CN109766887A - A kind of multi-target detection method based on cascade hourglass neural network - Google Patents
A kind of multi-target detection method based on cascade hourglass neural network Download PDFInfo
- Publication number
- CN109766887A CN109766887A CN201910038893.1A CN201910038893A CN109766887A CN 109766887 A CN109766887 A CN 109766887A CN 201910038893 A CN201910038893 A CN 201910038893A CN 109766887 A CN109766887 A CN 109766887A
- Authority
- CN
- China
- Prior art keywords
- confidence level
- network
- hourglass
- target
- cascade
- 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
Landscapes
- Image Analysis (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The present invention provides a kind of multi-target detection methods based on cascade hourglass neural network, it is intended to the technical issues of it is excessively slow to solve existing detection method speed, and is difficult to for Small object.The present invention is the following steps are included: step 1, acquisition training sample；Step 2 builds deep learning frame and constructs the backbone network cascade hourglass network of target detection；Step 3, project training sample label be confidence level thermal map；The loss function that step 4, design cascade hourglass network optimizes it；Step 5, training cascade hourglass network obtain detection model；Step 6, multi-target detection.The beneficial technical effect of the present invention lies in: it can quickly and accurately identify the target of plurality of classes, improve the recognition capability to Small object.
Description
Technical field
The present invention relates to multi-target detection technical fields, and in particular to a kind of multiple target based on cascade hourglass neural network
Detection method.
Background technique
Multi-target detection is an important directions in computer vision field, and main task is that sense is oriented from image
The target of interest and the specific category for judging each target.It is driven automatically in medical target detection, intelligent video monitoring, vehicle
It sails, pedestrian detection, vehicle flowrate etc. are widely applied.Traditional method is using histograms of oriented gradients, part two
The feature that the methods of value pattern feature extracts image uses support vector machines, random forest and neural network to classify again.But
Being it, there are speed is excessively slow and not high two disadvantages of precision.
Convolutional neural networks obtain extensive success in the picture in recent years.Girshick proposes RCNN and Fast-
RCNN greatly improves the detection accuracy of traditional detection method.Ren et al. has also been proposed Faster R-CNN and further improves
Detect speed.But speed is still undesirable；Joseph et al. proposes YOLO and YOLO9000 greatly and improves speed can be real
When handle picture to be detected, but for Small object, this method be easy to cause missing inspection and erroneous detection.
Summary of the invention
The technical problem to be solved in the present invention is to provide it is a kind of based on cascade hourglass neural network multi-target detection method,
Effect is poor and slow when solving the problems, such as that existing detection model is applied to identification Small object.
In order to solve the above technical problems, the present invention adopts the following technical scheme: a kind of based on cascade hourglass neural network
Multi-target detection method, comprising the following steps:
Step 1, acquisition training sample: by image capture device target image to be detected, target image is marked
And to pre-processing, image is made to meet call format, constructs training sample set；Image preprocessing uses image enhancement, including
(- 180 °, 180 °) of random angles rotations and random scaled (0.5 times -2 times).
Step 2 builds deep learning frame and constructs the backbone network cascade hourglass network of target detection: used grade
Connection hourglass neural network is formed by 4 hourglass cascades, and each hourglass network includes 4 up-sampling layers and 4 down-sampling layers
Constitute, have including 12 convolutional layers, 12 ReLU layer, 12 BatchNorm2d layers with 4 MaxPool2d layers.
Step 3, project training sample label be confidence level thermal map: the training sample label that step 1 is acquired and marked
It generates confidence level thermal map and is used for network training, the target of each type is in same layer confidence level thermal map.Confidence level generated
Thermal map generates M × 64 × 64 sizes confidence level thermal map according to the species number M of required detection, and the target of each type only exists
Affiliated layer maps a Gauss confidence level.
The loss function that step 4, design cascade hourglass network optimizes it: losing letter using least mean-square error
Number, carries out study optimization to network with Adam optimizer.
Step 5, training cascade hourglass network obtain detection model: the collected training sample of step 1 is raw by step 3
Detection model is obtained by cascading hourglass network training at after confidence level thermal map.
Step 6, multi-target detection: target detection, the confidence level heat of the different layers of output are carried out using cascade hourglass network
Figure represents different types, and confidence level position mapping in each layer represents the position of target to be detected.To confidence level heat
Figure carries out non-maxima suppression, and at its frame, will map back picture to be detected on confidence level thermal map with minimum rectangle is to detect
As a result.
The invention has the following advantages over the prior art:
(1) present invention does backbone network by cascade hourglass neural network, is detected by way of confidence level thermal map, is
A kind of relatively new detection method.
(2) present invention uses more cascade structures, can identify smaller target, improves detection accuracy.
(3) the invention belongs to the methods of one-stage, and the speed of service is faster.
Detailed description of the invention
Fig. 1 is acquired example images, wherein Fig. 1 (a) is multiple unmanned plane images that scene once acquires, Fig. 1 (b)
For the lower multiple unmanned plane images acquired of scene two, Fig. 1 (c) is the lower multiple unmanned plane images acquired of scene three, and Fig. 1 (d) is
The lower multiple unmanned plane images acquired of scene four；
Fig. 2 is cascade hourglass neural network schematic diagram；
Fig. 3 is that data set label switchs to Gauss confidence level thermal map schematic diagram, wherein Fig. 3 (a) is to be generated by data set label
First Gauss confidence level thermal map, Fig. 3 (b) is the second Gauss confidence level thermal map generated by data set label, Fig. 3 (c)
For the third Zhang Gaosi confidence level thermal map generated by data set label, Fig. 3 (d) is the 4th Zhang Gaosi generated by data set label
Confidence level thermal map, Fig. 3 (e) are the 5th Gauss confidence level thermal map generated by data set label, and Fig. 3 (f) is by data set mark
The 6th Gauss confidence level thermal map generated is signed, Fig. 3 (g) is the correspondence image data of data set；
Fig. 4 is overall flow figure of the present invention；
Fig. 5 is general frame figure of the present invention.
Specific embodiment
Specific embodiments of the present invention will be described in detail with reference to the accompanying drawing.But following embodiment is used only in detail
Illustrate the present invention, does not limit the scope of the invention in any way.
As shown in figure 4, a kind of multi-target detection method based on cascade hourglass neural network of the present invention, including walk as follows
It is rapid:
Step 1, as shown in Figure 1, acquisition training sample: by image capture device target image to be detected, by target figure
As being marked and to pre-processing, image is made to meet call format, constructs training sample set；Image preprocessing uses image
Enhancing, including (- 180 °, 180 °) of random angles rotations and random scaled (0.5 times -2 times).
Step 2 builds deep learning frame and constructs the backbone network cascade hourglass network of target detection: specific such as Fig. 2
Its shown network, used cascade hourglass neural network are formed by 4 hourglass cascades, and each hourglass network includes 4
Up-sample layer and 4 down-sampling layers constituted, have including 12 convolutional layers, 12 ReLU layer, 12 BatchNorm2d layers with 4
MaxPool2d layers, each network hourglass module is connected using residual error；
Step 3, project training sample label be confidence level thermal map: as shown in figure 3, the instruction that step 1 is acquired and marked
Practice sample label and generate confidence level thermal map for network training, the target of each type is in same layer confidence level thermal map.It gives birth to
At confidence level thermal map according to the species number M of required detection, generate M × 64 × 64 sizes confidence level thermal map, each type
Target only map a Gauss confidence level in affiliated layer.
Gaussion=(x-centerx) ^2/boxx+ (y-centery) ^2/boxy
Wherein, x, y are the coordinate of confidence map respectively, and centerx, centery are the center x of sample label respectively, and y is sat
Mark, boxx, boxy are the length and width of label respectively.
The loss function that step 4, design cascade hourglass network optimizes it: losing letter using least mean-square error
Number, carries out study optimization to network with Adam optimizer.Specifically, network is formed using 4 hourglass cascades, and one shares 4
A output is y_pred1-y_pred4 respectively, loss function be four outputs and；
Loss1=mse (y_pred1-label)；
Loss2=mse (y_pred2-label)；
Loss3=mse (y_pred3-label)；
Loss4=mse (y_pred4-label)；
Loss=loss1+loss2+loss3+loss4；
Step 5, training cascade hourglass network obtain detection model: as shown in figure 4, according to this method frame diagram, by step 1
Collected training sample obtains detection model by cascading hourglass network training after step 3 generates confidence level thermal map.
Step 6, multi-target detection: as shown in figure 3, target detection is carried out using cascade hourglass network, it is defeated with the last layer
The confidence level thermal map of different layers out represents different types, and confidence level position mapping in each layer represents mesh to be detected
Target position.Non-maxima suppression is carried out to confidence level thermal map, is mapped back on confidence level thermal map by its frame with minimum rectangle
Picture to be detected is testing result.
In order to verify effectiveness of the invention, selected from Faster-RCNN and YOLO as comparative example, using shown in Fig. 1
Data set come comparison result, the Average Accuracy (mAP) and real-time of comparison algorithm, comparison result such as table one.
| Method | MAP (%) | Real-time (frame/second) |
| Faster-RCNN | 82.4 | 5 |
| YOLO | 81.52 | 53 |
| The present invention | 89.36 | 32 |
Table one: the result of embodiment and comparative example on data set compares.
As shown in Table 1, algorithm improves Average Accuracy and real-time compared to Faster-RCNN, although comparing YOLO reality
When property decreases, but precision is very high.The result reflects the validity of this algorithm.
Program that is involved or relying on is the conventional program or simple program of the art, this field skill in embodiment
Art personnel can make conventional selection or are adaptively adjusted according to concrete application scene.
Claims (5)
1. a kind of multi-target detection method based on cascade hourglass neural network, which comprises the following steps:
Step 1, acquisition training sample: by image capture device target image to be detected, target image is marked and right
It pre-processes, image is made to meet call format, construct training sample set；
Step 2 builds deep learning frame and constructs the backbone network cascade hourglass network of target detection: cascade hourglass nerve net
Network is formed by multiple hourglass cascades, and each hourglass network contains multiple up-sampling layers, down-sampling layer, trans-regional connection
Layer and convolutional layer；
Step 3, project training sample label be confidence level thermal map: training sample label that step 1 is acquired and marked generates
Confidence level thermal map is used for network training, and the target of each type is in same layer confidence level thermal map；
The loss function that step 4, design cascade hourglass network optimizes it: using least mean-square error loss function, using
Adam optimizer carries out study optimization to network；
Step 5, training cascade hourglass network obtain detection model: the collected training sample of step 1 is set by step 3 generation
Detection model is obtained by cascading hourglass network training after reliability thermal map；
Step 6, multi-target detection: target detection, the confidence level thermal map generation of the different layers of output are carried out using cascade hourglass network
The different type of table, in each layer confidence level position mapping represents the position of target to be detected, to confidence level thermal map into
Row non-maxima suppression obtains obtaining mapping in final confidence level thermal map obtaining detection target.
2. a kind of multi-target detection method based on cascade hourglass neural network according to claim 1, which is characterized in that
The image preprocessing of step 1 uses image enhancement, including (- 180 °, 180 °) of random angles rotations and random scaled
(0.5 times -2 times).
3. a kind of multi-target detection method based on cascade hourglass neural network according to claim 1, which is characterized in that
Cascade hourglass neural network used by step 2 is formed by 4 hourglass cascades, and each hourglass network includes 4 up-samplings
Layer and 4 down-sampling layers are constituted, and are had including 12 convolutional layers, 12 ReLU layer, 12 BatchNorm2d layers with 4
MaxPool2d layers.
4. a kind of multi-target detection method based on cascade hourglass neural network according to claim 1, which is characterized in that
Step 3 confidence level thermal map generated generates M × 64 × 64 sizes confidence level thermal map according to the species number M of required detection,
The target of each type only maps a Gauss confidence level in affiliated layer.
5. a kind of multi-target detection method based on cascade hourglass neural network according to claim 1, which is characterized in that
Step 6 predicted to confidence level thermal map need using non-maximizations inhibition handle after, with minimum rectangle in confidence level thermal map
On by its frame, mapping back picture to be detected is testing result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910038893.1A CN109766887B (en) | 2019-01-16 | 2019-01-16 | Multi-target detection method based on cascaded hourglass neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910038893.1A CN109766887B (en) | 2019-01-16 | 2019-01-16 | Multi-target detection method based on cascaded hourglass neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109766887A true CN109766887A (en) | 2019-05-17 |
| CN109766887B CN109766887B (en) | 2022-11-11 |
Family
ID=66452239
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910038893.1A Active CN109766887B (en) | 2019-01-16 | 2019-01-16 | Multi-target detection method based on cascaded hourglass neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109766887B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110503063A (en) * | 2019-08-28 | 2019-11-26 | 东北大学秦皇岛分校 | Fall detection method based on hourglass convolution autocoding neural network |
| CN110826403A (en) * | 2019-09-27 | 2020-02-21 | 深圳云天励飞技术有限公司 | Tracking target determination method and related equipment |
| CN110889858A (en) * | 2019-12-03 | 2020-03-17 | 中国太平洋保险（集团）股份有限公司 | Automobile part segmentation method and device based on point regression |
| CN111062981A (en) * | 2019-12-13 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Image processing method, device and storage medium |
| CN111105590A (en) * | 2019-12-27 | 2020-05-05 | 深圳前海微众银行股份有限公司 | Alarm method and device |
| CN111523480A (en) * | 2020-04-24 | 2020-08-11 | 北京嘀嘀无限科技发展有限公司 | Method and device for detecting face obstruction, electronic equipment and storage medium |
| CN111913435A (en) * | 2020-07-30 | 2020-11-10 | 浙江科技学院 | Single/multi-target key point positioning method based on stacked hourglass network |
| CN112132839A (en) * | 2020-08-27 | 2020-12-25 | 杭州英歌智达科技有限公司 | Multi-scale rapid face segmentation method based on deep convolution cascade network |
| CN112598055A (en) * | 2020-12-21 | 2021-04-02 | 电子科技大学 | Helmet wearing detection method, computer-readable storage medium and electronic device |
| CN112966546A (en) * | 2021-01-04 | 2021-06-15 | 航天时代飞鸿技术有限公司 | Embedded attitude estimation method based on unmanned aerial vehicle scout image |
| CN113011365A (en) * | 2021-03-31 | 2021-06-22 | 中国科学院光电技术研究所 | Target detection method combined with lightweight network |
| WO2021217937A1 (en) * | 2020-04-27 | 2021-11-04 | 平安国际智慧城市科技股份有限公司 | Posture recognition model training method and device, and posture recognition method and device |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090161912A1 (en) * | 2007-12-21 | 2009-06-25 | Raviv Yatom | method for object detection |
| CN102665062A (en) * | 2012-03-16 | 2012-09-12 | 华为技术有限公司 | Method and device for stabilizing target object image in video |
| CN106295678A (en) * | 2016-07-27 | 2017-01-04 | 北京旷视科技有限公司 | Neural metwork training and construction method and device and object detection method and device |
| US20170124409A1 (en) * | 2015-11-04 | 2017-05-04 | Nec Laboratories America, Inc. | Cascaded neural network with scale dependent pooling for object detection |
| CN108416394A (en) * | 2018-03-22 | 2018-08-17 | 河南工业大学 | Multi-target detection model building method based on convolutional neural networks |
| CN108647742A (en) * | 2018-05-19 | 2018-10-12 | 南京理工大学 | Fast target detection method based on lightweight neural network |
| CN108875595A (en) * | 2018-05-29 | 2018-11-23 | 重庆大学 | A kind of Driving Scene object detection method merged based on deep learning and multilayer feature |
| CN108921225A (en) * | 2018-07-10 | 2018-11-30 | 深圳市商汤科技有限公司 | A kind of image processing method and device, computer equipment and storage medium |
| CN109003267A (en) * | 2017-08-09 | 2018-12-14 | 深圳科亚医疗科技有限公司 | From the computer implemented method and system of the automatic detected target object of 3D rendering |
| CN109190537A (en) * | 2018-08-23 | 2019-01-11 | 浙江工商大学 | A kind of more personage's Attitude estimation methods based on mask perceived depth intensified learning |
-
2019
- 2019-01-16 CN CN201910038893.1A patent/CN109766887B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090161912A1 (en) * | 2007-12-21 | 2009-06-25 | Raviv Yatom | method for object detection |
| CN102665062A (en) * | 2012-03-16 | 2012-09-12 | 华为技术有限公司 | Method and device for stabilizing target object image in video |
| US20170124409A1 (en) * | 2015-11-04 | 2017-05-04 | Nec Laboratories America, Inc. | Cascaded neural network with scale dependent pooling for object detection |
| CN106295678A (en) * | 2016-07-27 | 2017-01-04 | 北京旷视科技有限公司 | Neural metwork training and construction method and device and object detection method and device |
| CN109003267A (en) * | 2017-08-09 | 2018-12-14 | 深圳科亚医疗科技有限公司 | From the computer implemented method and system of the automatic detected target object of 3D rendering |
| CN108416394A (en) * | 2018-03-22 | 2018-08-17 | 河南工业大学 | Multi-target detection model building method based on convolutional neural networks |
| CN108647742A (en) * | 2018-05-19 | 2018-10-12 | 南京理工大学 | Fast target detection method based on lightweight neural network |
| CN108875595A (en) * | 2018-05-29 | 2018-11-23 | 重庆大学 | A kind of Driving Scene object detection method merged based on deep learning and multilayer feature |
| CN108921225A (en) * | 2018-07-10 | 2018-11-30 | 深圳市商汤科技有限公司 | A kind of image processing method and device, computer equipment and storage medium |
| CN109190537A (en) * | 2018-08-23 | 2019-01-11 | 浙江工商大学 | A kind of more personage's Attitude estimation methods based on mask perceived depth intensified learning |
Non-Patent Citations (2)
| Title |
|---|
| ANIL ATVAR等: "Classification of human poses and orientations with deep learning", 《2018 26TH SIGNAL PROCESSING AND COMMUNICATIONS APPLICATIONS CONFERENCE (SIU)》 * |
| 伍鹏瑛等: "多层卷积特征的真实场景下行人检测研究", 《智能系统学报》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110503063A (en) * | 2019-08-28 | 2019-11-26 | 东北大学秦皇岛分校 | Fall detection method based on hourglass convolution autocoding neural network |
| CN110503063B (en) * | 2019-08-28 | 2021-12-17 | 东北大学秦皇岛分校 | Falling detection method based on hourglass convolution automatic coding neural network |
| WO2021057309A1 (en) * | 2019-09-27 | 2021-04-01 | 深圳云天励飞技术股份有限公司 | Tracked target determination method and related device |
| CN110826403A (en) * | 2019-09-27 | 2020-02-21 | 深圳云天励飞技术有限公司 | Tracking target determination method and related equipment |
| CN110889858A (en) * | 2019-12-03 | 2020-03-17 | 中国太平洋保险（集团）股份有限公司 | Automobile part segmentation method and device based on point regression |
| CN111062981A (en) * | 2019-12-13 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Image processing method, device and storage medium |
| CN111062981B (en) * | 2019-12-13 | 2023-05-05 | 腾讯科技（深圳）有限公司 | Image processing method, device and storage medium |
| CN111105590A (en) * | 2019-12-27 | 2020-05-05 | 深圳前海微众银行股份有限公司 | Alarm method and device |
| CN111523480A (en) * | 2020-04-24 | 2020-08-11 | 北京嘀嘀无限科技发展有限公司 | Method and device for detecting face obstruction, electronic equipment and storage medium |
| WO2021217937A1 (en) * | 2020-04-27 | 2021-11-04 | 平安国际智慧城市科技股份有限公司 | Posture recognition model training method and device, and posture recognition method and device |
| CN111913435A (en) * | 2020-07-30 | 2020-11-10 | 浙江科技学院 | Single/multi-target key point positioning method based on stacked hourglass network |
| CN112132839A (en) * | 2020-08-27 | 2020-12-25 | 杭州英歌智达科技有限公司 | Multi-scale rapid face segmentation method based on deep convolution cascade network |
| CN112132839B (en) * | 2020-08-27 | 2024-04-30 | 杭州英歌智达科技有限公司 | Multi-scale rapid face segmentation method based on deep convolution cascade network |
| CN112598055A (en) * | 2020-12-21 | 2021-04-02 | 电子科技大学 | Helmet wearing detection method, computer-readable storage medium and electronic device |
| CN112598055B (en) * | 2020-12-21 | 2022-06-17 | 电子科技大学 | Helmet wearing detection method, computer-readable storage medium and electronic device |
| CN112966546A (en) * | 2021-01-04 | 2021-06-15 | 航天时代飞鸿技术有限公司 | Embedded attitude estimation method based on unmanned aerial vehicle scout image |
| CN113011365A (en) * | 2021-03-31 | 2021-06-22 | 中国科学院光电技术研究所 | Target detection method combined with lightweight network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109766887B (en) | 2022-11-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109766887A (en) | A kind of multi-target detection method based on cascade hourglass neural network | |
| CN112380952B (en) | Power equipment infrared image real-time detection and identification method based on artificial intelligence | |
| Song et al. | Kiwifruit detection in field images using Faster R-CNN with VGG16 | |
| CN109285179A (en) | A kind of motion target tracking method based on multi-feature fusion | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN108053419A (en) | Inhibited and the jamproof multiscale target tracking of prospect based on background | |
| CN111611874B (en) | Face mask wearing detection method based on ResNet and Canny | |
| CN107993215A (en) | A kind of weather radar image processing method and system | |
| CN112084869B (en) | Compact quadrilateral representation-based building target detection method | |
| CN112200143A (en) | Road disease detection method based on candidate area network and machine vision | |
| CN109543676B (en) | Character wheel type water meter digital character recognition method based on image processing | |
| CN103839265A (en) | SAR image registration method based on SIFT and normalized mutual information | |
| CN106056619A (en) | Unmanned aerial vehicle vision wire patrol method based on gradient constraint Radon transform | |
| CN104408482A (en) | Detecting method for high-resolution SAR (Synthetic Aperture Radar) image object | |
| CN104050681A (en) | Road vanishing point detection method based on video images | |
| CN116091951A (en) | Method and system for extracting boundary line between farmland and tractor-ploughing path | |
| CN109712071A (en) | Unmanned plane image mosaic and localization method based on track constraint | |
| CN109447036A (en) | A kind of segmentation of image digitization and recognition methods and system | |
| CN112861870B (en) | Pointer instrument image correction method, system and storage medium | |
| CN107563299A (en) | A kind of pedestrian detection method using ReCNN integrating context informations | |
| CN110084743A (en) | Image mosaic and localization method based on more air strips starting track constraint | |
| CN112949510A (en) | Human detection method based on fast R-CNN thermal infrared image | |
| Nair et al. | Flood water depth estimation—A survey | |
| CN113191204B (en) | Multi-scale blocking pedestrian detection method and system | |
| CN114581307A (en) | Multi-image stitching method, system, device and medium for target tracking identification |
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 |