample and negative sample is filtered out online
Ratio is 3:1, renewal Sample Storehouse, which is input on multilayer feature figure, proceeds iterative regression positioning;
(4) after iteration n times, candidate frame closer to object true encirclement frame, model training well after can carry out reality
The test of object.
Advantages and positive effects of the present invention are:
1st, picture is inputted in VGG-16 convolutional neural networks to obtain more image informations and extracts image by the present invention
Feature, then carries out multilayer to characteristics of image and merges to form multi-characteristic;In order to quickly obtain object candidate frame, in convolutional layer 5
Target candidate frame is generated according to certain length-width ratio and size on characteristic pattern, and is mapped on multi-characteristic and is cut;In order to obtain
The information of more candidate frames is obtained, result will be cut and carry out multiple features connection, and be input to full articulamentum;In order to realize precise elevation
Classification and positioning, carry out the classification of Softmax graders and the positioning of Overlap loss functions iterative regression, realize complete mesh
The classification and positioning of detection are marked, the Detection results better than other main flow target detection frameworks such as Fatser R-CNN have been obtained.
2nd, the present invention is reasonable in design, and it carries out multi-feature extraction using deep learning framework, obtains the multilayer feature of image
Represent, but realize more accurately classification;And employ a kind of new Overlap overlapping area loss functions in positioning,
The position of target-area object in input picture can be more accurately detected, has obtained good on the data set of target detection
Effect.
Brief description of the drawings
Fig. 1 is entire block diagram of the invention;
Fig. 2 generates fixed aspect ratio and the candidate frame of size for the present invention on the characteristic pattern of convolutional layer 5;
Fig. 3 is the Overlap loss functions between the candidate frame of the invention proposed in positioning and true encirclement frame;
Influences of the Fig. 4 for the different training iterations of the present invention to target detection precision;
Fig. 5 is the target detection accuracy table based on PASCAL VOC.
Embodiment
The embodiment of the present invention is further described below in conjunction with accompanying drawing.
It is a kind of to improve the method for target detection performance by improving target classification and positional accuracy, as shown in figure 1, first
First, in order to obtain more image informations, the picture with the true encirclement frame of object (Ground Truth) is input to VGG-
Characteristics of image is extracted in 16 convolutional neural networks, then carrying out multilayer to characteristics of image merges to form multi-characteristic;Then in order to
Object candidate frame quickly is obtained, target candidate frame is generated according to certain length-width ratio and size on the characteristic pattern of convolutional layer 5, and map
Cut on to multi-characteristic;Then in order to obtain the information of more candidate frames, result will be cut and carry out multiple features connection, and
It is input to full articulamentum;Last classification and positioning in order to realize precise elevation, carry out the classification of Softmax graders and
Overlap loss functions iterative regression is positioned, and realizes the classification and positioning of the target detection that complete precision is improved.Below with one
Individual instantiation is illustrated:
S1, based on VGG-16 convolutional neural networks framework extract characteristics of image, and by convolutional layer 1,2,3 and 5 carry out feature
Fusion, forms the characteristic pattern of multiple features;
S2, mesh generation is carried out on convolutional layer 5, the target candidate of fixed number and size is predicted in each network
Frame;
S3, candidate frame is mapped on characteristic pattern cut, then will cut result and carry out multiple features connection;、
After S4, the above results are by full articulamentum, characteristics of image is classified by Softmax sorting algorithms, is used in combination
Overlap space wastages function carries out online iterative regression positioning, obtains the result of final target detection.
In the present embodiment, the step S1 further comprises:
S1.1, the picture with the true encirclement frame of object is input to VGG-16 convolutional neural networks frameworks first, passed through
Caffe extracts feature of the image by convolutional neural networks different layers;
S1.2, the characteristics of image for exporting convolutional layer 1,2 carry out maximum pondization and operated, and the image that convolutional layer 5 is exported
Feature carries out deconvolution operation, realizes that output characteristic of the size of output all with convolutional layer 3 is in the same size;
S1.3, the feature for finally exporting convolutional layer 1,2,3 and 5 are merged, and obtain the spy of the multi-feature extraction of image
Levy figure.
Fig. 2 is given in the present invention and carries out mesh generation on the characteristic pattern of convolutional layer 5, and fixed length is generated in each grid
Wide 4 candidate frames than with size, the step S2 further comprises:
S2.1, convolutional layer 5 export characteristic pattern on carry out 6*6 mesh generation;
S2.2, predict 4 in the small lattice middle of each single network and may include the candidate frame of object, this 4 candidates
Frame has fixed size and length-width ratio, and length-width ratio is respectively 1:1、1:2 and 2:1, only for 1:The candidate frame of 1 length-width ratio, I
To set 2 kinds of candidate frame sizes be respectively 32*32 pixels and 64*64 pixels;
S2.3, during network training, we are matched the true encirclement frame and candidate frame of object, pass through two
The IOU overlapping areas of person are screened more than or equal to 0.7, and delete the candidate frame beyond image boundary;
S2.4, about 100 candidate frames are finally generated on the characteristic pattern of convolutional layer 5.
In the present embodiment, the step S3 further comprises:
100 candidate frames generated on S3.1, the characteristic pattern according to convolutional layer 5, map corresponding according to its position
On multilayer feature figure, and cut accordingly on multilayer feature figure;
S3.2, the characteristic pattern square after cutting done into 1*1 convolution, in order to retain preceding layer can the visual field, and subtract
Then convolution results are carried out 3*3 convolution and 5*5 convolution by few amount of calculation respectively;
S3.3, in order to obtain full text information, by multilayer feature figure by maximum pond layer, then by 1*1 convolutional layers with
Active coating, can halve amount of calculation;
S3.4, the convolution output result of 1*1 convolution, 3*3 convolution, 5*5 convolution and full text information entered according to tandem
Row connection, forms the multiple features connection of candidate frame.
In the present embodiment, the step S4 further comprises:
S4.1, above by convolutional layer and multiple features connection after result by 3 layers of full articulamentum after, pass through Softmax
Sorting algorithm is classified to characteristics of image, based on PASCAL VOC data sets, and classification results include 20 type objects, per type objects
With oneself corresponding precision;
S4.2, pass through Overlap space wastage function pairs candidate frame carry out recurrence positioning so that candidate frame is closer to thing
The true encirclement frame of body, the loss function is candidate frame and the common factor area divided by union area of true encirclement frame, and the numerical value is got over
Close to 1 explanation the two closer to;
S4.3, according to Softmax loss and Overlap penalty values be ranked up, positive sample and negative sample are filtered out online
Ratio be 3:1, renewal Sample Storehouse, which is input on multilayer feature figure, proceeds iterative regression positioning;
After S4.4, iteration n times, candidate frame closer to object true encirclement frame, model training well after can carry out
The test of actual object.
Fig. 3 shows the Overlap losses between the candidate frame proposed in the present invention in positioning and true encirclement frame
Function, the step S4.2 further comprises:
The true encirclement frame (Ground Truth) of object, its upper left corner and the lower right corner are included in S4.2.1, input picture
Coordinate constitute 4 dimensional vectors, beThe object candidate frame predicted by inventive algorithm,
The coordinate of its upper left corner () and the lower right corner () constitutes 4 dimensional vectors, is x=(x1,y1,x2,y2);
S4.2.2, traditional coordinate loss function are one-dimensional loss function, and the loss between each coordinate points is summed to come
Overall position skew loss is calculated, but traditional method individually makes a distinction coordinate, it is impossible to overall prediction is truly wrapped
Skew between peripheral frame and candidate frame is lost, and the formula of traditional one-dimensional coordinate loss function is:
S4.2.3, the skew loss predicted between true encirclement frame and candidate frame for entirety, we have proposed
4 dimension coordinates are carried out overall recurrence and calculated, calculate the area between true encirclement frame and candidate frame by Overlap loss functions
Skew loss.Wherein I represents common factor area therebetween, and U represents union area therebetween, by common factor area divided by
Union area evaluates position deviation therebetween, and the numerical value illustrates that position coordinates regression effect is better closer to 1.It is described
The formula of two-dimensional areas Overlap loss functions is:
I=(x2′-x1′)×(y2′-y1′)
Fig. 4 shows influence of the training iterationses different in the present invention to target detection precision, the step
S4.4 further comprises:
S4.4.1, in the training process according to Softmax loss and Overlap penalty values be ranked up, filter out 3:1
Positive and negative difficult sample, and the sample is re-entered into multilayer feature figure cut, multiple features connection is then proceeded by,
By screening difficult sample, the robustness and accuracy of detection for lifting system proposed by the present invention are realized;
S4.4.2, as seen in Figure 4, after successive ignition is trained, the nicety of grading of target detection has been obtained very
Big lifting, 1 Iterative classification precision proceeds iteration it is observed that precision is increased rapidly, but is passed through 42% or so
Cross after 4 iteration, the increasing degree of precision is smaller, therefore balance accuracy and speed, we select the repetitive exercise number of times of system
For 4, the nicety of grading of target detection and the overall raising of regression accuracy are obtained.
Method below as the present invention is tested, and illustrates the experiment effect of the present invention.
Test environment:MATLAB 2014b;Caffe;Ubuntu14.04 systems;NVIDIA GTX 1070p GPU
Cycle tests:Selected cycle tests and the true encirclement frame (Ground of its correspondence standard target detection object
Truth), be all from target detection PASCAL VOC data sets (M.Everingham, L.Van Gool, C.K.Williams,
J.Winn,and A.Zisserman,“The pascal visual object classes(voc)challenge,”
International journal of computer vision,vol.88,no.2,pp.303–338,2007.).Wherein wrap
The legend contained has 20 classifications, the respectively mankind;Animal (bird, cat, ox, dog, horse, sheep);The vehicles (aircraft, bicycle,
Ship, bus, car, motorcycle, train);Indoor (bottle, chair, dining table, potted plant, sofa, TV).From
Target be it is daily in most common object, be be exactly can preferably embody the practicality of algorithm, altogether comprising 9,963 figure
Piece, there is 24,640 labeled target objects.
Test index:Present invention uses two kinds of evaluation indexes, respectively precision mAP (mean average
) and speed fps (frames per second) precision.Wherein precision mAP is the bat of object detection results
Measurement, is compared with dreamboat testing result and is weighted average computation to all objects classification in database, to not
This parameter value is calculated with algorithm, it was demonstrated that inventive algorithm obtains preferable result in object detection field;Speed fps is target inspection
The measurement of the speed of result is surveyed, how many frame pictures can be handled by per second in test process and enters degree of testing the speed come evaluating target, it is right
Algorithms of different calculates this parameter value, it was demonstrated that superiority of the inventive algorithm in object detection field.
Test result as shown in figure 5, Fig. 5 be based in PASCAL VOC data sets all image category measuring accuracies it is flat
Equal result, it can be seen that inventive algorithm is significantly improved on mAP compared to other algorithm of target detection, wherein 4 generation of the invention
Table loop iteration is trained 4 times, and the present invention 6 represents loop iteration and trained 6 times.The best result Faster R- of current target detection
CNN mAP is 73.2%, and the mAP of the present invention 6 is 74.2%, and the accuracy of detection than Faster R-CNN improves 1.0%.And
And, in wisp detection, such as bottle, aircraft and plant, inventive algorithm obtain higher accuracy of detection, example than other algorithms
Wisp plant is such as directed to, inventive algorithm reaches 50.4%mAP, 11.6%mAP is higher by than Faster R-CNN.The above results
Object detection results produced by showing inventive algorithm possess higher precision, and can preferably solve small target deteection
Problem.
Target detection speed of the table 1 based on PASCAL VOC
Table 1 is the result based on all image category detection speeds in PASCAL VOC data sets, it can be seen that the present invention
Algorithm is significantly improved on fps compared to other algorithm of target detection, carries out full articulamentum to block SVD wherein compression is represented
(singular value decomposition) compresses.The best result Faster R-CNN of target detection speed is at present
7fps, the speed of the present invention 4 of compression is 12fps, accelerates 2fps than uncompressed convolutional layer, the speed of the present invention 6 of compression is
12fps, 2fps is accelerated than uncompressed convolutional layer;And the speed of inventive algorithm is 22 times of Fast R-CNN, almost close to real
When detect.Object detection results produced by the above results show the present invention possess higher speed, and in target detection
In two indexs of speed and precision, best object detection results can be reached, illustrate that inventive algorithm has frontier nature.
It is emphasized that embodiment of the present invention is illustrative, rather than it is limited, therefore present invention bag
Include and be not limited to embodiment described in embodiment, it is every by those skilled in the art's technique according to the invention scheme
The other embodiment drawn, also belongs to the scope of protection of the invention.
Claims (5)
1. a kind of improve the method for target detection performance by improving target classification and positional accuracy, it is characterised in that including with
Lower step:
Step 1, extract characteristics of image according to convolutional neural networks framework, and select before convolutional layer M layer output progress Fusion Features,
Form the characteristic pattern of multiple features;
Step 2, mesh generation is carried out on convolutional layer M, the target candidate frame of fixed number and size is predicted in each network;
Step 3, candidate frame is mapped on characteristic pattern cut, then will cut result and carry out multiple features connection;
Step 4, by the above results by full articulamentum after, characteristics of image is classified by Softmax sorting algorithms, is used in combination
Overlapping area loss function carries out online iterative regression positioning, obtains the result of final target detection.
2. according to claim 1 improve the method for target detection performance by improving target classification and positional accuracy,
It is characterized in that:The specific method of the step 1 comprises the following steps:
(1) the picture with the true encirclement frame of object is input in convolutional neural networks framework first, is extracted and schemed by Caffe
As the feature by convolutional neural networks different layers;
(2) the characteristics of image forward convolutional layer exported carries out maximum pondization and operated, and the characteristics of image that convolutional layer M is exported
Deconvolution operation is carried out, realizes that output characteristic of the size of output all with middle convolutional layer is in the same size;
(3) the last feature for exporting all convolutional layers is merged, and obtains the characteristic pattern of the multi-feature extraction of image.
3. according to claim 1 improve the method for target detection performance by improving target classification and positional accuracy,
It is characterized in that:The implementation method of the step 2 comprises the following steps:
(1) 6*6 mesh generation is carried out on the characteristic pattern that convolutional layer M is exported;
(2) the candidate frame of object may be included by predicting 4 in the small lattice middle of each single network, and this 4 candidate frames have solid
Fixed size and length-width ratio, length-width ratio is respectively 1:1、1:2 and 2:1, only for 1:The candidate frame of 1 length-width ratio, sets 2 kinds of candidates
Frame size 0.6 and 0.9;
(3) during network training, we are matched the true encirclement frame and candidate frame of object, pass through the IOU of the two
Overlapping area is screened more than or equal to 0.7, and deletes the candidate frame beyond image boundary;
(4) 100 candidate frames are finally generated on convolutional layer M characteristic pattern.
4. according to claim 1 improve the method for target detection performance by improving target classification and positional accuracy,
It is characterized in that:The implementation method of the step 3 comprises the following steps:
(1), according to 100 candidate frames generated on convolutional layer M characteristic pattern, corresponding multilayer feature is mapped according to its position
On figure, and cut accordingly on multilayer feature figure;
(2) the characteristic pattern square after cutting is done into 1*1 convolution, 3*3 convolution and 5*5 convolution are then carried out respectively to convolution results;
(3) in order to obtain full text information, by multilayer feature figure by maximum pond layer, then by 1*1 convolutional layers and active coating;
(4) the convolution output result of 1*1 convolution, 3*3 convolution, 5*5 convolution and full text information is attached according to tandem,
Form the multiple features connection of candidate frame.
5. the degree of accuracy according to claim 1 by improving target classification and positioning improves target detection performance, it is special
Levy and be:The concrete methods of realizing of the step 4 comprises the following steps:
By full articulamentum after, characteristics of image is classified by Softmax sorting algorithms, the data based on target detection
Collection, has oneself corresponding precision per type objects;
(2) recurrence positioning is carried out to candidate frame by overlapping area loss function so that true bag of the candidate frame closer to object
Peripheral frame, the loss function is candidate frame and the common factor area divided by union area of true encirclement frame;
(3) it is ranked up according to Softmax losses and overlapping area penalty values, positive sample and the ratio of negative sample is filtered out online
For 3:1, renewal Sample Storehouse, which is input on multilayer feature figure, proceeds iterative regression positioning;
(4) after iteration n times, candidate frame closer to object true encirclement frame, model training well after can carry out actual object
Test.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710450327.2A CN107316058A (en) | 2017-06-15 | 2017-06-15 | Improve the method for target detection performance by improving target classification and positional accuracy |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710450327.2A CN107316058A (en) | 2017-06-15 | 2017-06-15 | Improve the method for target detection performance by improving target classification and positional accuracy |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107316058A true CN107316058A (en) | 2017-11-03 |
Family
ID=60181717
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710450327.2A Pending CN107316058A (en) | 2017-06-15 | 2017-06-15 | Improve the method for target detection performance by improving target classification and positional accuracy |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107316058A (en) |
Cited By (38)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108171112A (en) * | 2017-12-01 | 2018-06-15 | 西安电子科技大学 | Vehicle identification and tracking based on convolutional neural networks |
CN108205687A (en) * | 2018-02-01 | 2018-06-26 | 通号通信信息集团有限公司 | Based on focus mechanism positioning loss calculation method and system in object detection system |
CN108229341A (en) * | 2017-12-15 | 2018-06-29 | 北京市商汤科技开发有限公司 | Sorting technique and device, electronic equipment, computer storage media, program |
CN108229477A (en) * | 2018-01-25 | 2018-06-29 | 深圳市商汤科技有限公司 | For visual correlation recognition methods, device, equipment and the storage medium of image |
CN108805064A (en) * | 2018-05-31 | 2018-11-13 | 中国农业大学 | A kind of fish detection and localization and recognition methods and system based on deep learning |
CN108805210A (en) * | 2018-06-14 | 2018-11-13 | 深圳深知未来智能有限公司 | A kind of shell hole recognition methods based on deep learning |
CN108830280A (en) * | 2018-05-14 | 2018-11-16 | 华南理工大学 | A kind of small target detecting method based on region nomination |
CN108830131A (en) * | 2018-04-10 | 2018-11-16 | 中科院微电子研究所昆山分所 | Traffic target detection and distance measuring method based on deep learning |
CN108830224A (en) * | 2018-06-19 | 2018-11-16 | 武汉大学 | A kind of high-resolution remote sensing image Ship Target Detection method based on deep learning |
CN108875577A (en) * | 2018-05-11 | 2018-11-23 | 深圳市易成自动驾驶技术有限公司 | Object detection method, device and computer readable storage medium |
CN109086779A (en) * | 2018-07-28 | 2018-12-25 | 天津大学 | A kind of attention target identification method based on convolutional neural networks |
CN109447066A (en) * | 2018-10-18 | 2019-03-08 | 中国人民武装警察部队海警学院 | A kind of quick accurately single phase object detection method and device |
CN109493370A (en) * | 2018-10-12 | 2019-03-19 | 西南交通大学 | A kind of method for tracking target based on spatial offset study |
CN109492685A (en) * | 2018-10-31 | 2019-03-19 | 中国矿业大学 | A kind of target object visible detection method for symmetrical feature |
CN109508672A (en) * | 2018-11-13 | 2019-03-22 | 云南大学 | A kind of real-time video object detection method |
CN109583483A (en) * | 2018-11-13 | 2019-04-05 | 中国科学院计算技术研究所 | A kind of object detection method and system based on convolutional neural networks |
CN109685008A (en) * | 2018-12-25 | 2019-04-26 | 云南大学 | A kind of real-time video object detection method |
CN109697464A (en) * | 2018-12-17 | 2019-04-30 | 环球智达科技(北京)有限公司 | Method and system based on the identification of the precision target of object detection and signature search |
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 |
CN109784349A (en) * | 2018-12-25 | 2019-05-21 | 东软集团股份有限公司 | Image object detection model method for building up, device, storage medium and program product |
CN109816012A (en) * | 2019-01-22 | 2019-05-28 | 南京邮电大学 | A kind of multiscale target detection method of integrating context information |
CN109918951A (en) * | 2019-03-12 | 2019-06-21 | 中国科学院信息工程研究所 | A kind of artificial intelligence process device side channel system of defense based on interlayer fusion |
CN110059667A (en) * | 2019-04-28 | 2019-07-26 | 上海应用技术大学 | Pedestrian counting method |
CN110110722A (en) * | 2019-04-30 | 2019-08-09 | 广州华工邦元信息技术有限公司 | A kind of region detection modification method based on deep learning model recognition result |
CN110222641A (en) * | 2019-06-06 | 2019-09-10 | 北京百度网讯科技有限公司 | The method and apparatus of image for identification |
CN110245675A (en) * | 2019-04-03 | 2019-09-17 | 复旦大学 | A kind of dangerous objects detection method based on millimeter-wave image human body contextual information |
WO2019179269A1 (en) * | 2018-03-21 | 2019-09-26 | 广州极飞科技有限公司 | Method and apparatus for acquiring boundary of area to be operated, and operation route planning method |
CN110348384A (en) * | 2019-07-12 | 2019-10-18 | 沈阳理工大学 | A kind of Small object vehicle attribute recognition methods based on Fusion Features |
CN110555354A (en) * | 2018-05-31 | 2019-12-10 | 北京深鉴智能科技有限公司 | Feature screening method and apparatus, target detection method and apparatus, electronic apparatus, and storage medium |
CN110610184A (en) * | 2018-06-15 | 2019-12-24 | 阿里巴巴集团控股有限公司 | Method, device and equipment for detecting salient object of image |
CN110874641A (en) * | 2018-08-29 | 2020-03-10 | 松下电器(美国)知识产权公司 | Information processing method and information processing system |
CN110909604A (en) * | 2019-10-23 | 2020-03-24 | 深圳市华讯方舟太赫兹科技有限公司 | Security image detection method, terminal device and computer storage medium |
CN110956060A (en) * | 2018-09-27 | 2020-04-03 | 北京市商汤科技开发有限公司 | Motion recognition method, driving motion analysis method, device and electronic equipment |
CN111126421A (en) * | 2018-10-31 | 2020-05-08 | 浙江宇视科技有限公司 | Target detection method, device and readable storage medium |
CN111160353A (en) * | 2019-12-27 | 2020-05-15 | 广州亚信技术有限公司 | License plate recognition method, device and equipment |
CN111325075A (en) * | 2018-12-17 | 2020-06-23 | 北京华航无线电测量研究所 | Video sequence target detection method |
CN111968087A (en) * | 2020-08-13 | 2020-11-20 | 中国农业科学院农业信息研究所 | Plant disease area detection method |
US10902314B2 (en) | 2018-09-19 | 2021-01-26 | Industrial Technology Research Institute | Neural network-based classification method and classification device thereof |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104881662A (en) * | 2015-06-26 | 2015-09-02 | 北京畅景立达软件技术有限公司 | Single-image pedestrian detection method |
CN105488758A (en) * | 2015-11-30 | 2016-04-13 | 河北工业大学 | Image scaling method based on content awareness |
CN106127204A (en) * | 2016-06-30 | 2016-11-16 | 华南理工大学 | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks |
US9501724B1 (en) * | 2015-06-09 | 2016-11-22 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
US20170083792A1 (en) * | 2015-09-22 | 2017-03-23 | Xerox Corporation | Similarity-based detection of prominent objects using deep cnn pooling layers as features |
CN106650699A (en) * | 2016-12-30 | 2017-05-10 | 中国科学院深圳先进技术研究院 | CNN-based face detection method and device |
CN106650725A (en) * | 2016-11-29 | 2017-05-10 | 华南理工大学 | Full convolutional neural network-based candidate text box generation and text detection method |
-
2017
- 2017-06-15 CN CN201710450327.2A patent/CN107316058A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9501724B1 (en) * | 2015-06-09 | 2016-11-22 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
US20160364633A1 (en) * | 2015-06-09 | 2016-12-15 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
CN104881662A (en) * | 2015-06-26 | 2015-09-02 | 北京畅景立达软件技术有限公司 | Single-image pedestrian detection method |
US20170083792A1 (en) * | 2015-09-22 | 2017-03-23 | Xerox Corporation | Similarity-based detection of prominent objects using deep cnn pooling layers as features |
CN105488758A (en) * | 2015-11-30 | 2016-04-13 | 河北工业大学 | Image scaling method based on content awareness |
CN106127204A (en) * | 2016-06-30 | 2016-11-16 | 华南理工大学 | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks |
CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
CN106650725A (en) * | 2016-11-29 | 2017-05-10 | 华南理工大学 | Full convolutional neural network-based candidate text box generation and text detection method |
CN106650699A (en) * | 2016-12-30 | 2017-05-10 | 中国科学院深圳先进技术研究院 | CNN-based face detection method and device |
Non-Patent Citations (1)
Title |
---|
任少卿: "基于特征共享的高效物体检测", 《中国博士学位论文全文数据库(信息科技辑)》 * |
Cited By (58)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108171112A (en) * | 2017-12-01 | 2018-06-15 | 西安电子科技大学 | Vehicle identification and tracking based on convolutional neural networks |
CN108171112B (en) * | 2017-12-01 | 2021-06-01 | 西安电子科技大学 | Vehicle identification and tracking method based on convolutional neural network |
CN108229341B (en) * | 2017-12-15 | 2021-08-06 | 北京市商汤科技开发有限公司 | Classification method and device, electronic equipment and computer storage medium |
CN108229341A (en) * | 2017-12-15 | 2018-06-29 | 北京市商汤科技开发有限公司 | Sorting technique and device, electronic equipment, computer storage media, program |
CN108229477A (en) * | 2018-01-25 | 2018-06-29 | 深圳市商汤科技有限公司 | For visual correlation recognition methods, device, equipment and the storage medium of image |
CN108229477B (en) * | 2018-01-25 | 2020-10-09 | 深圳市商汤科技有限公司 | Visual relevance identification method, device, equipment and storage medium for image |
CN108205687B (en) * | 2018-02-01 | 2022-04-01 | 通号通信信息集团有限公司 | Attention mechanism-based positioning loss calculation method and system in target detection system |
CN108205687A (en) * | 2018-02-01 | 2018-06-26 | 通号通信信息集团有限公司 | Based on focus mechanism positioning loss calculation method and system in object detection system |
WO2019179269A1 (en) * | 2018-03-21 | 2019-09-26 | 广州极飞科技有限公司 | Method and apparatus for acquiring boundary of area to be operated, and operation route planning method |
CN108830131A (en) * | 2018-04-10 | 2018-11-16 | 中科院微电子研究所昆山分所 | Traffic target detection and distance measuring method based on deep learning |
CN108830131B (en) * | 2018-04-10 | 2021-05-04 | 昆山微电子技术研究院 | Deep learning-based traffic target detection and ranging method |
CN108875577A (en) * | 2018-05-11 | 2018-11-23 | 深圳市易成自动驾驶技术有限公司 | Object detection method, device and computer readable storage medium |
CN108830280B (en) * | 2018-05-14 | 2021-10-26 | 华南理工大学 | Small target detection method based on regional nomination |
CN108830280A (en) * | 2018-05-14 | 2018-11-16 | 华南理工大学 | A kind of small target detecting method based on region nomination |
CN108805064A (en) * | 2018-05-31 | 2018-11-13 | 中国农业大学 | A kind of fish detection and localization and recognition methods and system based on deep learning |
CN110555354A (en) * | 2018-05-31 | 2019-12-10 | 北京深鉴智能科技有限公司 | Feature screening method and apparatus, target detection method and apparatus, electronic apparatus, and storage medium |
CN108805210B (en) * | 2018-06-14 | 2022-03-04 | 深圳深知未来智能有限公司 | Bullet hole identification method based on deep learning |
CN108805210A (en) * | 2018-06-14 | 2018-11-13 | 深圳深知未来智能有限公司 | A kind of shell hole recognition methods based on deep learning |
CN110610184A (en) * | 2018-06-15 | 2019-12-24 | 阿里巴巴集团控股有限公司 | Method, device and equipment for detecting salient object of image |
CN110610184B (en) * | 2018-06-15 | 2023-05-12 | 阿里巴巴集团控股有限公司 | Method, device and equipment for detecting salient targets of images |
CN108830224A (en) * | 2018-06-19 | 2018-11-16 | 武汉大学 | A kind of high-resolution remote sensing image Ship Target Detection method based on deep learning |
CN108830224B (en) * | 2018-06-19 | 2021-04-02 | 武汉大学 | High-resolution remote sensing image ship target detection method based on deep learning |
CN109086779A (en) * | 2018-07-28 | 2018-12-25 | 天津大学 | A kind of attention target identification method based on convolutional neural networks |
CN109086779B (en) * | 2018-07-28 | 2021-11-09 | 天津大学 | Attention target identification method based on convolutional neural network |
CN110874641A (en) * | 2018-08-29 | 2020-03-10 | 松下电器(美国)知识产权公司 | Information processing method and information processing system |
US10902314B2 (en) | 2018-09-19 | 2021-01-26 | Industrial Technology Research Institute | Neural network-based classification method and classification device thereof |
CN110956060A (en) * | 2018-09-27 | 2020-04-03 | 北京市商汤科技开发有限公司 | Motion recognition method, driving motion analysis method, device and electronic equipment |
CN109493370A (en) * | 2018-10-12 | 2019-03-19 | 西南交通大学 | A kind of method for tracking target based on spatial offset study |
CN109493370B (en) * | 2018-10-12 | 2021-07-02 | 西南交通大学 | Target tracking method based on space offset learning |
CN109447066B (en) * | 2018-10-18 | 2021-08-20 | 中国人民武装警察部队海警学院 | Rapid and accurate single-stage target detection method and device |
CN109447066A (en) * | 2018-10-18 | 2019-03-08 | 中国人民武装警察部队海警学院 | A kind of quick accurately single phase object detection method and device |
CN109492685A (en) * | 2018-10-31 | 2019-03-19 | 中国矿业大学 | A kind of target object visible detection method for symmetrical feature |
CN111126421A (en) * | 2018-10-31 | 2020-05-08 | 浙江宇视科技有限公司 | Target detection method, device and readable storage medium |
CN109492685B (en) * | 2018-10-31 | 2022-05-24 | 煤炭科学研究总院 | Target object visual detection method for symmetric characteristics |
CN109583483B (en) * | 2018-11-13 | 2020-12-11 | 中国科学院计算技术研究所 | Target detection method and system based on convolutional neural network |
CN109508672A (en) * | 2018-11-13 | 2019-03-22 | 云南大学 | A kind of real-time video object detection method |
CN109583483A (en) * | 2018-11-13 | 2019-04-05 | 中国科学院计算技术研究所 | A kind of object detection method and system based on convolutional neural networks |
CN109697464A (en) * | 2018-12-17 | 2019-04-30 | 环球智达科技(北京)有限公司 | Method and system based on the identification of the precision target of object detection and signature search |
CN111325075A (en) * | 2018-12-17 | 2020-06-23 | 北京华航无线电测量研究所 | Video sequence target detection method |
CN111325075B (en) * | 2018-12-17 | 2023-11-07 | 北京华航无线电测量研究所 | Video sequence target detection method |
CN109685008A (en) * | 2018-12-25 | 2019-04-26 | 云南大学 | A kind of real-time video object detection method |
CN109784349A (en) * | 2018-12-25 | 2019-05-21 | 东软集团股份有限公司 | Image object detection model method for building up, device, storage medium and program product |
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 |
CN109816012A (en) * | 2019-01-22 | 2019-05-28 | 南京邮电大学 | A kind of multiscale target detection method of integrating context information |
CN109816012B (en) * | 2019-01-22 | 2022-07-12 | 南京邮电大学 | Multi-scale target detection method fusing context information |
CN109918951A (en) * | 2019-03-12 | 2019-06-21 | 中国科学院信息工程研究所 | A kind of artificial intelligence process device side channel system of defense based on interlayer fusion |
CN110245675B (en) * | 2019-04-03 | 2023-02-10 | 复旦大学 | Dangerous object detection method based on millimeter wave image human body context information |
CN110245675A (en) * | 2019-04-03 | 2019-09-17 | 复旦大学 | A kind of dangerous objects detection method based on millimeter-wave image human body contextual information |
CN110059667A (en) * | 2019-04-28 | 2019-07-26 | 上海应用技术大学 | Pedestrian counting method |
CN110110722A (en) * | 2019-04-30 | 2019-08-09 | 广州华工邦元信息技术有限公司 | A kind of region detection modification method based on deep learning model recognition result |
CN110222641A (en) * | 2019-06-06 | 2019-09-10 | 北京百度网讯科技有限公司 | The method and apparatus of image for identification |
CN110348384B (en) * | 2019-07-12 | 2022-06-17 | 沈阳理工大学 | Small target vehicle attribute identification method based on feature fusion |
CN110348384A (en) * | 2019-07-12 | 2019-10-18 | 沈阳理工大学 | A kind of Small object vehicle attribute recognition methods based on Fusion Features |
CN110909604A (en) * | 2019-10-23 | 2020-03-24 | 深圳市华讯方舟太赫兹科技有限公司 | Security image detection method, terminal device and computer storage medium |
CN110909604B (en) * | 2019-10-23 | 2024-04-19 | 深圳市重投华讯太赫兹科技有限公司 | Security check image detection method, terminal equipment and computer storage medium |
CN111160353A (en) * | 2019-12-27 | 2020-05-15 | 广州亚信技术有限公司 | License plate recognition method, device and equipment |
CN111968087A (en) * | 2020-08-13 | 2020-11-20 | 中国农业科学院农业信息研究所 | Plant disease area detection method |
CN111968087B (en) * | 2020-08-13 | 2023-11-07 | 中国农业科学院农业信息研究所 | Plant disease area detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107316058A (en) | Improve the method for target detection performance by improving target classification and positional accuracy | |
CN107563381B (en) | Multi-feature fusion target detection method based on full convolution network | |
CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
Wang et al. | Autonomous garbage detection for intelligent urban management | |
Ouyang et al. | DeepID-Net: Object detection with deformable part based convolutional neural networks | |
CN108171112A (en) | Vehicle identification and tracking based on convolutional neural networks | |
Zhang et al. | Pedestrian detection method based on Faster R-CNN | |
CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
Wan et al. | Ceramic tile surface defect detection based on deep learning | |
CN107818302A (en) | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks | |
CN107451602A (en) | A kind of fruits and vegetables detection method based on deep learning | |
CN106446930A (en) | Deep convolutional neural network-based robot working scene identification method | |
CN109932730A (en) | Laser radar object detection method based on multiple dimensioned monopole three dimensional detection network | |
CN110321891A (en) | A kind of big infusion medical fluid foreign matter object detection method of combined depth neural network and clustering algorithm | |
CN107808376A (en) | A kind of detection method of raising one's hand based on deep learning | |
CN105243139A (en) | Deep learning based three-dimensional model retrieval method and retrieval device thereof | |
CN106127161A (en) | Fast target detection method based on cascade multilayer detector | |
CN110569926B (en) | Point cloud classification method based on local edge feature enhancement | |
CN105787488A (en) | Image feature extraction method and device realizing transmission from whole to local | |
Xu et al. | Occlusion problem-oriented adversarial faster-RCNN scheme | |
Yang et al. | Road crack detection using deep neural network with receptive field block | |
Ouadiay et al. | Simultaneous object detection and localization using convolutional neural networks | |
Zhang et al. | A precise apple leaf diseases detection using BCTNet under unconstrained environments | |
Wang et al. | A review of object detection based on convolutional neural networks and deep learning | |
Zhang et al. | Multiple Objects Detection based on Improved Faster R-CNN |
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 |
Application publication date: 20171103 |
|
RJ01 | Rejection of invention patent application after publication |