CN110135522B - Intelligent method for detecting and marking small target integration of remote sensing image - Google Patents
Intelligent method for detecting and marking small target integration of remote sensing image Download PDFInfo
- Publication number
- CN110135522B CN110135522B CN201910451869.0A CN201910451869A CN110135522B CN 110135522 B CN110135522 B CN 110135522B CN 201910451869 A CN201910451869 A CN 201910451869A CN 110135522 B CN110135522 B CN 110135522B
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- image
- sensing image
- target
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses an intelligent method for detecting and marking small targets of remote sensing images, and belongs to the technical field of small target detection and marking of high-resolution remote sensing images. The method of the invention comprises the following steps: self-adaptive image enhancement is carried out on the remote sensing image set marked with the target, and the method can carry out data enhancement and sample expansion; inputting the image set subjected to self-adaptive image enhancement into a target detection framework based on a fast convolutional neural network for target detection; after multiple training iterations, obtaining the model of the invention; and finally, migrating the remote sensing image to be marked to a remote sensing image set through a migration learning method. The method realizes the accurate detection and marking of the small target in the remote sensing image data through the convolutional neural network and the transfer learning method, is suitable for various processing environments, and has better anti-interference capability on complex environments.
Description
Technical Field
The invention relates to an intelligent method for detecting and marking small targets of remote sensing images, belonging to the technical field of small target detection and marking of high-resolution remote sensing images.
Background
The remote sensing image obtained by the modern remote sensing technology has high speed and high image quality, so the method is widely applied to the fields of battlefield environment simulation, hong Tao disaster prediction, military target reconnaissance, urban land planning, forest fire danger monitoring and the like. The labeling of the remote sensing image has become an important research direction for interpretation and analysis of the remote sensing image, and is concerned and researched in various aspects. With the increasing spatial resolution of remote sensing images and the increasing of remote sensing image platforms, the number of images is multiplied, and the information contained in the images is more and more abundant. How to accurately and quickly identify the target and mark the position of the target becomes the key point and difficulty of interpretation and analysis work of the remote sensing image.
In recent years, the deep learning method is different military processes driven by big data, and the method represented by a convolutional neural network is revolutionarily developed in the image field. Many scholars also apply convolutional neural networks to the identification, target detection and other directions in the field of remote sensing images. However, directions such as recognition and target detection based on the convolutional neural network require a large number of images with artificial labels for training, that is, what is the target and the position information of the target. In the remote sensing image, because the target of interest is small, the manual labeling cost is too high, and the convolution neural network can be subjected to overfitting. In addition, even if the remote sensing image data set is artificially labeled, due to the limited number of labels, the convolution neural network is overfitting in the training process. Aiming at the problem, the invention provides an intelligent method for automatically labeling small targets of remote sensing images
Disclosure of Invention
The invention solves the technical problems that: a method for automatically labeling a small target of a remote sensing image is provided. In order to solve the technical problems of small target labeling and detection, the technical scheme provided by the invention is to provide a remote sensing image small target automatic labeling method based on an image self-adaption enhancement strategy, a non-maximum suppression reset function and a transfer learning network model strategy on the basis of a convolutional neural network. The specific content is as follows:
And 2, step: inputting the image set subjected to self-adaptive image enhancement into a target detection framework based on a fast convolutional neural network for target detection; s1, using a network model of ImageNet trained on a VGG16 network as an initial weight to perform migration learning fine tuning; s2, selecting the size of a candidate frame anchor of a candidate area network in the convolutional neural network by adopting the semi-supervision method provided by the invention; and S3, selecting a candidate area in the last layer of the convolution layer of the VGG16 network, wherein after the area of interest is selected, the detection process enters a post-processing stage of detection, and in the post-processing stage, the candidate area is selected by adopting an improved non-maximum suppression algorithm.
And 3, step 3: and (3) carrying out the image self-adaptive enhancement on the remote sensing image set to be trained in the step (1), inputting the image into the convolutional neural network in the step (2) for training iteration to obtain a small target detection and labeling model.
And 4, step 4: and after obtaining the small target detection and labeling model, transferring the small target detection and labeling model to a remote sensing image set to be labeled by a transfer learning method.
Drawings
The invention will be further explained with reference to the drawings.
FIG. 1 is an overall flow chart of a training remote sensing image dataset employed by the present invention based on a convolutional neural network.
FIG. 2a is an original drawing;
FIG. 2b is a schematic illustration of a 60 degree rotation through adaptive image enhancement as described for an image of the present invention;
FIG. 2c is a schematic illustration of a 90 degree rotation through adaptive image enhancement as described for an image of the present invention;
FIG. 2d is a schematic representation of a 150 degree rotation through adaptive image enhancement as described for the images of the present invention;
FIG. 3 is a flow chart of a target detection framework using a VGG16 convolutional neural network;
FIG. 4a is an aircraft target in a remote sensing image scene detected using the method of the present invention;
FIG. 4b is a ship target in a remote sensing image scene detected using the method of the present invention;
FIG. 4c is a diagram of a ship and an oil drum in a remote sensing image scene detected using the method of the present invention;
FIG. 4d is a diagram of a very small target ship in a remote sensing image scene detected using the method of the present invention;
FIG. 5 is a diagram of the present invention for automatically labeling and detecting small targets in high resolution remote sensing images.
Detailed Description
The invention aims to provide a solution for the problems of low small target detection accuracy and less small target labeling data. The specific implementation of the invention can be seen in fig. 1. As shown in fig. 1, a remote sensing data set a with a small target is subjected to image adaptive enhancement, and then enters a detection framework based on a convolutional neural network, and then in a post-processing stage, the improved NMS (Non-Maximum Suppression) algorithm described in the present invention is used to perform the last check box cut, and finally the detection network model of the remote sensing data set a is obtained. And then, the detection network model of the data set A obtained by the invention is used for marking the remote sensing data set to be detected, and simultaneously, the remote sensing data set to be detected is input into a detection frame based on a convolutional neural network, and the target of the remote sensing data set to be detected and the accurate position of the target can be automatically marked because the detection network model of the data set A is supported by the weight.
The detailed implementation steps are as follows:
step 1: carrying out self-adaptive image enhancement on the labeled remote sensing image set;
the method comprises the steps of firstly, storing a remote sensing image set with marked targets, marking an object in each image in the image set, and marking the position of the object in the image, wherein the marking of the position adopts a marking method of a horizontal and vertical coordinate of the upper left corner and a horizontal and vertical coordinate of the lower right corner. As shown in (1) in fig. 1, the remote sensing image set labeled with the target is subjected to data enhancement and sample expansion by the adaptive image enhancement method of the invention.
The invention provides a self-adaptive image enhancement method, which specifically introduces the following steps: the class of the labeled remote sensing image dataset and the number of each class are first calculated (where the target data class generally includes, but is not limited to, aircraft, ship, oil drum, etc.). Assuming that a certain remote sensing image data set has {1,2,3, a 1 ,c 2 ,c 3 ,...,c n ,...,c m Calculating to obtain the labeling target contained in the nth class with the maximum value of c n . Comparing the number of maximum classes with the number of each class, the following sequence is obtained
The method for self-adaptive image enhancement adopts an image rotation mode: in the invention, a rotating dictionary dict = [180 °,150 °,120 °,90 °,60 °,30 ° °is created](dictionary internal numbering starts from 1 to 6), when ratio in any ratio sequence i ∈(0,2]Or ratio i ∈(0,1]Or ratio i ∈(1,2]Then, the image of this category is rotated by 180 degrees, i.e. the fact [1 ] is taken]. The rotation may be clockwise or counterclockwise. When ratio is i ∈(2,3]For the image of this category, respectively, perform dct [1 ]]And dct [2 ]]. When ratio is i ∈(3,4]Separately, perform dit [1 ] on the images of this category]，dict[2]，dict[3]And (4) rotating. When ratio is i ∈(4,5]For the image of this category, respectively, perform dct [1 ]]，dict[2]，dict[3]，dict[4]And (4) rotating. When ratio is i ∈(5,6]For the images in this category, the first 5 angular rotations in the dit dictionary are taken respectively. When ratio is i >And 6, rotating all the images of the category according to the angles in the fact dictionary. Meanwhile, because the length (w) and the width (h) of the image labeling frame are angularly changed compared with the original image after the rotation, the length and the width of the image labeling frame after the rotation are calculated by using the following formula in consideration of the change. W * ＝h*sin(θ)+w*cos(θ)，H * (= h cos (θ) + W sin (θ)), where W is * And H * The length and width after rotation are respectively shown, w and h are respectively shown as the length and width of the original image, and theta is the angle of rotation. The self-adaptive image enhancement method not only greatly enhances the remote sensing image set, but also solves the problem of imbalance among sample classes in the image set.
After the image set is subjected to the adaptive image enhancement method of the present invention, the number of test samples is greatly enhanced, and fig. 2b, fig. 2c, and fig. 2d show the 60-degree, 90-degree, and 150-degree rotated images of the image adaptive enhancement.
And 2, step: and inputting the image set subjected to the self-adaptive image enhancement into a target detection framework based on a fast convolutional neural network for target detection.
The convolutional neural network adopted by the invention is VGG16, the detailed flow of the target detection framework is shown in FIG. 3, and the detection flow is specifically described as follows:
the migratory learning tuning is first performed using the network model of ImageNet trained on the VGG16 network as initial weights. Then, the semi-supervised method provided by the invention selects the size of a candidate frame anchor of a candidate area network in the convolutional neural network. The specific description is as follows:
firstly, estimating by adopting a manual marking method; and then carrying out dimension clustering by adopting a K-nearest neighbor method, further correcting the size of the candidate box anchor, carrying out ground truth clustering by using a data set label by adopting a K-means clustering method, and carrying out size clustering by adopting a cross-over ratio (IoU). The actual metrology size can thus be given by the following equation:
d(prebox,boxcenter)＝1-IOU(prebox,boxcenter)
wherein prebox refers to the size of the manually selected candidate box anchor, and boxcenter refers to the center point of the group channel rectangular box. Through the clustering of the K-nearest neighbor method, a more accurate target detection frame can be selected.
Finally, at the last layer of the convolutional layer of the VGG16 network, the selection of the candidate region is entered, after the selection of the region of interest, the detection process enters a post-processing stage of detection, and in the post-processing stage, the present invention adopts the improved selection of the candidate region by the NMS, which is more favorable for detecting the small target, as shown in (2) in fig. 1.
The improved non-maximum suppression (NMS) algorithm formula in the invention is as follows:
b i represents a set of a series of detection boxes, B = { B1, B2,. Bi, … bn }, s i Indicates the corresponding detection box b i The fraction of (c). M is a box corresponding to the highest score, and iou represents the overlapping rate of the detection box and M. Where σ is the attenuation parameter. N is a radical of t Is a hard threshold, where N t When =0.7 and σ =0.6, the small target detection effect is best.
The improved non-maximum suppression algorithm is widely applied to the post-processing stage of target detection, and can eliminate detection frames which are generated by algorithms and overlapped in different degrees, but the existing non-maximum suppression algorithm belongs to a greedy algorithm and has the defect that if an object is within a preset overlap threshold value, the object to be detected, especially a small object, can not be detected. The present invention provides an improved algorithm for non-maxima suppression that avoids the above-mentioned drawbacks.
And 3, carrying out the image self-adaptive enhancement on the remote sensing image set to be trained in the step 1, inputting the image into the convolutional neural network in the step 2 for training iteration to obtain a small target detection and labeling model.
And (3) carrying out the image self-adaptive enhancement on the remote sensing image set to be trained in the step (1), inputting the image into the convolutional neural network in the step (2) for training iteration to obtain a small target detection and labeling model.
According to the invention, after about 60000 times of training iteration, a better model is obtained, and the model can well detect a small target in the remote sensing image. As shown in part (3) of fig. 1. The invention uses the model for detecting small targets in the remote sensing image, such as: the airplane, ship and the like can obtain good detection results, and the detection results are the airplane target in the detected remote sensing image scene as shown in FIG. 4 a; as shown in fig. 4b, a ship target in the detected remote sensing image scene; as shown in fig. 4c, the detected ship and oil drum in the remote sensing image scene is dense, and the present invention can also exhibit good performance; a tiny target ship in the sensed remote sensing image scene is shown in fig. 4 d.
The loss function employed by the convolutional neural network in the present invention is as follows:
wherein i represents the sequence of anchor, p i Representing the probability of the foreground at the time of prediction,
the predicted probability of a group truth is shown. t is t i Represents a prediction frame, <' > or>
And a ground route box corresponding to the anchor representing the corresponding foreground. Where the parameter lambda is a balance of the loss of the identified class and the regression loss. The regression loss here uses the following calculation formula:
and 4, after obtaining the small target detection and labeling model, transferring the small target detection and labeling model to a remote sensing image set to be labeled by a transfer learning method.
The goal of transfer learning is to use knowledge learned from one environment to assist in the learning task in a new environment. Most studies have adopted transfer learning for parameter tuning and the like.
The condition of transfer learning is to give a source domain D S Learning task T on the source domain S Target domain D T And a learning task T on the target domain T (ii) a The goal of transfer learning is to utilize D S And T S Learning a prediction function f (-) over the target domain; with the proviso that D S ≠D T ，T S ≠T T . The invention adopts two transfer learning methods on the basis of the convolutional neural network, and the target detection of the remote sensing data set is labeling. First, the invention employs deepAnd the degree migration method is used for training the remote sensing data set by using the ImageNet model trained on the VGG16 to carry out fine tuning, and the weight is fine tuned. This saves a lot of time and computer resources. Secondly, the model which is trained on a large remote sensing data set and has the capability of detecting the small target is adopted, the model is directly applied to the small target remote sensing data set with small data volume by adopting a model migration method of migration learning, and the target on the data set and the position of the target can be automatically marked.
According to the invention, an unmarked remote sensing data set is input again, the network model obtained by training in (3) in fig. 1 is used, and the network model is input into the automatic detection system based on the convolutional neural network, and after passing through the system, the automatic detection system based on the convolutional neural network can well mark the target of the unmarked remote sensing data set and the accurate position of the target, as shown in fig. 5, the automatic marking and the detected positions of the airplane target and the airplane target are shown in the invention. The method can greatly reduce the loss of manpower, material resources and time of manual annotation, and is meaningful work for the field of the remote sensing images.
Compared with the prior art, the invention has the following advantages:
firstly, the method comprises the following steps: the data enhancement of the invention is self-adaptive, can well balance the difference between object classes, and improve the average detection rate of small target detection in remote sensing data.
Secondly, the method comprises the following steps: the invention adopts the K-nearest neighbor dimension clustering and semi-supervised adjustment of the size of the rectangular frame Anchor, so that the position accuracy of the detected target can be improved.
Thirdly, the steps of: the invention provides an improved non-maximum suppression method which can improve the accuracy of small target detection in remote sensing images.
Fourthly: because manual labeling of the optical remote sensing image is time-consuming and labor-consuming, and an image sample set with small target labeling is lacking, the invention provides a method for automatically labeling the remote sensing image based on transfer learning and an improved non-maximum suppression method.
Claims (6)
1. An intelligent method for detecting and marking small targets of remote sensing images is characterized by comprising the following steps:
step 1: carrying out self-adaptive image enhancement on the labeled remote sensing image set;
firstly, a remote sensing image set with a marked target exists, the remote sensing image set with the marked target is subjected to data enhancement by using a self-adaptive image enhancement method, and a sample is expanded;
the self-adaptive image enhancement method specifically comprises the following steps:
a) Firstly, the category of the labeled remote sensing image data set and the quantity of each category are calculated, a certain remote sensing image data set is set to have {1,2,3, a 1 ,c 2 ,c 3 ,...,c n ,...,c m Calculating to obtain the labeling target contained in the nth class with the maximum value of c n Comparing the number of maximum classes with the number of each class, the following sequence is obtained
b) Creating a rotating dictionary, dit = [180 °,150 °,120 °,90 °,60 °,30 ° ], corresponding to the internal numbers of the dictionary starting from 1 to 6:
when ratio in any ratio sequence i ∈(0,2]The image in this category is rotated by 180 degrees, i.e. the fact [1 ] is taken]；
When ratio is i ∈(2,3]The images in this category are rotated by 180 and 150 degrees, respectively, i.e. the fact [1 ] is taken]And dct [2 ]]；
When ratio is i ∈(3,4]The images in this category are rotated by 180 degrees, 150 degrees and 120 degrees respectively, i.e. the fact [1 ] is taken]，dict[2]，dict[3]；
When ratio is i ∈(4,5]The images of this category are processed at 180 degrees and 150 degrees respectivelyRotation of 120 deg. and 90 deg. to obtain dit [1 ]]，dict[2]，dict[3]，dict[4]；
When ratio is i ∈(5,6]Respectively taking the first 5 angular rotations in the fact dictionary for the images of the category;
when ratio is i >6, rotating all the images in the category according to the angles in the fact dictionary;
c) Meanwhile, after the rotation, the length w and the width h of the image labeling frame are changed in an angle manner compared with the original image, and the length and the width of the image labeling frame after the rotation are calculated by using the following formula:
W * ＝h*sin(θ)+w*cos(θ)，
H * ＝h*cos(θ)+w*sin(θ)，
wherein W * And H * Respectively representing the length and width after rotation, w and h respectively representing the length and width of the original image, and theta represents the rotation angle;
step 2: inputting the image set subjected to self-adaptive image enhancement into a target detection framework based on a fast convolutional neural network for target detection;
and step 3: carrying out the image self-adaptive enhancement on a remote sensing image set to be trained in the step 1, and inputting the image self-adaptive enhancement into the convolutional neural network in the step 2 for training iteration to obtain a small target detection and labeling model;
and 4, step 4: and after the small target detection and labeling model is obtained, the small target detection and labeling model is migrated to a remote sensing image set to be labeled by a migration learning method.
2. The intelligent method for detecting and labeling the integration of the small target of the remote sensing image according to claim 1, which is characterized in that: the rotation described in step 1 is either clockwise rotation or counterclockwise rotation.
3. The intelligent method for detecting and labeling the integration of the small target of the remote sensing image according to claim 1, which is characterized in that: the convolutional neural network in the step 2 is VGG16, and the flow of the target detection framework is specifically as follows:
s1, using a network model of ImageNet trained on a VGG16 network as an initial weight to perform migration learning fine tuning;
s2, selecting the size of a candidate frame anchor of a candidate area network in the convolutional neural network by adopting a semi-supervision method, wherein the semi-supervision method is specifically described as follows:
a) Firstly, estimating by adopting a manual marking method;
b) Then, carrying out dimension clustering by adopting a K-nearest neighbor method, further correcting the size of the candidate frame anchor, wherein the K-means clustering method uses a group route clustering marked by a data set, and carries out size clustering by adopting a cross-over ratio IoU; the actual metrology size can thus be given by the following equation:
d(prebox,boxcenter)＝1-IOU(prebox,boxcenter)
the prebox refers to the size of a candidate frame anchor selected manually, the boxcenter refers to the central point of a ground channel rectangular frame, and more accurate target detection frames can be selected by clustering through a K-nearest neighbor method;
s3, selecting a candidate area in the last layer of the convolution layer of the VGG16 network, wherein after the region of interest is selected, the detection process enters a post-processing stage of detection, and the candidate area is selected by adopting an improved non-maximum suppression algorithm in the post-processing stage;
the formula of the improved non-maximum suppression algorithm is as follows:
b i represents a set of a series of detection boxes, B = { B1, B2,. Bi, … bn }, s = { B1, B2 }, bi, … bn }, s i Indicates the corresponding detection box b i A fraction of (d); m is a frame corresponding to the highest score, iou represents the overlapping rate of the detection frame and M, and sigma is an attenuation parameter; n is a radical of t Is a hard threshold.
4. The intelligent method for detecting and annotating small targets in remote sensing images according to claim 3, characterized in that: n is a radical of t When the value is 0.7 and the value is 0.6, the small target detection effect is best.
5. The intelligent method for detecting and labeling the integration of the small target of the remote sensing image according to claim 1, which is characterized in that: the model in the step 3 is obtained after at least about 60000 times of training iteration, and the model can well detect the small target in the remote sensing image.
6. The intelligent method for detecting and annotating small targets in remote sensing images according to claim 1, characterized in that: the loss function adopted by the convolutional neural network is as follows:
wherein i represents the sequence of anchor, p i Representing the probability of the foreground at the time of prediction,
the prediction probability of the group channel is shown; t is t i Represents a prediction frame, <' > or>
A group channel box corresponding to the anchor representing the corresponding foreground; wherein the parameter λ is a loss of the balanced recognition class and a regression loss; the regression loss here uses the following calculation formula:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910451869.0A CN110135522B (en) | 2019-05-28 | 2019-05-28 | Intelligent method for detecting and marking small target integration of remote sensing image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910451869.0A CN110135522B (en) | 2019-05-28 | 2019-05-28 | Intelligent method for detecting and marking small target integration of remote sensing image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110135522A CN110135522A (en) | 2019-08-16 |
| CN110135522B true CN110135522B (en) | 2023-03-28 |
Family
ID=67582323
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910451869.0A Active CN110135522B (en) | 2019-05-28 | 2019-05-28 | Intelligent method for detecting and marking small target integration of remote sensing image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110135522B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110866476B (en) * | 2019-11-06 | 2023-09-01 | 南京信息职业技术学院 | Dense stacking target detection method based on automatic labeling and transfer learning |
| CN111222462A (en) * | 2020-01-07 | 2020-06-02 | 河海大学 | Target detection-based intelligent labeling method for apparent feature monitoring data |
| CN111383267B (en) * | 2020-03-03 | 2024-04-05 | 重庆金山医疗技术研究院有限公司 | Target repositioning method, device and storage medium |
| CN111580947B (en) * | 2020-04-29 | 2021-10-01 | 中国科学院空天信息创新研究院 | Online collaborative remote sensing image annotation system based on artificial intelligence |
| CN111666983A (en) * | 2020-05-19 | 2020-09-15 | 华中科技大学鄂州工业技术研究院 | Method and device for marking abnormal behaviors |
| CN111797737A (en) * | 2020-06-22 | 2020-10-20 | 重庆高新区飞马创新研究院 | Remote sensing target detection method and device |
| CN111798435A (en) * | 2020-07-08 | 2020-10-20 | 国网山东省电力公司东营供电公司 | Image processing method, and method and system for monitoring invasion of engineering vehicle into power transmission line |
| CN112329721B (en) * | 2020-11-26 | 2023-04-25 | 上海电力大学 | Remote sensing small target detection method for model lightweight design |
| CN113496256B (en) * | 2021-06-24 | 2024-04-09 | 中汽创智科技有限公司 | Image annotation model training method, annotation method, device, equipment and medium |
| WO2023236008A1 (en) * | 2022-06-06 | 2023-12-14 | Intel Corporation | Methods and apparatus for small object detection in images and videos |
| CN116051985B (en) * | 2022-12-20 | 2023-06-23 | 中国科学院空天信息创新研究院 | Semi-supervised remote sensing target detection method based on multi-model mutual feedback learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107657279A (en) * | 2017-09-26 | 2018-02-02 | 中国科学院大学 | A kind of remote sensing target detection method based on a small amount of sample |
| CN108052940A (en) * | 2017-12-17 | 2018-05-18 | 南京理工大学 | SAR remote sensing images waterborne target detection methods based on deep learning |
| CN109063594A (en) * | 2018-07-13 | 2018-12-21 | 吉林大学 | Remote sensing images fast target detection method based on YOLOv2 |
| CN109147254A (en) * | 2018-07-18 | 2019-01-04 | 武汉大学 | A kind of video outdoor fire disaster smog real-time detection method based on convolutional neural networks |
| CN109359557A (en) * | 2018-09-25 | 2019-02-19 | 东北大学 | A kind of SAR remote sensing images Ship Detection based on transfer learning |
-
2019
- 2019-05-28 CN CN201910451869.0A patent/CN110135522B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107657279A (en) * | 2017-09-26 | 2018-02-02 | 中国科学院大学 | A kind of remote sensing target detection method based on a small amount of sample |
| CN108052940A (en) * | 2017-12-17 | 2018-05-18 | 南京理工大学 | SAR remote sensing images waterborne target detection methods based on deep learning |
| CN109063594A (en) * | 2018-07-13 | 2018-12-21 | 吉林大学 | Remote sensing images fast target detection method based on YOLOv2 |
| CN109147254A (en) * | 2018-07-18 | 2019-01-04 | 武汉大学 | A kind of video outdoor fire disaster smog real-time detection method based on convolutional neural networks |
| CN109359557A (en) * | 2018-09-25 | 2019-02-19 | 东北大学 | A kind of SAR remote sensing images Ship Detection based on transfer learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110135522A (en) | 2019-08-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110135522B (en) | Intelligent method for detecting and marking small target integration of remote sensing image | |
| CN109255317B (en) | Aerial image difference detection method based on double networks | |
| CN111507370B (en) | Method and device for obtaining sample image of inspection tag in automatic labeling image | |
| CN110766041B (en) | Deep learning-based pest detection method | |
| CN110889324A (en) | Thermal infrared image target identification method based on YOLO V3 terminal-oriented guidance | |
| CN108960135B (en) | Dense ship target accurate detection method based on high-resolution remote sensing image | |
| CN110837870A (en) | Sonar image target identification method based on active learning | |
| CN108038846A (en) | Transmission line equipment image defect detection method and system based on multilayer convolutional neural networks | |
| Li et al. | Automatic bridge crack identification from concrete surface using ResNeXt with postprocessing | |
| CN104463249B (en) | A kind of remote sensing images airfield detection method based on Weakly supervised learning framework | |
| CN110647802A (en) | Remote sensing image ship target detection method based on deep learning | |
| CN110796679B (en) | Target tracking method for aerial image | |
| CN113139594B (en) | Self-adaptive detection method for airborne image unmanned aerial vehicle target | |
| CN110443279B (en) | Unmanned aerial vehicle image vehicle detection method based on lightweight neural network | |
| CN113962960A (en) | Pavement disease detection method based on deep learning | |
| CN107909053B (en) | Face detection method based on hierarchical learning cascade convolution neural network | |
| CN108229524A (en) | A kind of chimney and condensing tower detection method based on remote sensing images | |
| CN111798417A (en) | SSD-based remote sensing image target detection method and device | |
| CN108133235A (en) | A kind of pedestrian detection method based on neural network Analysis On Multi-scale Features figure | |
| CN112329559A (en) | Method for detecting homestead target based on deep convolutional neural network | |
| CN108446613A (en) | A kind of pedestrian's recognition methods again based on distance centerization and projection vector study | |
| CN115147418B (en) | Compression training method and device for defect detection model | |
| CN111985325A (en) | Aerial small target rapid identification method in extra-high voltage environment evaluation | |
| CN115690627A (en) | Method and system for detecting aerial image rotating target | |
| CN115100497A (en) | Robot-based method, device, equipment and medium for routing inspection of abnormal objects in channel |
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 |