CN112464769A - High-resolution remote sensing image target detection method based on consistent multi-stage detection - Google Patents
High-resolution remote sensing image target detection method based on consistent multi-stage detection Download PDFInfo
- Publication number
- CN112464769A CN112464769A CN202011296355.1A CN202011296355A CN112464769A CN 112464769 A CN112464769 A CN 112464769A CN 202011296355 A CN202011296355 A CN 202011296355A CN 112464769 A CN112464769 A CN 112464769A
- Authority
- CN
- China
- Prior art keywords
- detection
- roi
- layer
- detection layer
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 201
- 238000000034 method Methods 0.000 claims abstract description 26
- 238000012545 processing Methods 0.000 claims abstract description 9
- 238000010586 diagram Methods 0.000 claims description 17
- 238000011176 pooling Methods 0.000 claims description 15
- 238000000605 extraction Methods 0.000 claims description 12
- 238000013135 deep learning Methods 0.000 abstract description 7
- 238000013461 design Methods 0.000 abstract description 4
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 230000008878 coupling Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 238000005859 coupling reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000001629 suppression Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000007670 refining Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a high-resolution remote sensing image target detection method based on consistent multi-stage detection, which aims to realize more robust and efficient remote sensing image multi-class target detection. The invention constructs a target detection network for high-resolution remote sensing images based on a deep learning method, particularly designs two different detection layers, namely a robust detection layer and a high-efficiency detection layer, which are respectively suitable for processing rough RoI and fine RoI, and can adapt to data with different complexity degrees by adopting reasonable combination. And an appropriate detector combination can be selected according to the target detection difficulty degree of the data so as to obtain higher detection precision.
Description
Technical Field
The invention belongs to the technical field of information processing, and particularly relates to a high-resolution remote sensing image target detection method based on consistent multi-stage detection.
Background
The high-resolution remote sensing target detection task aims at: on one hand, the frame of the position area where the target is located is predicted; and on the other hand predict the class within each box.
Existing target detection methods can be divided into two categories, manual feature-based and depth feature-based. The traditional method adopts the characteristics of artificial design, and obtains good effect in a period of time. However, with the rise of deep learning, the disadvantage that the conventional method relies on human factors gradually emerges, and the deep learning can objectively and autonomously learn favorable features, so that the Average Accuracy (AP) of target detection is greatly improved.
The method based on the depth features adopts a deep learning target detection framework, and generally comprises the following aspects: extracting image convolution characteristics, predicting a candidate Region (namely Region of Interest (RoI)), pooling the RoI and detecting. The detection includes position regression and category prediction. The position regression refers to refining the frame of the RoI, so that the predicted frame is closer to a real target. Both the RoI prediction and detection stages include detection modules (combination of location regression and category prediction), so the design of the detection modules has an important influence on the average accuracy AP of the whole target detection network framework. A completely new probabilistic model-based approach is proposed by Gidaris, Spyros and Komodakis, Nikos in "S.Gidaris and N.Komodakis, Locnet: Improling localization acquisition for object detection, in Computer Vision and Pattern Recognition, pp.789-798,2016. Compared to a conventional regression model, the model divides the candidate box into a grid of n by n and estimates the probability of the boundary of the target box at the corresponding abscissa and ordinate. Although the method improves the perception of spatial details and improves the expression of position regression by using the convolution characteristic, the method ignores the category prediction which is another part in the detection, and ignores the relationship between the category prediction and the position regression, which is not beneficial to further improving the detection precision. Li, Jiannan et al, in "J.Li, X.Liang, J.Li, Y.Wei, T.xu, J.Feng, and S.Yan, Multistage object detection with group recurrence leaving, IEEE Transactions on Multimedia, vol.20, No.7, pp.1645-1655,2017," propose a group recurrence network consisting of three cascade networks of weakly supervised object segmentation, RoI generation and recurrence detection. The recursive detection utilizes the detection module to carry out multiple detections, and the detection accuracy is gradually improved. However, in the process of neglecting recursive detection, the initial rough RoI is far from the position of the real target, so that classification errors are caused, and a large number of targets are missed. Cai, Zhaowei and Nuno, Vasconcolos in "Z.Cai and N.Vasconcolos, Cascade r-cnn: removing in high quality object detection, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),2018, pp.6154-6162" propose a multi-stage object detection framework, employing a plurality of cascaded detection modules, which are trained together. In addition, different Non-Maximum suppression (NMS) thresholds are used on different modules (Non-Maximum suppression is a RoI process to distinguish between positive and negative samples). Although the method makes the detectors in series have certain robustness to the initial rough RoI by controlling the NMS threshold, the method still misses more rough RoI.
Disclosure of Invention
In order to overcome the defects of the conventional method in initial rough RoI processing, the invention provides a high-resolution remote sensing image target detection method for consistent multi-stage detection, so as to realize more robust and efficient multi-class target detection of remote sensing images. The invention constructs a target detection network for high-resolution remote sensing images based on a deep learning method, particularly designs two different detection layers, namely a robust detection layer and a high-efficiency detection layer, which are respectively suitable for processing rough RoI and fine RoI, and can adapt to data with different complexity degrees by adopting reasonable sequencing and combination. And an appropriate detector combination can be selected according to the target detection difficulty degree of the data so as to obtain higher detection precision.
A high-resolution remote sensing image target detection method based on consistent multi-stage detection is characterized by comprising the following steps:
step 1: constructing a target detection network, which comprises a convolution feature extraction layer, a RoI prediction layer and a CMD detection layer, wherein the CMD detection layer comprises three cascaded RoI pooling layers and a detection layer, the three cascaded detection layers adopt a robust detection layer, a high-efficiency detection layer or a combination of the robust detection layer and the high-efficiency detection layer, if the three cascaded detection layers are the combination of the robust detection layer and the high-efficiency detection layer, the robust detection layer is required to be arranged in front of the high-efficiency detection layer, the number of the robust detection layers in the combination is selected according to data difficulty, and the greater the difficulty is, the greater the number;
the robust detection layer takes the RoI characteristic graph as input and obtains a prediction result of the RoI category through a classifier; meanwhile, obtaining a single prediction result of the RoI frame regression which is universal among different categories through a universal regressor, and updating the RoI; inputting the updated RoI into the RoI pooling layer to obtain an updated RoI characteristic diagram;
the high-efficiency detection layer takes the RoI characteristic diagram as input and obtains a prediction result of the RoI category through a classifier; meanwhile, a plurality of prediction results of the RoI frame regression aiming at different classes are obtained through a multi-class regression device; selecting the RoI frame regression corresponding to the category according to the category prediction result obtained by the classifier, and updating the RoI; inputting the updated RoI into the RoI pooling layer to obtain an updated RoI characteristic diagram;
step 2: inputting a high-resolution remote sensing image to a target detection network, and outputting a convolution characteristic diagram through a convolution characteristic extraction layer; the RoI prediction layer performs RoI prediction processing on the convolution characteristic graph to obtain initial RoI; and inputting the initial RoI into a RoI pooling layer of the CMD detection layer to obtain a RoI characteristic diagram, then obtaining updated RoI through the detection layer, and updating for three times in total to obtain a final target detection result.
Furthermore, three cascaded detection layers in the CMD detection layer are all robust detection layers.
Furthermore, three cascaded detection layers in the CMD detection layer are two robust detection layers and one efficient detection layer in sequence.
Furthermore, three cascaded detection layers in the CMD detection layer are a robust detection layer and two efficient detection layers in sequence.
Furthermore, three cascaded detection layers in the CMD detection layer are all high-efficiency detection layers.
The invention has the beneficial effects that: (1) the target detection network comprises a CMD detection layer, and two different detection layers, namely a robust detection layer and a high-efficiency detection layer, are defined, so that the target detection network can adapt to target candidate regions RoI with different roughness degrees by reasonably selecting the detector, and the average accuracy AP of the target is better improved. (2) Under the condition that the initial RoI is rough, the RoI characteristics obtained by the RoI pooling layer contain excessive interference information, and target information is too little, so that the classifier makes mistakes. (3) Under the condition that the initial RoI is fine enough, the target information contained in the RoI features obtained by the RoI pooling layer is enough, and the classifier can correctly classify, and at the moment, the invention adopts the high-efficiency detection layer, and can effectively utilize the category difference to carry out high-efficiency position regression. (4) Because the network framework reasonably recombines the robust detection layer and the efficient detection layer, the obtained CMD detection layer can better adapt to the variation trend of RoI from coarse to fine, and a reasonable form of the CMD detection layer can be selected according to the difficulty of a data set.
Drawings
FIG. 1 is a schematic diagram of an object detection network of the present invention;
FIG. 2 is a schematic diagram of an efficient detection layer in the object detection network of the present invention;
in the figure, w-image width, h-image height, c-channel number of input feature map, k-class number of target, 1-background class in k +1, (1,1,4) and 4-in (1,1,4k) include the following four items: the target center point horizontal coordinate, the target center point vertical coordinate, the target width and the target height;
FIG. 3 is a schematic diagram of a robust detection layer in the object detection network of the present invention;
in the figure, w-image width, h-image height, c-channel number of input feature map, k-class number of target, 1-background class in k +1, 4 in (1,1,4) contains the following four items: the target center point horizontal coordinate, the target center point vertical coordinate, the target width and the target height;
FIG. 4 is a sample image in a DIOR dataset;
fig. 5 is a sample image in a HRRSD dataset.
Detailed Description
The present invention will be further described with reference to the following drawings and examples, which include, but are not limited to, the following examples.
Aiming at the defects of the existing deep learning method in initial rough RoI processing, the invention provides a novel high-resolution remote sensing target detection network and a method thereof, so as to realize more robust and efficient target detection. In particular, the present invention aims to improve the following aspects: 1) in the existing deep learning target detection method, the coupling relation between a classifier and a regressor is ignored, and the classification precision is reduced when rough RoI is input in the traditional coupling mode; 2) the existing multi-stage detection method neglects the selection of detection layer structures in different stages, and adopts a single detection layer structure which cannot adapt to the data structures of different stages; 3) the existing multi-stage detection method ignores the change of the RoI roughness from coarse to fine, and the RoI roughness influences the selection of a detection layer structure; 4) the consistency between the detection layer structure and the RoI roughness is neglected in the existing multi-stage detection method.
As shown in fig. 1, the present invention provides a high-resolution remote sensing target Detection network based on Consistent Multi-stage Detection (CMD), which includes a convolution feature extraction layer, a RoI prediction layer, and a CMD Detection layer.
An input image passes through a convolution neural network of a convolution feature extraction layer to obtain a convolution feature map; the convolution feature extraction layer may be ResNet, DenseNet, ResNext, VGG, GoogleNet, AlexNet, or the like.
Obtaining an initial RoI image with a possible target by the convolution feature map through a RoI prediction layer by utilizing a RoI prediction method, such as region suggestion network (RPN), Selective Search (Selective Search) and the like;
the CMD detection layer includes three cascaded RoI pooling layers and detection layers, each of which may be a robust detection layer, an efficient detection layer, or a combination of both. Obtaining a RoI characteristic diagram through a RoI pooling layer by the RoI and convolution characteristic diagrams; obtaining updated RoI through the detection layer; and obtaining a target detection result through two groups of pooling layers and detection layers similar to the above.
The invention researches the consistency between the structure of a detection layer and the RoI roughness input by the detection layer in a cascaded detection module: as the number of cascaded layers increases, the variation of the roughness of the input RoI of the detection layer is from coarse to fine. Thus, the detection layer may select CMD-RRR, CMD-RRF, CMD-RFF, CMD-FFF, where R represents the robust detection layer, F represents the efficient convolutional layer, and XXX in CMD-XXX refers to the selection of three cascaded detection layers in CMD. Taking CMD-RFF as an example, RFF represents three cascaded detection layers, and all inputs pass through a robust detection layer and then pass through two efficient detection layers.
The selection of XXX in the CMD-XXX follows two principles: order consistency, quantity consistency. Order consistency, which means that any robust detection layer R must be in front of all efficient detection layers F, so as to maintain the consistency of the input RoI roughness and the detection layer structure; the quantity consistency means that the quantity selection of the robust detection layer R and the efficient detection layer F is determined according to the difficulty of data, the data is complex, the detection difficulty is high, more robust detection layers R are used, such as CMD-RRR and CMD-RRF, and otherwise, more efficient detection layers F are used, such as CMD-RFF and CMD-FFF. Overall, the data of different roughness levels can be processed by the most appropriate detection layer, and therefore higher detection accuracy can be obtained.
The high-efficiency detection layer is a traditional detection layer structure and comprises a multi-class position regression layer and a class prediction layer, and a plurality of different target class RoI frame regressions and class predictions are obtained respectively; selecting a single RoI frame regression corresponding to the category from the multiple types of RoI frame regressions by utilizing category prediction; and updating the RoI by using the selected RoI frame regression and the category prediction to obtain the updated RoI. This embodiment employs a class k classifier and a "class k" regressor. Wherein, the result of the k-type classifier is used as the output classification result; however, the result of the "k-type" regressor includes a plurality of regression results, and the classification result of the k-type classifier is used to select a regression result of the corresponding category from the plurality of regression results as an output regression result, as shown in fig. 2.
The robust detection layer is a proposed detection layer structure and comprises a general position regression layer and a category prediction layer, and single general RoI frame regression and category prediction are respectively obtained; and updating the RoI by utilizing the regression of the general RoI frame and the category prediction to obtain the updated RoI. This embodiment employs a class k classifier and a "single class" regressor. Wherein, the result of the k-type classifier is used as the output classification result; the results of the "single-class" regressor are directly used as the output regression results, as shown in FIG. 3.
Inputting a high-resolution remote sensing image to a target detection network, and outputting a convolution characteristic diagram through a convolution characteristic extraction layer; the RoI prediction layer performs RoI prediction processing on the convolution characteristic graph to obtain initial RoI; and inputting the initial RoI into a RoI pooling layer of the fine CMD fine target detection layer to obtain a RoI characteristic diagram, then obtaining an updated RoI through the fine target detection layer, and updating for three times in the way to obtain a final target detection result image.
The effect of the method of the present invention can be further illustrated by the following simulation experiments.
1. Simulation data set
Two data sets, DIOR, HRRSD were used in the experiment.
The DIOR data set comprises 23463 remote sensing images, 192472 targets, 20 target categories: airplanes, airports, baseball fields, basketball fields, bridges, chimneys, dams, highway service areas, highway toll stations, ports, golf courses, ground track fields, overpasses, ships, stadiums, storage tanks, tennis courts, train stations, vehicles, and windmills. The data set address is: http:// www.escience.cn/peoples/gongcheng/DIOR. html.
The HRRSD data set adopts satellite imagery of Google maps and Baidu maps, the sizes of the images are not uniform, and the HRRSD data set comprises 21761 remote sensing images, 55740 targets and 13 target categories: ships, bridges, track and field grounds, storage tanks, basketball courts, tennis courts, airplanes, baseball fields, ports, automobiles, seventeen intersections, T-intersections and parking lots. Each class contains-4000 samples, and the sample size is relatively balanced. The data set address is: https:// github. com/CrazyStoneRoad/TGRS-HRRSD-Dataset.
2. Emulated content
Firstly, the effect of the combination processing of the four CMD detection layers is compared on two high-resolution remote sensing image target detection data sets respectively. Fig. 4 and 5 are sample images of the DIOR and HRRSD data sets, respectively, in which airplan and the like represent different object types and boxes represent object positions.
Table 1 shows the accuracy AP values of different types of targets and the average accuracy mAP value of all types of targets calculated after target detection is carried out on a DIOR data set by adopting the method, wherein CMD-RRR, CMD-RRF, CMD-RFF and CMD-FFF represent four combination forms of a CMD detection layer. It can be seen that in the DIOR data set, APs of various target detection methods are generally low (less than 61%), and the DIOR can be considered to belong to a data set with high difficulty; and it can be seen that CMD-RRR achieves the maximum average accuracy mAP on DIOR. The experimental result shows that the robust detection layer is more effective to the data set with higher difficulty and coarser initial RoI generated by the RoI extraction layer.
Table 2 shows the accuracy AP values of different classes of targets and the average accuracy mAP values of all classes of targets calculated after target detection on the HRRSD dataset by using the method of the present invention. It can be seen that, in the HRRSD data set, APs of various target detection methods are generally higher (more than 88%), and it can be considered that the HRRSD belongs to the data set with less difficulty; and it can be seen that CMD-RRF and CMD-RFF achieved higher mAP on HRRSD. The experimental result shows that the high-efficiency detection layer is more effective on a data set which is low in difficulty and fine enough in initial RoI generated by the RoI extraction layer.
TABLE 1
TABLE 2
In addition, other existing methods are also selected to perform target detection processing on the two data sets respectively, and compared with the method of the invention, the method comprises the following steps: R-CNN, RICNN, FasterR-CNN _ R50+ FPN, FasterR-CNN _ R101+ FPN methods, the accuracy AP values of different classes of targets and the average accuracy mAP values of all classes of targets calculated by these methods after target detection on DIOR data set are given in Table 3. Wherein, Faster R-CNN represents a classical target detection framework; r50 and r100 represent convolution feature extraction layers employing 50 layers of ResNet50 and 101 layers of ResNet101, respectively; FPN represents a feature pyramid module; CMD-XXX represents the CMD detection layer proposed by the present invention. For example, FasterR-CNN _ R50+ FPN + CMD-RRR, using the FasterR-CNN framework, the convolution feature extraction layer, 50 layers of ResNet50, and the feature pyramid module FPN, and the CMD-RRR module. It can be seen that the method of the present invention achieves the best accuracy on the data set.
TABLE 3
Claims (5)
1. A high-resolution remote sensing image target detection method based on consistent multi-stage detection is characterized by comprising the following steps:
step 1: constructing a target detection network, which comprises a convolution feature extraction layer, a RoI prediction layer and a CMD detection layer, wherein the CMD detection layer comprises three cascaded RoI pooling layers and a detection layer, the three cascaded detection layers adopt a robust detection layer, a high-efficiency detection layer or a combination of the robust detection layer and the high-efficiency detection layer, if the three cascaded detection layers are the combination of the robust detection layer and the high-efficiency detection layer, the robust detection layer is required to be arranged in front of the high-efficiency detection layer, the number of the robust detection layers in the combination is selected according to data difficulty, and the greater the difficulty is, the greater the number;
the robust detection layer takes the RoI characteristic graph as input and obtains a prediction result of the RoI category through a classifier; meanwhile, obtaining a single prediction result of the RoI frame regression which is universal among different categories through a universal regressor, and updating the RoI; inputting the updated RoI into the RoI pooling layer to obtain an updated RoI characteristic diagram;
the high-efficiency detection layer takes the RoI characteristic diagram as input and obtains a prediction result of the RoI category through a classifier; meanwhile, a plurality of prediction results of the RoI frame regression aiming at different classes are obtained through a multi-class regression device; selecting the RoI frame regression corresponding to the category according to the category prediction result obtained by the classifier, and updating the RoI; inputting the updated RoI into the RoI pooling layer to obtain an updated RoI characteristic diagram;
step 2: inputting a high-resolution remote sensing image to a target detection network, and outputting a convolution characteristic diagram through a convolution characteristic extraction layer; the RoI prediction layer performs RoI prediction processing on the convolution characteristic graph to obtain initial RoI; and inputting the initial RoI into a RoI pooling layer of the CMD detection layer to obtain a RoI characteristic diagram, then obtaining updated RoI through the detection layer, and updating for three times in total to obtain a final target detection result.
2. The method for detecting the target of the high-resolution remote sensing image based on the consistent multi-stage detection as claimed in claim 1, wherein: and three cascaded detection layers in the CMD detection layer are robust detection layers.
3. The method for detecting the target of the high-resolution remote sensing image based on the consistent multi-stage detection as claimed in claim 1, wherein: three cascaded detection layers in the CMD detection layer are two robust detection layers and one efficient detection layer in sequence.
4. The method for detecting the target of the high-resolution remote sensing image based on the consistent multi-stage detection as claimed in claim 1, wherein: three cascaded detection layers in the CMD detection layer are a robust detection layer and two efficient detection layers in sequence.
5. The method for detecting the target of the high-resolution remote sensing image based on the consistent multi-stage detection as claimed in claim 1, wherein: and three cascaded detection layers in the CMD detection layer are all high-efficiency detection layers.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011296355.1A CN112464769A (en) | 2020-11-18 | 2020-11-18 | High-resolution remote sensing image target detection method based on consistent multi-stage detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011296355.1A CN112464769A (en) | 2020-11-18 | 2020-11-18 | High-resolution remote sensing image target detection method based on consistent multi-stage detection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112464769A true CN112464769A (en) | 2021-03-09 |
Family
ID=74837208
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011296355.1A Pending CN112464769A (en) | 2020-11-18 | 2020-11-18 | High-resolution remote sensing image target detection method based on consistent multi-stage detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112464769A (en) |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20110079455A (en) * | 2009-12-30 | 2011-07-07 | 전남대학교산학협력단 | Computer-readable recorded medium having color correction-coefficient calculating program and color correcting method for image input/output device |
| CN105975925A (en) * | 2016-05-03 | 2016-09-28 | 电子科技大学 | Partially-occluded pedestrian detection method based on joint detection model |
| CN110059667A (en) * | 2019-04-28 | 2019-07-26 | 上海应用技术大学 | Pedestrian counting method |
| CN110168477A (en) * | 2016-11-15 | 2019-08-23 | 奇跃公司 | Deep learning system for cuboid detection |
| CN110378242A (en) * | 2019-06-26 | 2019-10-25 | 南京信息工程大学 | A kind of remote sensing target detection method of dual attention mechanism |
| CN110399809A (en) * | 2019-07-08 | 2019-11-01 | 北京亮亮视野科技有限公司 | The face critical point detection method and device of multiple features fusion |
| CN111161311A (en) * | 2019-12-09 | 2020-05-15 | 中车工业研究院有限公司 | Visual multi-target tracking method and device based on deep learning |
| CN111275044A (en) * | 2020-02-21 | 2020-06-12 | 西北工业大学 | Weak supervision target detection method based on sample selection and self-adaptive hard case mining |
| CN111553212A (en) * | 2020-04-16 | 2020-08-18 | 中国科学院深圳先进技术研究院 | Remote sensing image target detection method based on smooth frame regression function |
| CN111582226A (en) * | 2020-05-19 | 2020-08-25 | 中国人民解放军国防科技大学 | Candidate frame redundancy removing method in target detection |
| CN111950488A (en) * | 2020-08-18 | 2020-11-17 | 山西大学 | Improved fast-RCNN remote sensing image target detection method |
-
2020
- 2020-11-18 CN CN202011296355.1A patent/CN112464769A/en active Pending
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20110079455A (en) * | 2009-12-30 | 2011-07-07 | 전남대학교산학협력단 | Computer-readable recorded medium having color correction-coefficient calculating program and color correcting method for image input/output device |
| CN105975925A (en) * | 2016-05-03 | 2016-09-28 | 电子科技大学 | Partially-occluded pedestrian detection method based on joint detection model |
| CN110168477A (en) * | 2016-11-15 | 2019-08-23 | 奇跃公司 | Deep learning system for cuboid detection |
| CN110059667A (en) * | 2019-04-28 | 2019-07-26 | 上海应用技术大学 | Pedestrian counting method |
| CN110378242A (en) * | 2019-06-26 | 2019-10-25 | 南京信息工程大学 | A kind of remote sensing target detection method of dual attention mechanism |
| CN110399809A (en) * | 2019-07-08 | 2019-11-01 | 北京亮亮视野科技有限公司 | The face critical point detection method and device of multiple features fusion |
| CN111161311A (en) * | 2019-12-09 | 2020-05-15 | 中车工业研究院有限公司 | Visual multi-target tracking method and device based on deep learning |
| CN111275044A (en) * | 2020-02-21 | 2020-06-12 | 西北工业大学 | Weak supervision target detection method based on sample selection and self-adaptive hard case mining |
| CN111553212A (en) * | 2020-04-16 | 2020-08-18 | 中国科学院深圳先进技术研究院 | Remote sensing image target detection method based on smooth frame regression function |
| CN111582226A (en) * | 2020-05-19 | 2020-08-25 | 中国人民解放军国防科技大学 | Candidate frame redundancy removing method in target detection |
| CN111950488A (en) * | 2020-08-18 | 2020-11-17 | 山西大学 | Improved fast-RCNN remote sensing image target detection method |
Non-Patent Citations (3)
| Title |
|---|
| S. LIU, G ET AL: "Light and Fast Hand Pose Estimation From Spatial-Decomposed Latent Heatmap", 《IEEE ACCESS》 * |
| ZHAOWEI CAI ET AL: "Cascade R-CNN：Delving into High Quality Object Detection", 《ARXIV:1712.00726V1》 * |
| 陈雨薇: "基于改进MTCNN模型的人脸检测与面部关键点定位", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10984532B2 (en) | Joint deep learning for land cover and land use classification | |
| EP3614308B1 (en) | Joint deep learning for land cover and land use classification | |
| CN109614985B (en) | Target detection method based on densely connected feature pyramid network | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN106778835B (en) | Remote sensing image airport target identification method fusing scene information and depth features | |
| CN107609525B (en) | Remote sensing image target detection method for constructing convolutional neural network based on pruning strategy | |
| CN108596055B (en) | Airport target detection method of high-resolution remote sensing image under complex background | |
| CN110796168A (en) | Improved YOLOv 3-based vehicle detection method | |
| CN111179217A (en) | Attention mechanism-based remote sensing image multi-scale target detection method | |
| CN108537824B (en) | Feature map enhanced network structure optimization method based on alternating deconvolution and convolution | |
| CN110163213B (en) | Remote sensing image segmentation method based on disparity map and multi-scale depth network model | |
| CN107633226B (en) | Human body motion tracking feature processing method | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN111814621A (en) | Multi-scale vehicle and pedestrian detection method and device based on attention mechanism | |
| CN106707296A (en) | Dual-aperture photoelectric imaging system-based unmanned aerial vehicle detection and recognition method | |
| CN107239762A (en) | Patronage statistical method in a kind of bus of view-based access control model | |
| CN111967313B (en) | Unmanned aerial vehicle image annotation method assisted by deep learning target detection algorithm | |
| CN109492596B (en) | Pedestrian detection method and system based on K-means clustering and regional recommendation network | |
| CN111738114B (en) | Vehicle target detection method based on anchor-free accurate sampling remote sensing image | |
| CN111126278A (en) | Target detection model optimization and acceleration method for few-category scene | |
| CN111833353B (en) | Hyperspectral target detection method based on image segmentation | |
| CN108648210B (en) | Rapid multi-target detection method and device under static complex scene | |
| CN113505670A (en) | Remote sensing image weak supervision building extraction method based on multi-scale CAM and super-pixels | |
| Li et al. | RailNet: An information aggregation network for rail track segmentation | |
| CN113610024B (en) | Multi-strategy deep learning remote sensing image small target detection method |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20210309 |
|
| WD01 | Invention patent application deemed withdrawn after publication |