CN109325449B - Convolutional neural network target detection framework based on sample updating - Google Patents
Convolutional neural network target detection framework based on sample updating Download PDFInfo
- Publication number
- CN109325449B CN109325449B CN201811112898.6A CN201811112898A CN109325449B CN 109325449 B CN109325449 B CN 109325449B CN 201811112898 A CN201811112898 A CN 201811112898A CN 109325449 B CN109325449 B CN 109325449B
- Authority
- CN
- China
- Prior art keywords
- sample
- remote sensing
- sensing image
- target detection
- detection
- 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.)
- Active
Links
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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a convolutional neural network target detection framework based on sample updating, which is used for solving the problem of target detection on a large-range remote sensing image. The frame comprises two stages: in the first stage, an initial sample set is used for training a basic model, the model is used for testing a large-range remote sensing image test set, and a large amount of false detections and missed detections are generated in a detection result; and in the second stage, generating an artificially synthesized sample by using a sample updating method, updating the initial sample set, and then finely adjusting the model parameters in the first stage by using the updated sample set to obtain a final target detection model. The method can effectively inhibit the phenomena of false detection and missed detection in the target detection of the remote sensing image in a large range, and embodies absolute effectiveness and superiority.
Description
Technical Field
The invention belongs to the technical field of remote sensing image processing, and relates to a convolutional neural network target detection framework based on sample updating, which is mainly applied to target detection of large-range remote sensing images.
Background
In recent years, more and more target detection methods based on deep learning are used for detecting a spatial target on an optical remote sensing image. Although the methods based on deep learning have made certain breakthrough progress in the field of target detection, it is not considered that these methods can achieve higher accuracy in a test sample, but cannot be directly and effectively applied to a wide range of remote sensing images. In practical application, the area to be detected is usually much larger than the range of a test sample, for example, a remote sensing image of a scene or a remote sensing image of an urban area, so that the remote sensing image of a large area contains a large number of surface feature targets and complex and variable environments, and a mode similar to the target to be detected inevitably exists, so that a large number of false detections and missed detections are generated, and the precision is greatly reduced.
In order to solve the problem, the invention provides a convolutional neural network target detection framework based on sample updating. The method comprises the steps of generating a certain number of artificially synthesized samples, taking an area which is far away from a target to be detected and is easy to be subjected to false detection and cannot be contained in an initial sample set as a negative sample, taking a missed detection target with a new shape, size and color in a large-range remote sensing image test set as a positive sample, and simultaneously training. Sample sets of targets such as airplanes, ships, oil storage tanks and the like are respectively manufactured and tested, and results show that the method has a remarkable improvement in test effect on large-range remote sensing images.
Disclosure of Invention
The invention aims to solve the problem that the precision is greatly reduced due to a large amount of false detections and missed detections when a large-range remote sensing image is tested by the conventional deep learning target detection method, and provides a convolutional neural network target detection framework based on sample updating, which is suitable for all the conventional deep learning target detection methods and can greatly improve the test effect of the conventional deep learning target detection methods on the large-range remote sensing image.
The technical scheme of the invention is a convolutional neural network target detection framework based on sample updating, which mainly comprises training and testing in a stage I, generation of an artificial synthesis sample and training and testing in a stage II:
stage one: training a basic model by using an initial data set, wherein the basic model can be any target detection model based on deep learning, and testing a large-range remote sensing image test set; generation of artificially synthesized samples: the artificially synthesized sample is synthesized by a background image and a positive sample, the background image is an image which is cut out by taking the false detection on the large-range remote sensing image test set as the center and has the same size with the sample in the initial data set, the positive sample is a missed detection target on a certain number of large-range remote sensing image test sets, and the positive sample is randomly placed on the background image to obtain the artificially synthesized sample; and a second stage: and updating the initial data set by using the artificial synthesis sample, finely adjusting the parameters of the stage-one basic model by using the updated data set to obtain a new basic model, and retesting the large-range remote sensing image test set by using the new basic model.
The invention has the following characteristics: 1) the framework is suitable for any target detection model based on deep learning, and is realized only by changing a basic model; 2) the framework provides a method for generating an artificially synthesized sample to integrate more background information and diversified target information, so that the number of false detection and missed detection in the process of testing a large-range remote sensing image is reduced, and the testing precision is improved; 3) The frame can greatly improve the testing precision by only adding a small amount of artificially synthesized samples.
Drawings
FIG. 1 is a general flow diagram of a sample update based convolutional neural network object detection framework implemented by the present invention;
FIG. 2 is a schematic illustration of the generation of a synthetic sample;
FIG. 3 is a graph of accuracy versus recall of the present invention and other methods.
Detailed Description
In order to make the purpose and technical solutions of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention. It should be noted that the embodiments and features of the embodiments in the present application may be arbitrarily combined with each other without conflict.
Fig. 1 illustrates a general flow diagram of a convolutional neural network object detection framework based on sample update proposed in the present invention. The framework consists of training and testing in stage one, generation of artificially synthesized samples, and training and testing in stage two. Stage one: training a basic model by using an initial data set, wherein the basic model can be any target detection model based on deep learning, and testing a large-range remote sensing image test set; generation of artificially synthesized samples: the artificially synthesized sample is synthesized by a background image and a positive sample, the background image is an image which is cut out by taking the false detection on the large-range remote sensing image test set as the center and has the same size with the sample in the initial data set, the positive sample is a missed detection target on a certain number of large-range remote sensing image test sets, and the positive sample is randomly placed on the background image to obtain the artificially synthesized sample; and a second stage: and updating the initial data set by using the artificial synthesis sample, fine-tuning the stage-one basic model parameters by using the updated data set to obtain a new basic model, and retesting the large-range remote sensing image test set by using the new basic model.
Fig. 2 illustrates the generation process of a synthetic sample. Firstly, taking a false detection target on a large-range remote sensing image test set as a center, and cutting an area with the same size as a sample in an initial sample set as a background image; secondly, selecting a missed detection target as a positive sample; and finally, randomly placing the positive sample in the background image to generate a synthesized sample.
Fig. 3 illustrates the accuracy-recall curve of the present invention and other methods in testing three exemplary remote sensing images over a wide range. The result shows that the method (SUCNN) of the invention is superior to other methods (SSD) on three example wide-range remote sensing images selected for each type of targets, such as three types of ground object targets, namely an airplane, a ship and an oil storage tank.
Claims (1)
1. A convolutional neural network target detection framework based on sample update is characterized in that:
the basic model in the target detection framework is a target detection model based on a convolutional neural network and is selected for a specific detection target;
training a selected basic model by using an initial sample set, wherein samples in the initial sample set are images with the length and width not more than 1000 pixels, testing a large-range remote sensing image testing set by using the trained basic model, and the images in the large-range remote sensing image testing set are images with the length and width more than 4000 pixels;
a second stage, manufacturing a manual synthesis sample based on false detection and missing detection on the large-range remote sensing image test set in the first stage, updating an initial sample set by using the manual synthesis sample, finely adjusting parameters of the basic model in the first stage by using the updated sample set to obtain a new basic model, and retesting the large-range remote sensing image test set by using the new basic model; the artificially synthesized sample meets the following requirements:
the size of the artificially synthesized sample is consistent with that of the sample in the initial sample set; each artificial synthesis sample is synthesized by a positive sample and a background image; the positive sample is a selected number of missed samples in the large-range remote sensing image test set; and intercepting the background image by taking the error detection target in the large-range remote sensing image test set as a center, and randomly placing the positive sample on the background image to obtain an artificially synthesized sample.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2018100061643 | 2018-01-04 | ||
| CN201810006164 | 2018-01-04 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109325449A CN109325449A (en) | 2019-02-12 |
| CN109325449B true CN109325449B (en) | 2022-06-21 |
Family
ID=65265791
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811112898.6A Active CN109325449B (en) | 2018-01-04 | 2018-09-25 | Convolutional neural network target detection framework based on sample updating |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109325449B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062885B (en) * | 2019-12-09 | 2023-09-12 | 中国科学院自动化研究所 | Mark detection model training and mark detection method based on multi-stage transfer learning |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101221623B (en) * | 2008-01-30 | 2011-11-23 | 北京中星微电子有限公司 | Object type on-line training and recognizing method and system thereof |
| US9633282B2 (en) * | 2015-07-30 | 2017-04-25 | Xerox Corporation | Cross-trained convolutional neural networks using multimodal images |
| US10140522B2 (en) * | 2015-12-16 | 2018-11-27 | Intel Corporation | Fully convolutional pyramid networks for pedestrian detection |
| CN105844228B (en) * | 2016-03-21 | 2019-02-19 | 北京航空航天大学 | A kind of remote sensing images cloud detection method of optic based on convolutional neural networks |
| CN106022307B (en) * | 2016-06-08 | 2019-09-27 | 中国科学院自动化研究所 | Remote sensing images ship detection method based on ship rotation rectangular space |
| CN107066995A (en) * | 2017-05-25 | 2017-08-18 | 中国矿业大学 | A kind of remote sensing images Bridges Detection based on convolutional neural networks |
| CN107273832B (en) * | 2017-06-06 | 2020-09-22 | 青海省交通科学研究院 | License plate recognition method and system based on integral channel characteristics and convolutional neural network |
-
2018
- 2018-09-25 CN CN201811112898.6A patent/CN109325449B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109325449A (en) | 2019-02-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110136170B (en) | Remote sensing image building change detection method based on convolutional neural network | |
| US11151403B2 (en) | Method and apparatus for segmenting sky area, and convolutional neural network | |
| CN110097129B (en) | Remote sensing target detection method based on profile wave grouping characteristic pyramid convolution | |
| CN110516539A (en) | Remote sensing image building extracting method, system, storage medium and equipment based on confrontation network | |
| Amitrano et al. | A new framework for SAR multitemporal data RGB representation: Rationale and products | |
| CN108960404B (en) | Image-based crowd counting method and device | |
| CN111738111A (en) | Road extraction method of high-resolution remote sensing image based on multi-branch cascade void space pyramid | |
| CN110555841B (en) | SAR image change detection method based on self-attention image fusion and DEC | |
| Santos et al. | Long‐term spatial dynamics in vegetated seascapes: fragmentation and habitat loss in a human‐impacted subtropical lagoon | |
| CN110163213A (en) | Remote sensing image segmentation method based on disparity map and multiple dimensioned depth network model | |
| CN110866494A (en) | Optical remote sensing image-based town group extraction method and system | |
| CN108364305B (en) | Vehicle-mounted camera video target tracking method based on improved DSST | |
| KR102496876B1 (en) | Apparatus and method for forecasting the seasonal forest fire risk | |
| CN110633633B (en) | Remote sensing image road extraction method based on self-adaptive threshold | |
| CN111582102B (en) | Remote sensing data refined classification method and device based on multi-mode end-to-end network | |
| CN106651917A (en) | Image target tracking algorithm based on neural network | |
| CN109325449B (en) | Convolutional neural network target detection framework based on sample updating | |
| CN110084181A (en) | A kind of remote sensing images Ship Target Detection method based on sparse MobileNetV2 network | |
| CN105913451B (en) | A kind of natural image superpixel segmentation method based on graph model | |
| CN109360231B (en) | Sea ice remote sensing image simulation method for generating confrontation network based on fractal depth convolution | |
| CN109448039B (en) | Monocular vision depth estimation method based on deep convolutional neural network | |
| CN107194917A (en) | In-orbit SAR image change detection based on DAP and ARELM | |
| CN117115669B (en) | Object-level ground object sample self-adaptive generation method and system with double-condition quality constraint | |
| David et al. | Mapping mangrove forest from LiDAR data using object-based image analysis and Support Vector Machine: The case of Calatagan, Batangas | |
| CN105741277A (en) | ViBe (Visual Background Extractor) algorithm and SLIC (Simple Linear Iterative Cluster) superpixel based background difference 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |