CN110378232A - The examination hall examinee position rapid detection method of improved SSD dual network - Google Patents
The examination hall examinee position rapid detection method of improved SSD dual network Download PDFInfo
- Publication number
- CN110378232A CN110378232A CN201910534569.9A CN201910534569A CN110378232A CN 110378232 A CN110378232 A CN 110378232A CN 201910534569 A CN201910534569 A CN 201910534569A CN 110378232 A CN110378232 A CN 110378232A
- Authority
- CN
- China
- Prior art keywords
- convolution
- ssd
- convolutional layer
- network
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000009977 dual effect Effects 0.000 title claims abstract description 43
- 238000001514 detection method Methods 0.000 title claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 113
- 238000012360 testing method Methods 0.000 claims abstract description 72
- 238000005070 sampling Methods 0.000 claims abstract description 65
- 238000000034 method Methods 0.000 claims abstract description 32
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000012956 testing procedure Methods 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 21
- 230000000644 propagated effect Effects 0.000 claims description 20
- 238000011478 gradient descent method Methods 0.000 claims description 10
- 238000011156 evaluation Methods 0.000 claims description 7
- 230000001629 suppression Effects 0.000 claims description 7
- 238000003475 lamination Methods 0.000 claims description 4
- 238000005259 measurement Methods 0.000 abstract description 2
- 238000012795 verification Methods 0.000 abstract description 2
- 238000007689 inspection Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000012935 Averaging Methods 0.000 description 4
- 241000208340 Araliaceae Species 0.000 description 3
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 3
- 235000003140 Panax quinquefolius Nutrition 0.000 description 3
- 230000006399 behavior Effects 0.000 description 3
- 235000008434 ginseng Nutrition 0.000 description 3
- 238000012800 visualization Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/20—Education
- G06Q50/205—Education administration or guidance
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Educational Administration (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Tourism & Hospitality (AREA)
- Strategic Management (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Educational Technology (AREA)
- Computational Linguistics (AREA)
- Multimedia (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
Abstract
A kind of examination hall examinee position rapid detection method of improved SSD dual network by image preprocessing, building dynamic threshold SSD network, building up-sampling SSD network, the improved SSD dual network of training and forms test sample image progress testing procedure.On the basis of SSD network structure, the characteristics of according to examination hall examinee's data set, building dynamic adjustment hands over and compares threshold method, increase the positive sample amount of SSD network Small Target, up-sampling layer is added in SSD network, enhance Small object characteristics of image, the thought of integrating parallel dual network structure improves image object verification and measurement ratio.Compared with prior art, the good, high accuracy for examination with robustness is suitable for detecting examinee and invigilator teacher under examination scene.
Description
Technical field
The invention belongs to image procossing and target detection technique fields, particularly relate in standardization examination hall monitor video
The identification, positioning and demographics of examinee in the single-frame images of acquisition.
Background technique
Examination is one of education activities important link, is the major way for measuring, judging student's attainment level, is me
The important means they state all types of talents examination and selected.In all kinds of important examinations of country in recent years, standardization examination hall has been played
Important function.It, inevitably can be because of for a long time, uninterruptedly if only relying on the artificial monitor video of observation in real time carries out identification judgement
Continuous work bring fatigue and there is the case where judging by accident or omitting, therefore advanced computer vision technique is incorporated into mark
Standardization examination hall monitors big data, intelligently analyze that examinee's behavior and being aided with manually invigilates using video monitoring as main means newly
Invigilator's form will be more more and more urgent.
It is the premise and basis of intellectual analysis examinee's behavior to the accurate detection of examinee position in standardization examination hall, and is based on
Examinee's behavior of single frames scene is related to the target detection technique in image, includes target identification and positioning.Examination hall monitor video list
The orientation of student of frame image and counting are an important application of the target detection technique in standardization examination hall, and process is related to very
More image procossings and image analysis technology, such as the extraction of the body characteristics of people, detection zone in image are handed over and compare calculating etc.
Aspect.Currently, image object detection method both domestic and external includes object detection method, base of the conventional method such as based on image threshold
In inter-frame difference and the object detection method of bilinear interpolation etc., the object detection method based on deep learning has Faster-
RNN, YOLO, SSD etc..When above-mentioned target identification and localization method are applied to the single-frame images of standardization examination hall monitor video, institute
Existing technical problem underlying is that object recognition rate is not high, and positional accuracy is low, or even a large amount of missing inspections occurs or can not calculate
Problem.This cause existing method standardize examination hall monitoring scene under by student's unusual checking for the purpose of using it is micro- its
It is micro-.
Summary of the invention
Technical problem to be solved by the present invention lies in overcoming the disadvantage that above-mentioned prior art speed is slow, accuracy is not high,
The examination hall examinee position for providing the improved SSD dual network that a kind of robustness is good, detection accuracy is high, speed is fast, accuracy is high is fast
Fast detection method.
Technical solution used by above-mentioned technical problem is solved to be made of following step:
(1) image preprocessing
It is concentrated from image data and chooses 600~800 training sample images and 80~280 test sample images, with double
The image of selection is normalized to 250 × 250~500 × 500 by pixel size by linear interpolation method.
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
Input layer is the training sample image that 600~800 Zhang great little are 250 × 250~500 × 500, is connected after input layer
The convolution 1 of VGG16, head and the tail sequentially connect 2~convolution of convolution 5 after convolution 1, and head and the tail sequentially connect full convolutional layer 6, complete after convolution 5
Convolutional layer 7, head and the tail sequentially connect 8~convolution of convolution 11, convolution 4, full convolutional layer 7, convolution 8, convolution 9, convolution after full convolutional layer 7
10 and convolution 11 after connect with output layer.
(b) building pre-selection frame matching process
In initial configuration, convolutional layer 4-3, the full convolutional layer 7, the convolutional layer 8-2 in convolution 8, convolution in convolution 4 are selected
The convolutional layer 10-2 in convolutional layer 9-2, convolution 10 in 9, the convolutional layer 11 in convolution 11 totally 6 convolutional layers, in 6 convolutional layers
On take the pre-selection frame of different scale to be matched with true tag frame, divide pre-selection the positive and negative sample set of frame, matching process be SSD net
The area of pre-selection frame and true tag frame in network is handed over and is compared greater than friendship and than threshold value IOU, is denoted as positive sample, be otherwise negative sample
This.
Determine friendship and the method than threshold value IOU are as follows: according to the ratio of the area of true tag frame and input image pixels size
Value determines to hand over and the threshold value IOU of ratio, dynamic adjust threshold value IOU and divide positive negative sample, true tag frame area and input picture picture
Plain size ratio≤0.01 hands over and is 0.4 than threshold value IOU, otherwise, hands over and is 0.5 than threshold value IOU, determine true tag as the following formula
The area S of frameGTbWith friendship and than threshold value IOU:
SGTb=(ymax-ymin)×(xmax-xmin)
Wherein xmax、xmin、ymax、yminThe respectively maxima and minima of the abscissa of true tag frame and ordinate;
SinputFor the area of input picture, trained network is constituted by dynamic threshold SSD network according to the matching process.
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is the training sample image that 600~800 Zhang great little are 250 × 250~500 × 500, is connected after input layer
The convolution 1 of VGG16, head and the tail sequentially connect convolution 2, convolution 3 after convolution 1, and head and the tail sequentially connect the convolutional layer of convolution 4 after convolution 3
4-1, convolutional layer 4-2, the up-sampling layer 4-4 that connection size is 76 × 76 after convolutional layer 4-2, connect convolution after up-sampling layer 4-4
After the prediction interval 4-5 that core number is 512, convolution kernel size is 3 × 3, convolutional layer 4-2 after connection convolutional layer 4-3, convolutional layer 4-3
Convolution 5 is connected, full convolutional layer 6, full convolutional layer 7 is linked in sequence after convolution 5 from beginning to end, head and the tail sequentially roll up by connection after full convolutional layer 7
8~convolution 11 is accumulated, is connect with output layer after convolutional layer 4-5, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11, structure
At up-sampling SSD network structure.
(b) the pre-selection frame matching process of building up-sampling SSD network
The convolution in prediction interval 4-5, full convolutional layer 7, convolution 8 in up-sampling SSD network structure, in selection convolution 4
Layer 8-2, the convolutional layer 9-2 in convolution 9, the convolutional layer 10-2 in convolution 10, totally 6 convolutional layers of the convolutional layer 11 in convolution 11,
Take the pre-selection frame of different scale to be matched on 6 convolutional layers with true tag frame, will pre-selection frame be divided into positive sample collection and
Negative sample collection, specific matching principle are to preselect the area of frame and true tag frame to hand over and compare and be greater than friendship and than threshold value IOU, are denoted as just
Otherwise sample is negative sample, wherein set friendship and than threshold value IOU as 0.5.
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Suddenly are as follows:
(a) setting classification number is 2, the t raining period of training set is 8~20 times, training batch size is 15~32, basis
Study rate parameter is 0.0002~0.001, learning rate decay factor is 0.7~0.95.
(b) pre-training is carried out to improved SSD dual network using VOC2007 data set, by the weight after pre-training and partially
Set the initial weight as improved SSD dual network and biasing.
(c) pretreated 600~800 Zhang great little training sample image for being 250 × 250~500 × 500 is input to
In dynamic threshold SSD network after pre-training, sets and intersect entropy function as loss function, reduce penalty values using gradient descent method
Propagated forward and backpropagation, iterative cycles propagated forward and backpropagation are carried out, and updates weighted value and the biasing of network,
Training 8~20 times, obtains trained dynamic threshold SSD network model.
(d) pretreated 600~800 Zhang great little training sample image for being 250 × 250~500 × 500 is input to
In up-sampling SSD network after pre-training, sets and intersect entropy function as loss function, reduce penalty values using gradient descent method and come
Propagated forward and backpropagation, iterative cycles propagated forward and backpropagation are carried out, and updates weighted value and the biasing of network, is instructed
Practice 8~20 times, obtains trained up-sampling SSD network model.
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.45~0.65.
(b) pretreated 80~280 test sample images are input to trained dynamic threshold SSD network simultaneously
It is tested in up-sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each
The confidence level of the position coordinates of target, classification and generic.
It (c) is more than registration threshold value with position registration of the non-maxima suppression algorithm to target in 2 groups of testing results
0.35~0.6 testing result carries out duplicate removal, retains the higher position coordinates of confidence level and its classification, obtains finally detecting knot
Fruit chooses average precision mAP value as evaluation index, evaluates testing result.
The building dynamic threshold SSD network step (2) of the invention the step of in (b), the pre-selection of the different scale
The width ω of framekWith height hkAre as follows:
Wherein, SkFor size basis value, SminMost preferably 0.1, SmaxMost preferably 0.9, respectively indicate smallest dimension and maximum
Scale, m be the convolutional layer number of prediction, most preferably 6, k be (1, m] integer, a be pre-selection different the ratio of width to height of frame, a ∈ { 1/
3,1/2,1,2,3}。
Training of the invention improved SSD dual network step (4) the step of in (a), the t raining period of training set is best
It is 10 times, training batch size most preferably 16.
In the step of carrying out testing procedure (5) to test sample image (a) of the invention, the classification confidence level threshold of target
Value most preferably 0.5.
In the step of carrying out testing procedure (5) to test sample image (c) of the invention, position is sat in testing result
Target registration threshold value most preferably 0.4.
The present invention is on the basis of SSD network, and the characteristics of according to examination hall examinee's data set, building dynamic adjustment, which is handed over, simultaneously compares threshold
Value method increases the positive sample amount of SSD network Small Target, and up-sampling layer is added in SSD network, and enhancing Small object image is special
Sign, the thought of integrating parallel dual network structure improve image object verification and measurement ratio.Compared with prior art, good with robustness, inspection
The advantages that precision is high is surveyed, under scene of taking an examination, the student classroom or teacher detection.
Detailed description of the invention
Fig. 1 is the flow chart of the embodiment of the present invention 1.
Fig. 2 is the SSD primitive network structure chart of embodiment 1.
Fig. 3 is the part-structure figure of the up-sampling SSD network of embodiment 1.
Fig. 4 is the original image that number is 302 in 1 examination hall examinee's data set of embodiment.
Fig. 5 is examinee's locating effect figure in Fig. 4.
Fig. 6 is the original image that number is 814 in 4 examination hall examinee's data set of embodiment.
Fig. 7 is examinee's locating effect figure of Fig. 6.
Fig. 8 is the original image that number is 475 in 5 examination hall examinee's data set of embodiment.
Fig. 9 is examinee's locating effect figure in Fig. 8.
Specific embodiment
The present invention is further detailed with example with reference to the accompanying drawing, but the present invention is not limited to following implementations
Example.
Embodiment 1
The image of the present embodiment carrys out examination hall monitor video under the standardization examination hall environment that freely modes such as online crawler obtain
Examination hall examinee's data set for constituting of single-frame images, the present embodiment uses 700 images in the data set as training set,
180 images are as test set, and training set is not be overlapped with test set.
In Fig. 1~4, the examination hall examinee position rapid detection method of the improved SSD dual network of the present embodiment, by following
Step composition:
(1) image preprocessing
It is concentrated from image data and chooses 700 training sample images and 180 test sample images, use bilinear interpolation
The image of selection is normalized to 300 × 300 by pixel size.
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
The SSD network structure of the present embodiment is selected from paper " the Single Shot Multibox of CVPR meeting in 2016
Detector ", SSD network structure is by input layer, convolution 1, convolution 2, convolution 3, convolution 4, convolution 5, full convolutional layer 6, full convolution
Layer 7, convolutional layer 8, convolutional layer 9, convolutional layer 10, convolutional layer 11 connect and compose.
Input layer is the training sample image that 700 Zhang great little are 300 × 300, and the convolution 1 of VGG16, volume are connected after input layer
Convolution 2 is connected after product 1, convolution 3 is connected after convolution 2, and convolution 4 is connected after convolution 3, convolution 5 is connected after convolution 4, is connected after convolution 5
Full convolutional layer 6 connects full convolutional layer 7 after full convolutional layer 6, convolution 8 is connected after full convolutional layer 7, connects convolution 9, convolution after convolution 8
Convolution 10 is connected after 9, connects convolution 11 after convolution 10, after convolution 4, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11
It is connect with output layer.
(b) building pre-selection frame matching process
In initial configuration, convolutional layer 4-3, the full convolutional layer 7, the convolutional layer 8-2 in convolution 8, convolution in convolution 4 are selected
The convolutional layer 10-2 in convolutional layer 9-2, convolution 10 in 9, the convolutional layer 11 in convolution 11 totally 6 convolutional layers, in 6 convolutional layers
On take the pre-selection frame of different scale to be matched with true tag frame, divide pre-selection the positive and negative sample set of frame, matching process be SSD net
The area of pre-selection frame and true tag frame in network is handed over and is compared greater than friendship and than threshold value IOU, is denoted as positive sample, be otherwise negative sample
This.
The width ω of the pre-selection frame of different scalekWith height hkAre as follows:
Wherein, SkFor size basis value, SminIt is 0.1, SmaxIt is 0.9, respectively indicates smallest dimension and out to out, m is
The convolutional layer number of prediction, be 6, k be (1, m] integer, a be pre-selection different the ratio of width to height of frame, a ∈ { 1/3,1/2,1,2,3 }.
Determine friendship and the method than threshold value IOU are as follows: determine according to the area of true tag frame and the ratio of input picture area
The threshold value IOU of fixed friendship and ratio, dynamic adjust threshold value IOU and divide positive negative sample, true tag frame area and input picture area ratio
≤ 0.01, it hands over and is 0.4 than threshold value IOU, otherwise, hand over and be 0.5 than threshold value IOU, determined by formula (1) and formula (2) true
The area S of label frameGTbWith friendship and than threshold value IOU:
SGTb=(ymax-ymin)×(xmax-xmin)
Wherein xmax、xmin、ymax、yminThe respectively maxima and minima of the abscissa of true tag frame and ordinate;
SinputFor the area of input picture, trained network is constituted by dynamic threshold SSD network according to the matching process.
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is the training sample image that 700 Zhang great little are 300 × 300, and the convolution 1 of VGG16, volume are connected after input layer
Head and the tail sequentially connect convolution 2, convolution 3 after product 1, and head and the tail sequentially connect convolutional layer 4-1, the convolutional layer 4-2 of convolution 4, volume after convolution 3
The up-sampling layer 4-4 that size is 76 × 76 is connected after lamination 4-2, connection convolution kernel number is 512, convolution after up-sampling layer 4-4
The prediction interval 4-5 that core size is 3 × 3 connects convolutional layer 4-3 after convolutional layer 4-2, connects convolution 5 after convolutional layer 4-3, after convolution 5
Full convolutional layer 6, full convolutional layer 7 is linked in sequence in head and the tail, and head and the tail sequentially connect 8~convolution of convolution 11, convolutional layer after full convolutional layer 7
It is connect with output layer after 4-5, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11, constitutes up-sampling SSD network knot
Structure;
(b) the pre-selection frame matching process of building up-sampling SSD network
The convolution in prediction interval 4-5, full convolutional layer 7, convolution 8 in up-sampling SSD network structure, in selection convolution 4
Layer 8-2, the convolutional layer 9-2 in convolution 9, the convolutional layer 10-2 in convolution 10, totally 6 convolutional layers of the convolutional layer 11 in convolution 11,
Take the pre-selection frame of different scale to be matched on 6 convolutional layers with true tag frame, will pre-selection frame be divided into positive sample collection and
Negative sample collection, specific matching principle are to preselect the area of frame and true tag frame to hand over and compare and be greater than friendship and than threshold value IOU, are denoted as just
Otherwise sample is negative sample, wherein set friendship and than threshold value IOU as 0.5.
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Suddenly are as follows:
(a) setting classification number is 2, the t raining period of training set is 10 times, training batch size is 16, basic learning rate ginseng
Number is 0.0003, learning rate decay factor is 0.9.
(b) pre-training is carried out to improved SSD dual network using VOC2007 data set, by the weight after pre-training and partially
Set the initial weight as improved SSD dual network and biasing.
(c) the pretreated 700 Zhang great little training sample image for being 300 × 300 is input to the dynamic after pre-training
In threshold value SSD network, sets and intersect entropy function as loss function, reduce penalty values using gradient descent method to carry out propagated forward
And backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 10 times obtains
Trained dynamic threshold SSD network model.
(d) by the pretreated 700 Zhang great little training sample image for being 300 × 300 be input to after pre-training on adopt
In sample SSD network, set intersect entropy function as loss function, using gradient descent method reduce penalty values come carry out propagated forward with
Backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 10 times is instructed
The up-sampling SSD network model perfected.
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.5.
(b) pretreated 180 test sample images are input to trained dynamic threshold SSD network and upper simultaneously
It is tested in sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each target
Position coordinates, classification and generic confidence level.
It (c) is more than registration threshold value 0.4 with position registration of the non-maxima suppression algorithm to target in 2 groups of testing results
Testing result carry out duplicate removal, retain the higher position coordinates of confidence level and its classification, obtain final detection result, choose average
Precision mAP value evaluates testing result as evaluation index, and mAP value reaches 85.11%, every image averaging inspection
Surveying the time is 0.13 second.For convenience of observation testing result, the present embodiment chooses a Zhang great little from 180 test sample images and is
704 × 576, the image (such as Fig. 4) that number is 312 is visualized, and visualization result is as shown in Figure 5.As seen from Figure 5, in figure
The position of examinee can be detected, and the confidence value of each examinee's classification is higher, and effect is preferable.
Embodiment 2
The image of the present embodiment carrys out examination hall monitor video under the standardization examination hall environment that freely modes such as online crawler obtain
Examination hall examinee's data set for constituting of single-frame images, the present embodiment uses 700 images in the data set as training set,
180 images are as test set, and training set is not be overlapped with test set.
The examination hall examinee position rapid detection method of the improvement SSD dual network of the present embodiment is made of following step:
(1) image preprocessing
It is concentrated from image data and chooses 700 training sample images and 180 test sample images, use bilinear interpolation
The image of selection is normalized to 250 × 250 by pixel size.
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
The SSD network structure of the present embodiment is by input layer, convolution 1, convolution 2, convolution 3, convolution 4, convolution 5, full convolutional layer
6, full convolutional layer 7, convolutional layer 8, convolutional layer 9, convolutional layer 10, convolutional layer 11 connect and compose.
Input layer is the training sample image that 700 Zhang great little are 250 × 250, and the convolution 1 of VGG16, volume are connected after input layer
Convolution 2 is connected after product 1, convolution 3 is connected after convolution 2, and convolution 4 is connected after convolution 3, convolution 5 is connected after convolution 4, is connected after convolution 5
Full convolutional layer 6 connects full convolutional layer 7 after full convolutional layer 6, convolution 8 is connected after full convolutional layer 7, connects convolution 9, convolution after convolution 8
Convolution 10 is connected after 9, connects convolution 11 after convolution 10, after convolution 4, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11
It is connect with output layer.
(b) building pre-selection frame matching process
The step is same as Example 1.
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is the training sample image that 700 Zhang great little are 250 × 250, and the convolution 1 of VGG16, volume are connected after input layer
Head and the tail sequentially connect convolution 2, convolution 3 after product 1, and head and the tail sequentially connect convolutional layer 4-1, the convolutional layer 4-2 of convolution 4, volume after convolution 3
The up-sampling layer 4-4 that size is 76 × 76 is connected after lamination 4-2, connection convolution kernel number is 512, convolution after up-sampling layer 4-4
The prediction interval 4-5 that core size is 3 × 3 connects convolutional layer 4-3 after convolutional layer 4-2, connects convolution 5 after convolutional layer 4-3, after convolution 5
Full convolutional layer 6, full convolutional layer 7 is linked in sequence in head and the tail, and head and the tail sequentially connect 8~convolution of convolution 11, convolutional layer after full convolutional layer 7
It is connect with output layer after 4-5, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11, constitutes up-sampling SSD network knot
Structure;
(b) the pre-selection frame matching process of building up-sampling SSD network
The step is same as Example 1.
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Suddenly are as follows:
(a) setting classification number is 2, the t raining period of training set is 8 times, training batch size is 15, basic learning rate ginseng
Number is 0.0002, learning rate decay factor is 0.7.
(b) pre-training is carried out to improved SSD dual network using VOC2007 data set, by the weight after pre-training and partially
Set the initial weight as improved SSD dual network and biasing.
(c) the pretreated 700 Zhang great little training sample image for being 250 × 250 is input to the dynamic after pre-training
In threshold value SSD network, sets and intersect entropy function as loss function, reduce penalty values using gradient descent method to carry out propagated forward
And backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 8 times is instructed
The dynamic threshold SSD network model perfected.
(d) by the pretreated 700 Zhang great little training sample image for being 250 × 250 be input to after pre-training on adopt
In sample SSD network, set intersect entropy function as loss function, using gradient descent method reduce penalty values come carry out propagated forward with
Backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 8 times is trained
Good up-sampling SSD network model.
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.45.
(b) pretreated 180 test sample images are input to trained dynamic threshold SSD network and upper simultaneously
It is tested in sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each target
Position coordinates, classification and generic confidence level.
It (c) is more than registration threshold value with position registration of the non-maxima suppression algorithm to target in 2 groups of testing results
0.35 testing result carries out duplicate removal, retains the higher position coordinates of confidence level and its classification, obtains final detection result, chooses
Average precision mAP value evaluates testing result as evaluation index, and mAP value reaches 85.11%, and every image is flat
Equal detection time is 0.13 second.
Embodiment 3
The image of the present embodiment carrys out examination hall monitor video under the standardization examination hall environment that freely modes such as online crawler obtain
Examination hall examinee's data set for constituting of single-frame images, the present embodiment uses 700 images in the data set as training set,
180 images are as test set, and training set is not be overlapped with test set.
The examination hall examinee position rapid detection method of the improvement SSD dual network of the present embodiment is made of following step:
(1) image preprocessing
It is concentrated from image data and chooses 700 training sample images and 180 test sample images, use bilinear interpolation
The image of selection is normalized to 500 × 500 by pixel size.
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
The SSD network structure of the present embodiment is by input layer, convolution 1, convolution 2, convolution 3, convolution 4, convolution 5, full convolutional layer
6, full convolutional layer 7, convolutional layer 8, convolutional layer 9, convolutional layer 10, convolutional layer 11 connect and compose.
Input layer is the training sample image that 700 Zhang great little are 500 × 500, and the convolution 1 of VGG16, volume are connected after input layer
Convolution 2 is connected after product 1, convolution 3 is connected after convolution 2, and convolution 4 is connected after convolution 3, convolution 5 is connected after convolution 4, is connected after convolution 5
Full convolutional layer 6 connects full convolutional layer 7 after full convolutional layer 6, convolution 8 is connected after full convolutional layer 7, connects convolution 9, convolution after convolution 8
Convolution 10 is connected after 9, connects convolution 11 after convolution 10, after convolution 4, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11
It is connect with output layer.
(b) building pre-selection frame matching process
The step is same as Example 1.
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is the training sample image that 700 Zhang great little are 500 × 500, and the convolution 1 of VGG16, volume are connected after input layer
Head and the tail sequentially connect convolution 2, convolution 3 after product 1, and head and the tail sequentially connect convolutional layer 4-1, the convolutional layer 4-2 of convolution 4, volume after convolution 3
The up-sampling layer 4-4 that size is 76 × 76 is connected after lamination 4-2, connection convolution kernel number is 512, convolution after up-sampling layer 4-4
The prediction interval 4-5 that core size is 3 × 3 connects convolutional layer 4-3 after convolutional layer 4-2, connects convolution 5 after convolutional layer 4-3, after convolution 5
Full convolutional layer 6, full convolutional layer 7 is linked in sequence in head and the tail, and head and the tail sequentially connect 8~convolution of convolution 11, convolutional layer after full convolutional layer 7
It is connect with output layer after 4-5, full convolutional layer 7, convolution 8, convolution 9, convolution 10 and convolution 11, constitutes up-sampling SSD network knot
Structure;
(b) the pre-selection frame matching process of building up-sampling SSD network
The step is same as Example 1.
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Suddenly are as follows:
(a) setting classification number is 2, the t raining period of training set is 20 times, training batch size is 32, basic learning rate ginseng
Number is 0.001, learning rate decay factor is 0.95.
(b) pre-training is carried out to improved SSD dual network using VOC2007 data set, by the weight after pre-training and partially
Set the initial weight as improved SSD dual network and biasing.
(c) the pretreated 700 Zhang great little training sample image for being 500 × 500 is input to the dynamic after pre-training
In threshold value SSD network, sets and intersect entropy function as loss function, reduce penalty values using gradient descent method to carry out propagated forward
And backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 20 times obtains
Trained dynamic threshold SSD network model.
(d) by the pretreated 700 Zhang great little training sample image for being 500 × 500 be input to after pre-training on adopt
In sample SSD network, set intersect entropy function as loss function, using gradient descent method reduce penalty values come carry out propagated forward with
Backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 20 times is instructed
The up-sampling SSD network model perfected.
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.65.
(b) pretreated 180 test sample images are input to trained dynamic threshold SSD network and upper simultaneously
It is tested in sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each target
Position coordinates, classification and generic confidence level.
It (c) is more than registration threshold value 0.6 with position registration of the non-maxima suppression algorithm to target in 2 groups of testing results
Testing result carry out duplicate removal, retain the higher position coordinates of confidence level and its classification, obtain final detection result, choose average
Precision mAP value evaluates testing result as evaluation index, and mAP value reaches 85.11%, every image averaging inspection
Surveying the time is 0.13 second.
Embodiment 4
In above Examples 1 to 3, the present embodiment uses 600 images in the data set as training set, and 280
Image is opened as test set, and training set is not be overlapped with test set.
The examination hall examinee position rapid detection method of the improvement SSD dual network of the present embodiment is made of following step:
(1) image preprocessing
From other in image data concentration 600 training sample images of selection and 280 test sample images, the step
Step is identical as corresponding embodiment.
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
The SSD network structure of the present embodiment is by input layer, convolution 1, convolution 2, convolution 3, convolution 4, convolution 5, full convolutional layer
6, full convolutional layer 7, convolutional layer 8, convolutional layer 9, convolutional layer 10, convolutional layer 11 connect and compose.
Input layer is 600 training sample images, and the pixel of training sample image is identical as corresponding embodiment.The step
Other steps it is same as Example 1.
(b) building pre-selection frame matching process
The step is same as Example 1.
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is 600 training sample images, and the pixel of training sample image is identical as corresponding embodiment.The step
Other steps it is same as Example 1.
(b) the pre-selection frame matching process of building up-sampling SSD network
The step is same as Example 1.
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Suddenly are as follows:
Step (a) is identical as corresponding embodiment as step (b).
(c) by pretreated 600 training sample images, in the dynamic threshold SSD network after being input to pre-training, instruction
The pixel for practicing sample image is identical as corresponding embodiment.Other steps are identical as corresponding embodiment.
(d) by pretreated 600 training sample images, in the up-sampling SSD network after being input to pre-training, training
The pixel of sample image is identical as corresponding embodiment.Other steps are identical as corresponding embodiment.
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.5.
(b) pretreated 280 test sample images are input to trained dynamic threshold SSD network and upper simultaneously
It is tested in sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each target
Position coordinates, classification and generic confidence level.
It (c) is more than registration threshold value 0.4 with position registration of the non-maxima suppression algorithm to target in 2 groups of testing results
Testing result carry out duplicate removal, retain the higher position coordinates of confidence level and its classification, obtain final detection result, choose average
Precision mAP value evaluates testing result as evaluation index, and mAP value reaches 82.01%, every image averaging inspection
Surveying the time is 0.131 second.For convenience of observation testing result, the present embodiment chooses a Zhang great little from 280 test sample images
The image (such as Fig. 6) for being 814 for 704 × 576, number is visualized, visualization result such as Fig. 7.As seen from Figure 7, it is examined in figure
Raw position can be detected, and the confidence value of each examinee's classification is higher, and effect is preferable.
Embodiment 5
In above Examples 1 to 3, the present embodiment uses 800 images in the data set as training set, and 80
Image is as test set, and training set is not be overlapped with test set.
The examination hall examinee position rapid detection method of the improvement SSD dual network of the present embodiment is made of following step:
(1) image preprocessing
From other in image data concentration 800 training sample images of selection and 80 test sample images, the step
Step is identical as corresponding embodiment.
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
The SSD network structure of the present embodiment is by input layer, convolution 1, convolution 2, convolution 3, convolution 4, convolution 5, full convolutional layer
6, full convolutional layer 7, convolutional layer 8, convolutional layer 9, convolutional layer 10, convolutional layer 11 connect and compose.
Input layer is 800 training sample images, and the pixel of the image of training sample is identical as corresponding embodiment.The step
Other rapid steps are same as Example 1.
(b) building pre-selection frame matching process
The step is same as Example 1.
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is 800 training sample images, and the pixel of training sample image is identical as corresponding embodiment.The step
Other steps it is same as Example 1.
(b) the pre-selection frame matching process of building up-sampling SSD network
The step is same as Example 1.
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Suddenly are as follows:
Step (a) is identical as corresponding embodiment as step (b).
(c) by pretreated 800 training sample images, in the dynamic threshold SSD network after being input to pre-training, instruction
The pixel for practicing sample image is identical as corresponding embodiment.Other steps are identical as corresponding embodiment.
(d) by pretreated 800 training sample images, in the up-sampling SSD network after being input to pre-training, training
The pixel of sample image is identical as corresponding embodiment.Other steps are identical as corresponding embodiment.
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.5.
(b) pretreated 80 test sample images are input to trained dynamic threshold SSD network and upper simultaneously
It is tested in sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each target
Position coordinates, classification and generic confidence level.
It (c) is more than registration threshold value 0.4 with position registration of the non-maxima suppression algorithm to target in 2 groups of testing results
Testing result carry out duplicate removal, retain the higher position coordinates of confidence level and its classification, obtain final detection result, choose average
Precision mAP value evaluates testing result as evaluation index, and mAP value reaches 80.21%, every image averaging inspection
Surveying the time is 0.129 second.For convenience of observation testing result, the present embodiment chooses a Zhang great little from 80 test sample images and is
704 × 576, the image (such as Fig. 8) that number is 475 is visualized, visualization result such as Fig. 9.As seen from Figure 9, light compared with
Under conditions of dark, the position of examinee can preferably be detected in figure, missing inspection do not occur, the classification confidence value of examinee compared with
It is high.
Claims (5)
1. a kind of examination hall examinee position rapid detection method of improved SSD dual network, it is characterised in that be made of following step:
(1) image preprocessing
It is concentrated from image data and chooses 600~800 training sample images and 80~280 test sample images, use bilinearity
The image of selection is normalized to 250 × 250~500 × 500 by pixel size by interpolation method;
(2) dynamic threshold SSD network is constructed
(a) using SSD network structure as the initial configuration of dynamic threshold SSD network
Input layer is the training sample image that 600~800 Zhang great little are 250 × 250~500 × 500, is connected after input layer
The convolution (1) of VGG16, head and the tail sequentially connect convolution (2)~convolution (5) to convolution (1) afterwards, and sequentially connection is complete for head and the tail afterwards for convolution (5)
Convolutional layer (6), full convolutional layer (7), head and the tail sequentially connect convolution (8)~convolution (11) to full convolutional layer (7) afterwards, convolution (4), full volume
Lamination (7), convolution (8), convolution (9), convolution (10) and convolution (11) are connect with output layer afterwards;
(b) building pre-selection frame matching process
Convolutional layer (the 8- in convolutional layer (4-3), full convolutional layer (7), convolution (8) in initial configuration, in selection convolution (4)
2), the convolutional layer (9-2) in convolution (9), the convolutional layer (10-2) in convolution (10), the convolutional layer (11) in convolution (11) totally 6
A convolutional layer takes the pre-selection frame of different scale to be matched with true tag frame on 6 convolutional layers, divides the pre-selection positive and negative sample of frame
This collection, matching process are that the area of the pre-selection frame and true tag frame in SSD network is handed over and compared greater than friendship and than threshold value IOU, note
It is otherwise negative sample for positive sample;
Determine friendship and the method than threshold value IOU are as follows: determine according to the area of true tag frame and the ratio of input image pixels size
The threshold value IOU of fixed friendship and ratio, dynamic adjust threshold value IOU and divide positive negative sample, and true tag frame area and input image pixels are big
Small ratio≤0.01 hands over and is 0.4 than threshold value IOU, otherwise, hands over and is 0.5 than threshold value IOU, determine true tag frame as the following formula
Area SGTbWith friendship and than threshold value IOU:
SGTb=(ymax-ymin)×(xmax-xmin)
Wherein xmax、xmin、ymax、yminThe respectively maxima and minima of the abscissa of true tag frame and ordinate;Sinput
For the area of input picture, trained network is constituted by dynamic threshold SSD network according to the matching process;
(3) building up-sampling SSD network
(a) building up-sampling SSD network structure
Input layer is the training sample image that 600~800 Zhang great little are 250 × 250~500 × 500, is connected after input layer
The convolution (1) of VGG16, head and the tail sequentially connect convolution (2), convolution (3) to convolution (1) afterwards, and head and the tail sequentially connect convolution to convolution (3) afterwards
(4) convolutional layer (4-1), convolutional layer (4-2), convolutional layer (4-2) connect the up-sampling layer (4-4) that size is 76 × 76 afterwards, on
Sample level (4-4) connects the prediction interval (4-5) that convolution kernel number is 512, convolution kernel size is 3 × 3 afterwards, after convolutional layer (4-2)
Connect convolutional layer (4-3), convolutional layer (4-3) connects convolution (5) afterwards, convolution (5) be linked in sequence from beginning to end afterwards full convolutional layer (6),
Full convolutional layer (7), head and the tail sequentially connect convolution (8)~convolution (11) to full convolutional layer (7) afterwards, convolutional layer (4-5), full convolutional layer
(7), convolution (8), convolution (9), convolution (10) and convolution (11) are connect with output layer afterwards, constitute up-sampling SSD network structure;
(b) the pre-selection frame matching process of building up-sampling SSD network
In up-sampling SSD network structure, in the prediction interval (4-5), full convolutional layer (7), convolution (8) in selection convolution (4)
The convolutional layer (10-2) in convolutional layer (9-2), convolution (10) in convolutional layer (8-2), convolution (9), the convolution in convolution (11)
Layer (11) totally 6 convolutional layers, take the pre-selection frame of different scale to be matched on 6 convolutional layers with true tag frame, will preselect
Frame is divided into positive sample collection and negative sample collection, and specific matching principle is to preselect frame and the area of true tag frame is handed over and compare greater than friendship
And than threshold value IOU, it is denoted as positive sample, is otherwise negative sample, wherein sets friendship and than threshold value IOU as 0.5;
(4) the improved SSD dual network of training
Improved SSD dual network is made of dynamic threshold SSD network and up-sampling SSD network, simultaneously training, trained step
Are as follows:
(a) setting classification number is 2, the t raining period of training set is 8~20 times, training batch size is 15~32, basic learning
Rate parameter is 0.0002~0.001, learning rate decay factor is 0.7~0.95;
(b) pre-training is carried out to improved SSD dual network using VOC2007 data set, by the weight and biasing work after pre-training
Initial weight and biasing for improved SSD dual network;
(c) pretreated 600~800 Zhang great little training sample image for being 250 × 250~500 × 500 is input to pre- instruction
In dynamic threshold SSD network after white silk, set intersect entropy function as loss function, using gradient descent method reduction penalty values come into
Row propagated forward and backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training
8~20 times, obtain trained dynamic threshold SSD network model;
(d) pretreated 600~800 Zhang great little training sample image for being 250 × 250~500 × 500 is input to pre- instruction
In up-sampling SSD network after white silk, sets and intersect entropy function as loss function, reduce penalty values using gradient descent method to carry out
Propagated forward and backpropagation, iterative cycles propagated forward and backpropagation, and weighted value and the biasing of network are updated, training 8
~20 times, obtain trained up-sampling SSD network model;
(5) test sample image is tested
(a) the classification confidence threshold value that target is arranged is 0.45~0.65;
(b) pretreated 80~280 test sample images are input to trained dynamic threshold SSD network and upper simultaneously
It is tested in sampling SSD network, obtains 2 groups of different image object testing results, wherein every group of result includes each target
Position coordinates, classification and generic confidence level;
(c) with non-maxima suppression algorithm to the position registration of target in 2 groups of testing results be more than registration threshold value 0.35~
0.6 testing result carries out duplicate removal, retains the higher position coordinates of confidence level and its classification, obtains final detection result, chooses
Average precision mAP value evaluates testing result as evaluation index.
2. the examination hall examinee position rapid detection method according to claim 1 for improving SSD dual network, it is characterised in that
In the step of constructing dynamic threshold SSD network step (2) (b), the width ω of the pre-selection frame of the different scalekWith height hk
Are as follows:
Wherein, SkFor size basis value, SminIt is 0.1, SmaxIt is 0.9, respectively indicates smallest dimension and out to out, m is prediction
Convolutional layer number, be 6, k be (1, m] integer, a be pre-selection different the ratio of width to height of frame, a ∈ { 1/3,1/2,1,2,3 }.
3. the examination hall examinee position rapid detection method according to claim 1 for improving SSD dual network, it is characterised in that:
In the step of training improved SSD dual network step (4) (a), the t raining period of training set is 10 times, training batch ruler
Very little is 16.
4. the examination hall examinee position rapid detection method according to claim 1 for improving SSD dual network, it is characterised in that:
In described the step of carrying out testing procedure (5) to test sample image (a), the classification confidence threshold value of target is 0.5.
5. a kind of examination hall examinee position rapid detection method for improving SSD dual network according to claim 1, feature exist
In: in the step of of carrying out testing procedure (5) to test sample image (c), the registration of position coordinates in testing result
Threshold value is 0.4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534569.9A CN110378232B (en) | 2019-06-20 | 2019-06-20 | Improved test room examinee position rapid detection method of SSD dual-network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534569.9A CN110378232B (en) | 2019-06-20 | 2019-06-20 | Improved test room examinee position rapid detection method of SSD dual-network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110378232A true CN110378232A (en) | 2019-10-25 |
CN110378232B CN110378232B (en) | 2022-12-27 |
Family
ID=68250365
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910534569.9A Expired - Fee Related CN110378232B (en) | 2019-06-20 | 2019-06-20 | Improved test room examinee position rapid detection method of SSD dual-network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110378232B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110827432A (en) * | 2019-11-11 | 2020-02-21 | 深圳算子科技有限公司 | Class attendance checking method and system based on face recognition |
CN110852222A (en) * | 2019-10-31 | 2020-02-28 | 上海交通大学 | Campus corridor scene intelligent monitoring method based on target detection |
CN111144417A (en) * | 2019-12-27 | 2020-05-12 | 创新奇智(重庆)科技有限公司 | Intelligent container small target detection method and detection system based on teacher student network |
CN111259844A (en) * | 2020-01-21 | 2020-06-09 | 陕西师范大学 | Real-time monitoring method for examinees in standardized examination room |
CN112132140A (en) * | 2020-09-23 | 2020-12-25 | 平安国际智慧城市科技股份有限公司 | Vehicle brand identification method, device, equipment and medium based on artificial intelligence |
CN112418263A (en) * | 2020-10-10 | 2021-02-26 | 上海鹰瞳医疗科技有限公司 | Medical image focus segmentation and labeling method and system |
CN113221768A (en) * | 2021-05-18 | 2021-08-06 | 北京百度网讯科技有限公司 | Recognition model training method, recognition method, device, equipment and storage medium |
CN113408342A (en) * | 2021-05-11 | 2021-09-17 | 深圳大学 | Target detection method for determining intersection ratio threshold based on features |
CN114120127A (en) * | 2021-11-30 | 2022-03-01 | 济南博观智能科技有限公司 | Target detection method, device and related equipment |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190065901A1 (en) * | 2017-08-29 | 2019-02-28 | Vintra, Inc. | Systems and methods for a tailored neural network detector |
CN109711377A (en) * | 2018-12-30 | 2019-05-03 | 陕西师范大学 | Standardize examinee's positioning and method of counting in the single-frame images of examination hall monitoring |
CN109784476A (en) * | 2019-01-12 | 2019-05-21 | 福州大学 | A method of improving DSOD network |
-
2019
- 2019-06-20 CN CN201910534569.9A patent/CN110378232B/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190065901A1 (en) * | 2017-08-29 | 2019-02-28 | Vintra, Inc. | Systems and methods for a tailored neural network detector |
CN109711377A (en) * | 2018-12-30 | 2019-05-03 | 陕西师范大学 | Standardize examinee's positioning and method of counting in the single-frame images of examination hall monitoring |
CN109784476A (en) * | 2019-01-12 | 2019-05-21 | 福州大学 | A method of improving DSOD network |
Non-Patent Citations (1)
Title |
---|
周晓彦等: "基于深度学习的目标检测算法综述", 《电子测量技术》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110852222A (en) * | 2019-10-31 | 2020-02-28 | 上海交通大学 | Campus corridor scene intelligent monitoring method based on target detection |
CN110827432A (en) * | 2019-11-11 | 2020-02-21 | 深圳算子科技有限公司 | Class attendance checking method and system based on face recognition |
CN111144417A (en) * | 2019-12-27 | 2020-05-12 | 创新奇智(重庆)科技有限公司 | Intelligent container small target detection method and detection system based on teacher student network |
CN111259844A (en) * | 2020-01-21 | 2020-06-09 | 陕西师范大学 | Real-time monitoring method for examinees in standardized examination room |
CN111259844B (en) * | 2020-01-21 | 2023-02-24 | 陕西师范大学 | Real-time monitoring method for examinees in standardized examination room |
CN112132140A (en) * | 2020-09-23 | 2020-12-25 | 平安国际智慧城市科技股份有限公司 | Vehicle brand identification method, device, equipment and medium based on artificial intelligence |
CN112132140B (en) * | 2020-09-23 | 2022-08-12 | 平安国际智慧城市科技股份有限公司 | Vehicle brand identification method, device, equipment and medium based on artificial intelligence |
CN112418263A (en) * | 2020-10-10 | 2021-02-26 | 上海鹰瞳医疗科技有限公司 | Medical image focus segmentation and labeling method and system |
CN113408342A (en) * | 2021-05-11 | 2021-09-17 | 深圳大学 | Target detection method for determining intersection ratio threshold based on features |
CN113221768A (en) * | 2021-05-18 | 2021-08-06 | 北京百度网讯科技有限公司 | Recognition model training method, recognition method, device, equipment and storage medium |
CN114120127A (en) * | 2021-11-30 | 2022-03-01 | 济南博观智能科技有限公司 | Target detection method, device and related equipment |
CN114120127B (en) * | 2021-11-30 | 2024-06-07 | 济南博观智能科技有限公司 | Target detection method, device and related equipment |
Also Published As
Publication number | Publication date |
---|---|
CN110378232B (en) | 2022-12-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110378232A (en) | The examination hall examinee position rapid detection method of improved SSD dual network | |
CN107644415B (en) | A kind of text image method for evaluating quality and equipment | |
CN106446930A (en) | Deep convolutional neural network-based robot working scene identification method | |
CN104851099B (en) | A kind of image interfusion method based on expression study | |
CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
CN108875794A (en) | Image visibility detection method based on transfer learning | |
CN108334847A (en) | A kind of face identification method based on deep learning under real scene | |
CN108830285A (en) | A kind of object detection method of the reinforcement study based on Faster-RCNN | |
CN108596046A (en) | A kind of cell detection method of counting and system based on deep learning | |
CN104463249B (en) | A kind of remote sensing images airfield detection method based on Weakly supervised learning framework | |
CN106845357A (en) | A kind of video human face detection and recognition methods based on multichannel network | |
CN113762133A (en) | Self-weight fitness auxiliary coaching system, method and terminal based on human body posture recognition | |
CN106709958A (en) | Gray scale gradient and color histogram-based image quality evaluation method | |
CN116863274B (en) | Semi-supervised learning-based steel plate surface defect detection method and system | |
CN106779054B (en) | A kind of PM2.5 methods of estimation based on Misty Image | |
CN113435282B (en) | Unmanned aerial vehicle image ear recognition method based on deep learning | |
CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
US20220315243A1 (en) | Method for identification and recognition of aircraft take-off and landing runway based on pspnet network | |
CN109711377A (en) | Standardize examinee's positioning and method of counting in the single-frame images of examination hall monitoring | |
CN108764134A (en) | A kind of automatic positioning of polymorphic type instrument and recognition methods suitable for crusing robot | |
CN109784278A (en) | The small and weak moving ship real-time detection method in sea based on deep learning | |
CN110133049A (en) | Tea grades fast non-destructive detection method based on electronic nose and machine vision | |
CN108257154A (en) | Polarimetric SAR Image change detecting method based on area information and CNN | |
CN109635634A (en) | A kind of pedestrian based on stochastic linear interpolation identifies data enhancement methods again | |
CN109657600A (en) | A kind of video area removes altering detecting method and device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20221227 |