CN113449666A - Remote sensing image multi-scale target detection method based on data fusion and feature selection - Google Patents
Remote sensing image multi-scale target detection method based on data fusion and feature selection Download PDFInfo
- Publication number
- CN113449666A CN113449666A CN202110766948.8A CN202110766948A CN113449666A CN 113449666 A CN113449666 A CN 113449666A CN 202110766948 A CN202110766948 A CN 202110766948A CN 113449666 A CN113449666 A CN 113449666A
- Authority
- CN
- China
- Prior art keywords
- image
- fused
- training
- remote sensing
- data fusion
- 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
- 230000004927 fusion Effects 0.000 title claims abstract description 59
- 238000001514 detection method Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 71
- 230000002708 enhancing effect Effects 0.000 claims abstract description 4
- 238000000034 method Methods 0.000 claims description 25
- 238000012360 testing method Methods 0.000 claims description 17
- 238000000605 extraction Methods 0.000 claims description 8
- 238000010586 diagram Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 6
- 239000000126 substance Substances 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 14
- 230000006870 function Effects 0.000 description 6
- 238000012800 visualization Methods 0.000 description 5
- 230000006872 improvement Effects 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Images
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a remote sensing image multi-scale target detection method based on data fusion and feature selection, which comprises the following steps: step S1: acquiring a training set, and counting the number of bounding boxes of each category in the whole training set; step S2: enhancing and filling the number difference of the bounding boxes of each category in the training set through data fusion, so that the number of the bounding boxes of each category in the training set is balanced, and a final training set for model training is obtained; step S3: and training the training set processed in the step S2 by using a FoveaBox network, and identifying the targets with different sizes on different scale feature maps. The combination of the multi-scale feature expression and selection and the data fusion enhancement can cope with images with relatively complex backgrounds in the remote sensing images, reduce the influence of class imbalance, better conform to the scenes used by the remote sensing images, improve the performance of the model and improve the effect of identifying different-scale targets in the remote sensing images.
Description
Technical Field
The invention relates to the technical field of remote sensing image target detection, in particular to a remote sensing image multi-scale target detection method based on data fusion and feature selection.
Background
The remote sensing image target detection algorithm based on deep learning has the problem of sample imbalance, meanwhile, in the multi-task remote sensing image target detection, detection objects with different sizes exist, and the model has a certain limit on the generalization of ground feature scales, so that the sample imbalance and the diversity of the remote sensing image scales bring great battles for the target detection task.
To mitigate the negative effects of sample imbalances, Pang J et al propose a simple and efficient balance learning target detection framework Libra R-CNN. It integrates three new components: IoU equalized sample, equalized feature pyramid, and equalized L1loss functions, respectively, for reducing imbalances at the sample, feature, and target levels. Shrivastava A et al propose a method of "on-line hard case mining" (OHEM) to select some samples which are difficult to predict as training samples, thereby improving the problem of poor detection effect caused by unbalanced sample number. Cao Y et al propose a simple and effective sampling and learning strategy (PISA) that directs the emphasis of the training process to important samples selected from the basic samples, which are more effective in training focused on the original samples. An image pyramid Scale Normalization (SNIP) training scheme selectively back-propagates gradients of different sized target instances according to changes in image scale.
A large number of targets with different scales and small samples exist in the optical remote sensing image, and the multi-scale feature fusion can effectively improve the detection effect of the small targets and the different targets. The SSD algorithm carries out prediction of different scales by carrying out regression on a target box on a plurality of feature maps of different sizes of VGG16, and finally the prediction effect of the small target is superior to that of other single-stage networks. The FPN network provides a top-down structure with feature layer fusion on the basis of a fast RCNN frame, and the structure can effectively extract feature information of different scales of a picture. Since the remote sensing image has targets of various scales, the multi-scale fusion structure has excellent effect in remote sensing target detection, and meanwhile, the structure becomes the most common multi-scale feature extraction network. The subsequent improvements of YOLO v2 and YOLO v3 also add multi-scale improvement to the detection precision. The RetinaNet model adopts FPN as a feature extraction network, provides a Focal local Loss function to reduce the influence of positive and negative samples on the precision, and adds a bottom-up fusion structure on the basis of the FPN in the PaNet model.
Although great progress is made in the detection of the remote sensing image target, the following defects still exist:
1. first, the improved solutions for sample imbalance on the network structure above are not universally applicable to all target detection frameworks.
2. Secondly, although the commonly used multi-scale method can take target information of different scales into account on the remote sensing image, the feature layers of different scales cannot be well reflected on the target information of each scale, and the commonly used multi-scale method identifies targets of all sizes on each feature map, so that the target identification effect with weak feature information is poor.
In summary, a remote sensing image multi-scale target detection method based on data fusion and feature selection is urgently needed to solve the problems in the prior art.
Disclosure of Invention
The invention aims to provide a remote sensing image multi-scale target detection method based on data fusion and feature selection, aims to solve the problem that the target detection task is greatly challenged due to sample imbalance and the diversity of remote sensing image scales, and adopts the following specific technical scheme:
a remote sensing image multi-scale target detection method based on data fusion and feature selection comprises the following steps:
step S1: acquiring a training set, and counting the number of bounding boxes of each category in the whole training set;
step S2: enhancing and filling the number difference of the bounding boxes of each category in the training set through data fusion, so that the number of the bounding boxes of each category in the training set is balanced, and a final training set for model training is obtained;
step S3: and training the training set processed in the step S2 by using a FoveaBox network, and identifying the targets with different sizes on different scale feature maps.
Preferably, in the above technical solution, in the step S1, the remote sensing image target detection data set is proportionally divided into a training set and a test set.
Preferably, in the above technical solution, the data fusion enhancement in step S2 specifically includes:
step S2.1: taking an image to be enhanced of a category to be enhanced, then taking a background image which does not contain any target, and obtaining the maximum height h and the maximum width w of the image to be enhanced and the background image;
step S2.2: creating a fusion image with the height h, the width w, the channel number 3 and the pixel value 0;
step S2.3: and fusing the fused image and the image to be enhanced to obtain a fused image fused with the image to be enhanced, and fusing the fused image fused with the image to be enhanced and the background image to obtain a final fused image.
Preferably, in the above technical solution, in step S2.3, the fused image and the image to be enhanced are fused according to formula (1) to obtain a fused image fused with the image to be enhanced, and the fused image fused with the image to be enhanced is fused with the background image according to formula (2) to obtain a final fused image;
wherein the content of the first and second substances,
the representation matrixes are added according to corresponding coordinates, and x represents the multiplication of the matrixes; coefficient theta is artificially set to [0, 1 ]]A constant of (d); i represents an image to be enhanced; v represents a fused image before fusion; m represents a fused image into which an image to be enhanced is fused; pkRepresenting a background image; v represents the final fused image.
Preferably, in the above technical solution, the step S3 specifically includes:
step S3.1: training the training set obtained in the step S2 by adopting a FoveaBox network, and obtaining a plurality of feature maps with different scales by using the training images in the training set through an FPN feature extraction network;
step S3.2: each feature map respectively selects a plurality of targets in different bounding box area ranges for prediction, so that the feature maps in different scales can predict the targets in different sizes, and one target can be predicted on the plurality of feature maps according to the area;
step S3.3: the feature diagram is trained through a classification branch and a boundary box branch, the classification branch is trained by using a Focal Loss function, and the boundary box branch is trained by using a Smooth L1Loss function; after training is finished, the classification branch and the boundary box branch of the model respectively predict the classification and the boundary box of the target in the characteristic diagram to obtain a final output result.
Preferably, in the above technical solution, in the process of training the boundary box branch, an area smaller than the real boundary box is adopted for training.
Preferably, in step S3.1, 5 feature maps with different sizes are obtained from the training pictures in the training set through the FPN feature extraction network.
Preferably, in the above technical solution, in step S3.2, 5 feature maps with different scales are respectively selected and predicted for targets in (1,64), (32,128), (64,256), (128,512), (256,2048)) bounding box pixel area ranges.
The technical scheme of the invention has the following beneficial effects:
in order to solve the problem of unbalanced sample number, a data enhancement strategy based on image fusion is provided, the diversity of a sample library is improved through image fusion enhancement (namely data fusion enhancement), the influence of class imbalance can be reduced, and the result generalization of model training is better; secondly, carrying out feature expression and selection of multi-scale target detection on the optical remote sensing image by adopting a FoveaBox, identifying a target with a corresponding scale by utilizing a feature map with a proper scale, and simultaneously allowing the target with the same scale to be predicted on an adjacent feature map; the combination of the multi-scale feature expression and selection and the data fusion enhancement method can deal with the image with a relatively complex background in the remote sensing image, reduce the influence of class imbalance, better conform to the scene used by the remote sensing image, improve the performance of the model and improve the effect of identifying the target with different scales in the remote sensing image.
The image fusion enhancement provided by the invention is to carry out sample unbalance processing on a data layer, can be suitable for all target detection frames, and solves the problem that the existing method cannot be generally suitable for all target detection frames. In addition, the target with different sizes is predicted on the feature maps with different scales by the feature expression and the selected multi-scale network, the target features contained in different feature layers are fully utilized for better target identification, and the problem that the common multi-scale method identifies the targets with all sizes on each feature map, so that the target identification effect with weaker feature information is poor is solved.
In addition to the objects, features and advantages described above, other objects, features and advantages of the present invention are also provided. The present invention will be described in further detail below with reference to the drawings.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate embodiments of the invention and, together with the description, serve to explain the invention and not to limit the invention. In the drawings:
FIG. 1 is a schematic process flow diagram of the process of the present invention;
FIG. 2a is an original image before image fusion;
FIG. 2b is a background image I before image fusion;
FIG. 2c is a second background image before image fusion;
FIG. 2d is a background image three before image fusion;
FIG. 3a is a graph of the effect of FIG. 2a without fusion;
FIG. 3b is a graph showing the effect of FIG. 2a fused with FIG. 2 b;
FIG. 3c is a graph of the effect of FIG. 2a fused with FIG. 2 c;
FIG. 3d is a graph showing the effect of FIG. 2a fused with FIG. 2 d;
FIG. 4 is a schematic diagram of candidate box prediction based on a target center region;
FIG. 5a is a realistic annotation of scenario one in a test set test result visualization;
FIG. 5b is a realistic annotation of scenario two in a test set test result visualization;
FIG. 5c is a realistic annotation of scenario three in a test set test result visualization;
FIG. 5d is a realistic annotation of scenario four in the test set test result visualization;
FIG. 6a is the result of FIG. 5a using FPN network prediction;
FIG. 6b is the result of FIG. 5b using FPN network prediction;
FIG. 6c is the result of FIG. 5c using FPN network prediction;
FIG. 6d is the result of FIG. 5d using FPN network prediction;
FIG. 7a is the results of FIG. 5a using RetinaNet prediction;
FIG. 7b is the result of FIG. 5b predicted using RetinaNet;
FIG. 7c is the results of FIG. 5c using RetinaNet prediction;
FIG. 7d is the results of FIG. 5d using RetinaNet prediction;
FIG. 8a is the result of FIG. 5a using FoveaBox prediction;
FIG. 8b is the result of FIG. 5b using FoveaBox prediction;
FIG. 8c is the result of FIG. 5c using FoveaBox prediction;
FIG. 8d is the result of FIG. 5d using FoveaBox prediction;
FIG. 9a is the predicted result of FIG. 5a using FoveaBox and data fusion to enhance binding;
FIG. 9b is the predicted result of FIG. 5b using FoveaBox and data fusion to enhance binding;
FIG. 9c is the predicted result of FIG. 5c using FoveaBox and data fusion to enhance binding;
FIG. 9d is the predicted result of FIG. 5d using FoveaBox and data fusion to enhance binding.
Detailed Description
In order that the invention may be more fully understood, there now follows a more detailed description of the invention, and there is shown preferred embodiments of the invention. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
Example 1:
referring to fig. 1, a remote sensing image multi-scale target detection method based on data fusion and feature selection includes the following steps:
step S1: acquiring a training set, and counting the number of bounding boxes of each category in the whole training set;
preferably, in step S1, the remote sensing image target detection data set is divided into a training set and a test set according to a ratio of 8:2, where the training set is used for the processing in step S1, and the test set is used for the subsequent precision test.
Step S2: enhancing and filling the number difference of the bounding boxes of each category in the training set through data fusion, so that the number of the bounding boxes of each category in the training set is balanced (namely the number of the bounding boxes is consistent), and a final training set for model training is obtained;
referring to fig. 2a to fig. 3d, the data fusion enhancement in step S2 specifically includes:
step S2.1: taking an image I to be enhanced which needs to be enhanced, and then taking a background image P which does not contain any targetkAcquiring the maximum height h and the maximum width w of the image to be enhanced and the background image;
namely: h is max (h)I,hk)，w＝max(wI,wk) Wherein h isI、wIRespectively representing the height and width, h, of the image to be enhancedk、 wkRespectively representing the height and width of the background image;
step S2.2: creating a fused image v with height h, width w, number of channels 3 and pixel value 0, namely: v ═ zero (h, w, 3);
step S2.3: fusing the fused image and the image to be enhanced according to the formula (1) to obtain a fused image m fused with the image to be enhanced, and fusing the fused image fused with the image to be enhanced and the background image according to the formula (2) to obtain a final fused image V;
wherein the content of the first and second substances,
the representation matrixes are added according to corresponding coordinates, and x represents the multiplication of the matrixes; coefficient theta is artificially set to [0, 1 ]]A constant of (d); i represents an image to be enhanced; v represents a fused image before fusion; m represents the fusionMerging the fused images after the images to be enhanced are combined; pkRepresenting a background image; v represents the final fused image.
Step S3: and training the training set processed in the step S2 by using a FoveaBox network, and identifying the targets with different sizes on different scale feature maps.
Preferably, the step S3 specifically includes:
step S3.1: training the training set obtained in the step S2 by using a FoveaBox network, and obtaining a plurality of feature maps with different scales by using the training images in the training set through an FPN feature extraction network (namely an FPN feature pyramid network);
step S3.2: each feature map respectively selects a plurality of targets in different bounding box area ranges for prediction, so that the feature maps in different scales can predict the targets in different sizes, and one target can be predicted on the plurality of feature maps according to the area;
step S3.3: the feature diagram is trained through a classification branch and a boundary box branch, the classification branch is trained by using a Focal Loss function, and the boundary box branch is trained by using a Smooth L1Loss function; after training is finished, the classification branch and the boundary box branch of the model respectively predict the classification and the boundary box of the target in the characteristic diagram to obtain a final output result.
Preferably, in the process of training the boundary box branches, an area smaller than the real boundary box is adopted for training; as shown in fig. 4, the bright frame is a supervised training boundary frame used during training, the dark frame is a real labeled boundary frame, and the region model smaller than the real boundary frame can learn more representative features, so that the robustness of the model is improved.
Preferably, in step S3.1, 5 feature maps with different sizes are obtained from the training pictures in the training set through the FPN feature extraction network.
Preferably, in this embodiment, in step S3.2, 5 feature maps with different scales are selected to predict the target of the (1,64), (32,128), (64,256), (128,512), (256,2048)) bounding box pixel area range.
Referring to fig. 5a to fig. 9d, this embodiment further provides a comparison between the method in this embodiment and the commonly used cases of processing four scenes by FPN, RetinaNet, FoveaBox, and the like.
After dividing the nwuvhr 10 data set into test sets, the method of this embodiment is used to compare the test accuracy with the methods of FPN, RetinaNet, FoveaBox, etc. on the test sets, and the specific comparison results are shown in table 1.
TABLE 1 test accuracy contrast table
As can be seen from fig. 5a to 9d and table 1, the average accuracies of the FoveaBox network and the data fusion enhancement combination (i.e., the FoveaBox & fusion items in table 1) are respectively 6.84%, 7.17% and 4.78% higher than those of FPN, RetinaNet and FoveaBox (the FPN, RetinaNet and FoveaBox are all names of commonly used multi-scale networks), and the accuracies of the ship, baseball field, tennis court, basketball court, harbor, bridge and vehicle categories after the network is selected and expressed by image fusion and feature selection are all improved.
As can be seen from the visualization results in fig. 5a to fig. 9d, the multi-scale selection and expression model with image fusion enhancement (i.e. data fusion enhancement) has better recognition effect on the airplane category, ship, harbor, and vehicle category than other multi-scale networks, especially reduces false detection rate on tennis courts, vehicles, and other categories, and contains the image of the ship on the bank, and FPN, RetinaNet, and FoveaBox are difficult to distinguish the bank from the ship well, while the method combining FoveaBox and data fusion enhancement is relatively more accurate in recognizing the target under the complex background.
The result shows that the image fusion enhancement (namely, data fusion enhancement) can eliminate the problem of class imbalance in the training data, and the effectiveness of the image fusion enhancement can be seen in comparison of the FoveaBox method without data fusion enhancement and the FoveaBox method combined with the image fusion enhancement, so that the performance and the robustness of the model are improved.
In conclusion, the diversity of sample scenes is enhanced through image fusion, and the result of selecting and expressing the multi-scale features on the remote sensing image integrally shows that the targets with different scales and sizes in the optical remote sensing image can be reasonably predicted.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (8)
1. A remote sensing image multi-scale target detection method based on data fusion and feature selection is characterized by comprising the following steps:
step S1: acquiring a training set, and counting the number of bounding boxes of each category in the whole training set;
step S2: enhancing and filling the number difference of the bounding boxes of each category in the training set through data fusion, so that the number of the bounding boxes of each category in the training set is balanced, and a final training set for model training is obtained;
step S3: and training the training set processed in the step S2 by using a FoveaBox network, and identifying the targets with different sizes on different scale feature maps.
2. The method for detecting the multi-scale target of the remote sensing image based on the data fusion and the feature selection as claimed in claim 1, wherein in the step S1, the target detection data set of the remote sensing image is proportionally divided into a training set and a test set.
3. The method for detecting the remote sensing image multi-scale target based on the data fusion and the feature selection as claimed in claim 1, wherein the data fusion enhancement in the step S2 is specifically as follows:
step S2.1: taking an image to be enhanced of a category to be enhanced, then taking a background image which does not contain any target, and obtaining the maximum height h and the maximum width w of the image to be enhanced and the background image;
step S2.2: creating a fusion image with the height h, the width w, the channel number 3 and the pixel value 0;
step S2.3: and fusing the fused image and the image to be enhanced to obtain a fused image fused with the image to be enhanced, and fusing the fused image fused with the image to be enhanced and the background image to obtain a final fused image.
4. The method for detecting the remote sensing image multi-scale target based on the data fusion and the feature selection according to claim 3, wherein in step S2.3, the fused image and the image to be enhanced are fused according to formula (1) to obtain a fused image fused with the image to be enhanced, and the fused image fused with the image to be enhanced is fused with the background image according to formula (2) to obtain a final fused image;
wherein the content of the first and second substances,
the representation matrixes are added according to corresponding coordinates, and x represents the multiplication of the matrixes; coefficient theta is artificially set to [0, 1 ]]A constant of (d); i represents an image to be enhanced; v represents a fused image before fusion; m represents a fused image into which an image to be enhanced is fused; pkRepresenting a background image; v represents the final fused image.
5. The method for detecting the remote sensing image multi-scale target based on the data fusion and the feature selection according to claim 1, wherein the step S3 specifically comprises:
step S3.1: training the training set obtained in the step S2 by using a FoveaBox network, and obtaining a plurality of feature maps with different scales by using the training pictures in the training set through an FPN feature extraction network;
step S3.2: each feature map respectively selects a plurality of targets in different bounding box area ranges for prediction, so that the feature maps in different scales can predict the targets in different sizes, and one target can be predicted on the plurality of feature maps according to the area;
step S3.3: the feature diagram is trained through a classification branch and a boundary box branch, the classification branch is trained by using a FocalLoss Loss function, and the boundary box branch is trained by using a SmoothL1Loss Loss function; after the training is finished, the classification branch and the boundary box branch of the model respectively predict the class and the boundary box of the target in the feature map to obtain a final output result.
6. The method for detecting the remote sensing image multi-scale target based on the data fusion and the feature selection as claimed in claim 5, characterized in that in the process of the bounding box branch training, an area smaller than a real bounding box is adopted for training.
7. The method for detecting the remote sensing image multi-scale target based on the data fusion and the feature selection as claimed in claim 5, wherein in step S3.1, 5 feature maps with different scales are obtained from training pictures in a training set through an FPN feature extraction network.
8. The method for detecting the multi-scale target of the remote sensing image based on the data fusion and the feature selection as claimed in claim 7, wherein in step S3.2, 5 feature maps with different scales are respectively selected to predict the target in the (1,64), (32,128), (64,256), (128,512), (256,2048)) bounding box pixel area range.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110766948.8A CN113449666A (en) | 2021-07-07 | 2021-07-07 | Remote sensing image multi-scale target detection method based on data fusion and feature selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110766948.8A CN113449666A (en) | 2021-07-07 | 2021-07-07 | Remote sensing image multi-scale target detection method based on data fusion and feature selection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113449666A true CN113449666A (en) | 2021-09-28 |
Family
ID=77815312
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110766948.8A Pending CN113449666A (en) | 2021-07-07 | 2021-07-07 | Remote sensing image multi-scale target detection method based on data fusion and feature selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113449666A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111368776A (en) * | 2020-03-13 | 2020-07-03 | 长安大学 | High-resolution remote sensing image classification method based on deep ensemble learning |
| CN111563473A (en) * | 2020-05-18 | 2020-08-21 | 电子科技大学 | Remote sensing ship identification method based on dense feature fusion and pixel level attention |
| CN111738114A (en) * | 2020-06-10 | 2020-10-02 | 杭州电子科技大学 | Vehicle target detection method based on anchor-free accurate sampling remote sensing image |
| CN111832615A (en) * | 2020-06-04 | 2020-10-27 | 中国科学院空天信息创新研究院 | Sample expansion method and system based on foreground and background feature fusion |
| US20210027056A1 (en) * | 2019-07-23 | 2021-01-28 | Regents Of The University Of Minnesota | Remote-sensing-based detection of soybean aphid induced stress in soybean |
| CN112348765A (en) * | 2020-10-23 | 2021-02-09 | 深圳市优必选科技股份有限公司 | Data enhancement method and device, computer readable storage medium and terminal equipment |
-
2021
- 2021-07-07 CN CN202110766948.8A patent/CN113449666A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210027056A1 (en) * | 2019-07-23 | 2021-01-28 | Regents Of The University Of Minnesota | Remote-sensing-based detection of soybean aphid induced stress in soybean |
| CN111368776A (en) * | 2020-03-13 | 2020-07-03 | 长安大学 | High-resolution remote sensing image classification method based on deep ensemble learning |
| CN111563473A (en) * | 2020-05-18 | 2020-08-21 | 电子科技大学 | Remote sensing ship identification method based on dense feature fusion and pixel level attention |
| CN111832615A (en) * | 2020-06-04 | 2020-10-27 | 中国科学院空天信息创新研究院 | Sample expansion method and system based on foreground and background feature fusion |
| CN111738114A (en) * | 2020-06-10 | 2020-10-02 | 杭州电子科技大学 | Vehicle target detection method based on anchor-free accurate sampling remote sensing image |
| CN112348765A (en) * | 2020-10-23 | 2021-02-09 | 深圳市优必选科技股份有限公司 | Data enhancement method and device, computer readable storage medium and terminal equipment |
Non-Patent Citations (2)
| Title |
|---|
| TAO KONG ET AL.: "《FoveaBox: Beyound Anchor-Based Object Detection》", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》, vol. 29, 23 June 2020 (2020-06-23), pages 1 - 10 * |
| 陆保国 等: "《光学遥感影像飞机目标识别与分类方法》", 《指挥信息系统与技术》, vol. 11, no. 5, 31 October 2020 (2020-10-31) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111126472B (en) | SSD (solid State disk) -based improved target detection method | |
| CN111126202B (en) | Optical remote sensing image target detection method based on void feature pyramid network | |
| CN108288075A (en) | A kind of lightweight small target detecting method improving SSD | |
| CN109376603A (en) | A kind of video frequency identifying method, device, computer equipment and storage medium | |
| CN106504233A (en) | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN | |
| CN111353544B (en) | Improved Mixed Pooling-YOLOV 3-based target detection method | |
| CN112560675B (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy | |
| CN110738160A (en) | human face quality evaluation method combining with human face detection | |
| CN109934081A (en) | A kind of pedestrian's attribute recognition approach, device and storage medium based on deep neural network | |
| CN112488229A (en) | Domain self-adaptive unsupervised target detection method based on feature separation and alignment | |
| CN115457415A (en) | Target detection method and device based on YOLO-X model, electronic equipment and storage medium | |
| CN114897802A (en) | Metal surface defect detection method based on improved fast RCNN algorithm | |
| CN114331946A (en) | Image data processing method, device and medium | |
| CN111967399A (en) | Improved fast RCNN behavior identification method | |
| CN112819008B (en) | Method, device, medium and electronic equipment for optimizing instance detection network | |
| CN113962900A (en) | Method, device, equipment and medium for detecting infrared dim target under complex background | |
| CN113095404A (en) | X-ray contraband detection method based on front and back background partial convolution neural network | |
| CN117495718A (en) | Multi-scale self-adaptive remote sensing image defogging method | |
| Funt et al. | Does colour really matter? Evaluation via object classification | |
| CN110033443B (en) | Display panel defect detection method | |
| CN114494893B (en) | Remote sensing image feature extraction method based on semantic reuse context feature pyramid | |
| CN113449666A (en) | Remote sensing image multi-scale target detection method based on data fusion and feature selection | |
| Zhao et al. | Object detector based on enhanced multi-scale feature fusion pyramid network | |
| CN116385876A (en) | Optical remote sensing image ground object detection method based on YOLOX | |
| CN116310688A (en) | Target detection model based on cascade fusion, and construction method, device and application thereof |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210928 |