CN112016512A - Remote sensing image small target detection method based on feedback type multi-scale training - Google Patents
Remote sensing image small target detection method based on feedback type multi-scale training Download PDFInfo
- Publication number
- CN112016512A CN112016512A CN202010934966.8A CN202010934966A CN112016512A CN 112016512 A CN112016512 A CN 112016512A CN 202010934966 A CN202010934966 A CN 202010934966A CN 112016512 A CN112016512 A CN 112016512A
- Authority
- CN
- China
- Prior art keywords
- module
- small target
- remote sensing
- scale
- type multi
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 45
- 238000012549 training Methods 0.000 title claims abstract description 45
- 238000000034 method Methods 0.000 claims abstract description 16
- 238000013528 artificial neural network Methods 0.000 claims abstract description 9
- 230000008569 process Effects 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims description 23
- 238000011176 pooling Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 6
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 230000006872 improvement Effects 0.000 description 4
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 3
- 230000001965 increasing effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012937 correction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/20—Scenes; Scene-specific elements in augmented reality scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a remote sensing image small target detection method based on feedback type multi-scale training, which comprises the following steps: constructing a feedback type multi-scale convolution neural network formed by a detection module and a feedback multi-scale training module, inputting original image data and training the original image data in an end-to-end mode; the feedback multi-scale training module calculates a proportion value of the small target according to the loss of the current iteration process output by the detection module; comparing the calculated ratio value of the small target with a preset threshold, and when the ratio value of the small target is smaller than the preset threshold, using the spliced image data as the input of the next iteration, or else, using the original image data as the input; and obtaining a trained feedback type multi-scale convolution neural network, inputting the remote sensing image to be detected, and outputting a recognition result. The detection capability of the small targets in the remote sensing image is enhanced, the overfitting phenomenon and the class imbalance phenomenon are inhibited, and the small target detection method has better effect and robustness on the small target detection in the remote sensing image.
Description
Technical Field
The invention relates to the technical field of remote sensing image target detection, in particular to remote sensing image target detection by utilizing a neural network model, and specifically relates to a remote sensing image small target detection method based on feedback type multi-scale training.
Background
The remote sensing image target detection problem is always an important research direction in the field of computer vision and pattern recognition research. Although a plurality of deep learning algorithms based on target detection appear in recent years and bring strong improvement to target detection, aiming at the detection of small targets in remote sensing images, a plurality of places which need improvement exist.
The model performance is degraded due to the possibility of introducing more false positives due to the complex background in the remote sensing image. At the same time, the detection performance of small objects (less than 32 x 32 pixels) in the remote sensing image is the least satisfactory.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a remote sensing image small target detection method based on feedback type multi-scale training.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a remote sensing image small target detection method based on feedback type multi-scale training is characterized by comprising the following steps:
and 4, obtaining the trained feedback type multi-scale convolution neural network, inputting the remote sensing image to be detected, and outputting a recognition result.
Further, the detection module is built by the following steps:
step 11, constructing a feature extraction network, wherein the feature extraction network is used for extracting high-level semantic features and low-level semantic features of an input image, normalizing the size through upsampling, and obtaining an added feature map through average pooling;
step 12, constructing a prior frame acquisition module, wherein the prior frame acquisition module is used for extracting a feature map output by a network according to features and generating a prior frame by adopting a prior frame strategy;
step 13, constructing a classification task module, wherein the classification task module is used for classifying the target classes of the prior frames;
step 14, constructing a regression task module, wherein the regression task module is used for positioning the target frame;
and step 15, constructing a network optimization module, wherein the network optimization module is used for optimizing the network by utilizing the multitask loss function.
Further, the feature extraction network is built based on a residual error structure and comprises a first residual error module, a second residual error module and a third residual error module, the first residual error module, the second residual error module and the third residual error module successively perform down-sampling feature extraction on an input image, and feature extraction results output by the first residual error module, the second residual error module and the third residual error module are subjected to feature fusion to obtain the feature map.
Further, the first residual module, the second residual module, and the third residual module each include three convolution groups, a maximum pooling layer, and a convolutional layer, where a sliding window of the maximum pooling layer is 3 × 3, a step size is 2, a convolutional kernel size of the convolutional layer is 1 × 1, and a channel number is 128.
Further, the specific process of the prior frame acquisition module in step 12 using the prior frame strategy to generate the prior frame is as follows:
a1, acquiring central points of a feature map output by a feature extraction network;
step A2, setting the size and the length-width ratio of each prior frame based on the small target size;
and A3, obtaining a prior frame of the feature map according to a prior frame strategy.
Further, the functional formula of the multitask loss function is as follows:
wherein L is(x,c,p,g)For multi-tasking loss, N is the number of positive samples of the prior frame; l isconf(x, c) confidence error loss, x ∈ {1,0} being an indication parameter; c is a category confidence degree predicted value; l isloc(x, p, g) is the position error loss, and p is the position predicted value of the boundary frame corresponding to the prior frame; g is a position parameter of the real label; and alpha is a weight coefficient.
Further, the calculation formula of the proportional value of the small target is as follows:
wherein,is a proportional value of a small target, LtIs a position error Lloc(x,p,g),Is no greater than AsSmall target position error.
Further, the step of acquiring the stitched image data is as follows:
step 31, storing the original image data in a cache form;
step 32, obtaining four images in the current iteration image, scaling the size of each image to 1/2 of the original size through size conversion, and splicing the four scaled images into one image;
step 33: acquiring a data label of a current iteration image, performing same scale transformation and scaling on the data labels corresponding to the four scaled images, and storing the data labels into a new label;
and step 34, taking the spliced image after the scale transformation and the new label corresponding to the spliced image as spliced image data.
The invention has the following remarkable effects:
1. the invention designs a feedback multi-scale training module which carries out feedback correction by using the loss of network output and determines the image sequence to be input in the next iteration according to the small target proportion value of the current iteration, thereby not only enhancing the detection capability of small targets in a remote sensing image, but also inhibiting the overfitting phenomenon and the class imbalance phenomenon, greatly improving the accuracy of a target detection result and having better effect and robustness on the detection of the small targets in the remote sensing image;
2. the traditional multi-scale training method is generally time-consuming or occupies a large amount of memory, and is only used on a common image, but the method can be applied to remote sensing images, almost does not need to increase redundant memory consumption and time consumption, and is a rapid multi-scale training mode;
3. the method uses a feature extraction network to extract features of an input image, so that an obtained feature map contains feature information of high-level semantics and low-level semantics; and then, a great number of prior frames are generated by utilizing a prior frame strategy to find out a sensitive target area, so that the target recall rate is greatly increased.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a schematic structural diagram of the feedback-type multi-scale convolutional neural network;
FIG. 3 is a block diagram of the feature extraction network;
fig. 4 is a schematic diagram of the conversion of the stitched image and the original image.
Detailed Description
The following provides a more detailed description of the embodiments and the operation of the present invention with reference to the accompanying drawings.
As shown in fig. 1, a method for detecting a small target of a remote sensing image based on feedback-type multi-scale training specifically comprises the following steps:
the detection module is built by the following steps:
step 11, constructing a feature extraction network shown in fig. 3, wherein the feature extraction network is used for extracting high-level semantic features and low-level semantic features of an input image, normalizing the size through upsampling, and obtaining an added feature map through average pooling;
the feature extraction network is built based on a residual error structure and comprises a first residual error module, a second residual error module and a third residual error module, the first residual error module, the second residual error module and the third residual error module successively carry out down-sampling feature extraction on an input image, and feature extraction results output by the first residual error module, the second residual error module and the third residual error module are subjected to feature fusion to obtain the feature map.
As shown in fig. 4, the first Residual block (Residual-1) comprises three convolution groups, and each convolution group forms a Residual structure by using a shortcut connection (linear addition), wherein each convolution group is defined as two convolution networks of [3 × 3,32], 3 × 3 is the convolution kernel size, and 32 represents the number of channels. The third convolution group is followed by a maximal pooling (sliding window is 3 multiplied by 3, step length is 2) and a [1 multiplied by 1,64] convolution network for changing the size of the characteristic diagram, the number of channels and increasing the nonlinear capability of the network, and the characteristic diagram is output at the layer for characteristic fusion;
the second Residual block (Residual-2) contains three convolution groups, each convolution group being defined as two [3 × 3,64] convolution networks. Then, a maximal pooling (sliding window is 3 multiplied by 3, step length is 2) and a [1 multiplied by 1,128] convolution network are connected behind the third convolution group, and a feature map is output at the layer for feature fusion;
the third Residual block (Residual-3) contains three convolution groups, each defined as two convolution networks of [3 × 3,128 ]. Then, a maximal pooling (sliding window is 3 multiplied by 3, step length is 2) and a [1 multiplied by 1,128] convolution network (the number of channels in the layer is unchanged) are connected behind the third convolution group, and a feature map is output in the layer for feature fusion;
the feature maps output by each residual module are subjected to upsampling normalization size, and then are added by using average pooling (a sliding window is 3 multiplied by 3, and the step length is 2), so that the obtained feature maps contain high-level semantic information (containing stronger object feature information but weaker object position information) and low-level semantic information (containing stronger object position information but unobvious object feature information), namely, the fused feature maps contain stronger object position information and object feature information.
Step 12, constructing a prior frame acquisition module, wherein the prior frame acquisition module is used for extracting a feature map output by a network according to features and generating a prior frame by adopting a prior frame strategy;
a1, acquiring central points of a feature map output by a feature extraction network;
that is, an n × n feature map can be obtained through the feature extraction network, and the n × n feature map corresponds to n × n central points, wherein each central point generates a prior frame;
step A2, setting the size and the length-width ratio of each prior frame based on the small target size;
that is, the size of each prior box is set to (8, 16, 32) and the aspect ratio is set to (0.5, 1, 2). The size and length of each prior box can be determined by the target size of a specific data set, and if the size of a small target is generally lower than 32 × 32 pixels, the size of the prior box can be generally set to (8, 16, 32), and the aspect ratio is set to (0.5, 1, 2), so that the real label of each small target can be ensured to correspond to one of the prior boxes. And the parameters of the size or the length-width ratio can be set to be 4 or 5, but if the set parameters are too much, the training and reasoning speed is seriously influenced, so that the parameters are generally defaulted to be 3.
And A3, obtaining a prior frame of the feature map according to a prior frame strategy.
Therefore, the center point of each feature map will generate 9 prior frames (3 x 3), i.e. an n × n feature map will generate 9 × n2A priori block, and thus used for the training task of classifying branches and regressing branches.
Step 13, constructing a classification task module, wherein the classification task module is used for classifying the target classes of the prior frames;
step 14, constructing a regression task module, wherein the regression task module is used for positioning the target frame;
and step 15, constructing a network optimization module, wherein the network optimization module is used for optimizing the network by utilizing the multitask loss function.
The functional formula of the multitask loss function is as follows:
wherein L is(x,c,p,g)For multi-tasking loss, N is the number of positive samples of the prior frame; l isconf(x, c) confidence error loss, x ∈ {1,0} being an indication parameter; c is a category confidence degree predicted value; l isloc(x, p, g) is the position error loss, and p is the position predicted value of the boundary frame corresponding to the prior frame; g is a position parameter of the real label; and alpha is a weight coefficient.
The classification task will produce a confidence error loss Lconf(x, c), the calculation formula is as follows:
Lconf(c,x)=-log[x*c+(1-x)(1-c)],
the regression task will generate a position error Lloc(x, p, g), the calculation formula is as follows:
Lloc(x,p,g)=x*R(p-g),
where R is the Smooth L1 function, which is shown below:
in order to enhance the detection capability of the small target, the invention determines the input of the next iteration according to the feedback value of the current traversal, and the feedback of the current traversal is composed of the loss proportion of the small target.
the calculation process of the proportional value of the small target is as follows:
a) the object area is defined, which is strictly speaking generally determined by its mask region, but this is only available in the segmentation task. For object detection, the area of the object is typically replaced with a box area, as shown in equation (5):
a0≈h0×w0,
wherein h is0,w0Respectively representing the height and width of the object box.
b) Defining small target losses, small target lossesMeans from an area not greater than As(1024 in the COCO protocol), the proportion formula of the small target is:
wherein L istIs a position error Lloc(x,p,g),Is no greater than AsIs detected by the position error of the small target,is expressed as a proportional value of a small target, and
And splicing the image data by utilizing every four images loaded by the current memory, changing the sizes of the images and the corresponding labels, and if each image of the original image is (h, w), changing the image into a spliced image after the image data splicing operation is carried outThe method comprises the following steps:
step 31, storing original image data in a cache form, such as a. npy file or a binary file in numpy, to improve the loading speed;
step 32, firstly unifying the scales of the four images into (h, w), and scaling the size of each image into 1/2 size of the original size by scaling, so that the size of each image after scaling becomes the size of the image after scalingSplicing the four zoomed images into one image, wherein the size of the spliced image is (h, w);
step 33: acquiring data labels (target type and real target frame coordinates) of a current batch, firstly carrying out same scale transformation and scaling on labels corresponding to four loaded images (the real target frame coordinates are transformed, and the target type is unchanged), and then storing the labels corresponding to the four images into a new label to ensure that a generated spliced image contains the labels of the original four images;
and step 34, taking the spliced image after the scale transformation and the new label corresponding to the spliced image as spliced image data.
And 4, obtaining the trained feedback type multi-scale convolution neural network, inputting the remote sensing image to be detected, and outputting a recognition result.
The embodiment provides a method specially aiming at small target detection of remote sensing images, which is implemented by building a feedback type multi-scale convolution neural network (FBMS-Net) composed of a detection module and a feedback multi-scale training module, wherein the detection module comprises a feature extraction network, a priori frame strategy, a classification branch and a target frame regression branch. Firstly, extracting the characteristics of high-level semantics and low-level semantics by adopting a characteristic extraction network; using a priori frame strategy to carry out region proposal and generate a priori frame; performing a classification branch task and a target frame regression branch task in parallel; initializing a feedback multi-scale module to enable the feedback multi-scale module to be in a feedback receiving state (namely, training is carried out by adopting original picture data during initialization training); and finally, outputting the classification loss and the target frame loss by the detection module, wherein the loss is used as the input of the feedback multi-scale training module, and judging whether to activate the feedback multi-scale training module. If the module is in an activated state, the spliced image data is adopted for training; otherwise, the original image data is used for training. By the method, the small target detection in the remote sensing image has better effect and robustness.
The technical solution provided by the present invention is described in detail above. The principles and embodiments of the present invention are explained herein using specific examples, which are presented only to assist in understanding the method and its core concepts. It should be noted that, for those skilled in the art, it is possible to make various improvements and modifications to the present invention without departing from the principle of the present invention, and those improvements and modifications also fall within the scope of the claims of the present invention.
Claims (8)
1. A remote sensing image small target detection method based on feedback type multi-scale training is characterized by comprising the following steps:
step 1, constructing a feedback type multi-scale convolution neural network formed by a detection module and a feedback multi-scale training module, initializing a network weight by using a pre-training model, inputting original image data and training the original image data in an end-to-end mode;
step 2, the feedback multi-scale training module calculates a proportion value of a small target according to the loss of the current iteration process output by the detection module;
step 3, the feedback multi-scale training module compares the calculated ratio value of the small target with a preset threshold, and when the ratio value is smaller than the preset threshold, the spliced image data is used as the input of the next iteration, otherwise, the original image data is used as the input;
and 4, obtaining the trained feedback type multi-scale convolution neural network, inputting the remote sensing image to be detected, and outputting a recognition result.
2. The remote sensing image small target detection method based on feedback type multi-scale training of claim 1, characterized in that: the target detection convolutional neural network is constructed by the following steps:
step 11, constructing a feature extraction network, wherein the feature extraction network is used for extracting high-level semantic features and low-level semantic features of an input image, normalizing the size through upsampling, and obtaining an added feature map through average pooling;
step 12, constructing a prior frame acquisition module, wherein the prior frame acquisition module is used for extracting a feature map output by a network according to features and generating a prior frame by adopting a prior frame strategy;
step 13, constructing a classification task module, wherein the classification task module is used for classifying the target classes of the prior frames;
step 14, constructing a regression task module, wherein the regression task module is used for positioning the target frame;
and step 15, constructing a network optimization module, wherein the network optimization module is used for optimizing the network by utilizing the multitask loss function.
3. The remote sensing image small target detection method based on feedback type multi-scale training of claim 2, characterized in that: the feature extraction network is built based on a residual error structure and comprises a first residual error module, a second residual error module and a third residual error module, the first residual error module, the second residual error module and the third residual error module successively carry out down-sampling feature extraction on an input image, and feature extraction results output by the first residual error module, the second residual error module and the third residual error module are subjected to feature fusion to obtain the feature map.
4. The remote sensing image small target detection method based on feedback type multi-scale training of claim 3, characterized in that: the first residual module, the second residual module and the third residual module respectively comprise three convolution groups, a maximum pooling layer and a convolution layer, wherein the sliding window of the maximum pooling layer is 3 multiplied by 3, the step length is 2, the convolution kernel size of the convolution layer is 1 multiplied by 1, and the number of channels is 128.
5. The remote sensing image small target detection method based on feedback type multi-scale training of claim 2, characterized in that: in step 12, the specific process of the prior frame obtaining module generating the prior frame by using the prior frame strategy is as follows:
a1, acquiring central points of a feature map output by a feature extraction network;
step A2, setting the size and the length-width ratio of each prior frame based on the small target size;
and A3, obtaining a prior frame of the feature map according to a prior frame strategy.
6. The remote sensing image small target detection method based on feedback type multi-scale training of claim 2, characterized in that: the functional formula of the multitask loss function is as follows:
wherein L is(x,c,p,g)For multi-tasking loss, N is the number of positive samples of the prior frame; l isconf(x, c) confidence error loss, x ∈ {1,0} being an indication parameter; c is a category confidence degree predicted value; l isloc(x, p, g) is the position error loss, and p is the position predicted value of the boundary frame corresponding to the prior frame; g is a position parameter of the real label; and alpha is a weight coefficient.
7. The remote sensing image small target detection method based on feedback type multi-scale training of claim 1, characterized in that: the calculation formula of the proportional value of the small target is as follows:
8. The remote sensing image small target detection method based on feedback type multi-scale training of claim 1, characterized in that: the step of acquiring the spliced image data is as follows:
step 31, storing the original image data in a cache form;
step 32, obtaining four images in the current iteration image, scaling the size of each image to 1/2 of the original size through size conversion, and splicing the four scaled images into one image;
step 33: acquiring a data label of a current iteration image, performing same scale transformation and scaling on the data labels corresponding to the four scaled images, and storing the data labels into a new label;
and step 34, taking the spliced image after the scale transformation and the new label corresponding to the spliced image as spliced image data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010934966.8A CN112016512A (en) | 2020-09-08 | 2020-09-08 | Remote sensing image small target detection method based on feedback type multi-scale training |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010934966.8A CN112016512A (en) | 2020-09-08 | 2020-09-08 | Remote sensing image small target detection method based on feedback type multi-scale training |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112016512A true CN112016512A (en) | 2020-12-01 |
Family
ID=73516343
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010934966.8A Pending CN112016512A (en) | 2020-09-08 | 2020-09-08 | Remote sensing image small target detection method based on feedback type multi-scale training |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112016512A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112766184A (en) * | 2021-01-22 | 2021-05-07 | 东南大学 | Remote sensing target detection method based on multi-level feature selection convolutional neural network |
CN113076962A (en) * | 2021-05-14 | 2021-07-06 | 电子科技大学 | Multi-scale target detection method based on micro neural network search technology |
CN113159300A (en) * | 2021-05-15 | 2021-07-23 | 南京逸智网络空间技术创新研究院有限公司 | Image detection neural network model, training method thereof and image detection method |
CN113221855A (en) * | 2021-06-11 | 2021-08-06 | 中国人民解放军陆军炮兵防空兵学院 | Small target detection method and system based on scale sensitive loss and feature fusion |
CN115272814A (en) * | 2022-09-28 | 2022-11-01 | 南昌工学院 | Long-distance space self-adaptive multi-scale small target detection method |
CN115761383A (en) * | 2023-01-06 | 2023-03-07 | 北京匠数科技有限公司 | Image classification method and device, electronic equipment and medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109344821A (en) * | 2018-08-30 | 2019-02-15 | 西安电子科技大学 | Small target detecting method based on Fusion Features and deep learning |
CN109815799A (en) * | 2018-12-18 | 2019-05-28 | 南京理工大学 | A kind of vehicle detecting algorithm of quickly taking photo by plane based on SSD |
CN109977997A (en) * | 2019-02-13 | 2019-07-05 | 中国科学院自动化研究所 | Image object detection and dividing method based on convolutional neural networks fast robust |
CN111079604A (en) * | 2019-12-06 | 2020-04-28 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Method for quickly detecting tiny target facing large-scale remote sensing image |
-
2020
- 2020-09-08 CN CN202010934966.8A patent/CN112016512A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109344821A (en) * | 2018-08-30 | 2019-02-15 | 西安电子科技大学 | Small target detecting method based on Fusion Features and deep learning |
CN109815799A (en) * | 2018-12-18 | 2019-05-28 | 南京理工大学 | A kind of vehicle detecting algorithm of quickly taking photo by plane based on SSD |
CN109977997A (en) * | 2019-02-13 | 2019-07-05 | 中国科学院自动化研究所 | Image object detection and dividing method based on convolutional neural networks fast robust |
CN111079604A (en) * | 2019-12-06 | 2020-04-28 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Method for quickly detecting tiny target facing large-scale remote sensing image |
Non-Patent Citations (3)
Title |
---|
JUNSUO QU 等: "Dilated Convolution and Feature Fusion SSD Network for Small Object Detection in Remote Sensing Images", 《IEEE ACCESS》 * |
YUKANG CHEN 等: "Stitcher: Feedback-driven Data Provider for Object Detection", 《ARXIV》 * |
王冬丽 等: "基于特征融合的SSD视觉小目标检测", 《计算机工程与应用》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112766184A (en) * | 2021-01-22 | 2021-05-07 | 东南大学 | Remote sensing target detection method based on multi-level feature selection convolutional neural network |
CN112766184B (en) * | 2021-01-22 | 2024-04-16 | 东南大学 | Remote sensing target detection method based on multi-level feature selection convolutional neural network |
CN113076962A (en) * | 2021-05-14 | 2021-07-06 | 电子科技大学 | Multi-scale target detection method based on micro neural network search technology |
CN113159300A (en) * | 2021-05-15 | 2021-07-23 | 南京逸智网络空间技术创新研究院有限公司 | Image detection neural network model, training method thereof and image detection method |
CN113159300B (en) * | 2021-05-15 | 2024-02-27 | 南京逸智网络空间技术创新研究院有限公司 | Image detection neural network model, training method thereof and image detection method |
CN113221855A (en) * | 2021-06-11 | 2021-08-06 | 中国人民解放军陆军炮兵防空兵学院 | Small target detection method and system based on scale sensitive loss and feature fusion |
CN115272814A (en) * | 2022-09-28 | 2022-11-01 | 南昌工学院 | Long-distance space self-adaptive multi-scale small target detection method |
CN115272814B (en) * | 2022-09-28 | 2022-12-27 | 南昌工学院 | Long-distance space self-adaptive multi-scale small target detection method |
CN115761383A (en) * | 2023-01-06 | 2023-03-07 | 北京匠数科技有限公司 | Image classification method and device, electronic equipment and medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109299274B (en) | Natural scene text detection method based on full convolution neural network | |
CN112016512A (en) | Remote sensing image small target detection method based on feedback type multi-scale training | |
CN109977918B (en) | Target detection positioning optimization method based on unsupervised domain adaptation | |
CN111126472B (en) | SSD (solid State disk) -based improved target detection method | |
CN111027493B (en) | Pedestrian detection method based on deep learning multi-network soft fusion | |
CN111126359B (en) | High-definition image small target detection method based on self-encoder and YOLO algorithm | |
CN111898432B (en) | Pedestrian detection system and method based on improved YOLOv3 algorithm | |
CN112150493B (en) | Semantic guidance-based screen area detection method in natural scene | |
CN110647829A (en) | Bill text recognition method and system | |
CN111696110B (en) | Scene segmentation method and system | |
CN111914698B (en) | Human body segmentation method, segmentation system, electronic equipment and storage medium in image | |
CN112287941B (en) | License plate recognition method based on automatic character region perception | |
CN113673338A (en) | Natural scene text image character pixel weak supervision automatic labeling method, system and medium | |
CN111160407A (en) | Deep learning target detection method and system | |
CN112364931A (en) | Low-sample target detection method based on meta-feature and weight adjustment and network model | |
CN113487610B (en) | Herpes image recognition method and device, computer equipment and storage medium | |
CN116645592B (en) | Crack detection method based on image processing and storage medium | |
CN113159215A (en) | Small target detection and identification method based on fast Rcnn | |
CN113537085A (en) | Ship target detection method based on two-time transfer learning and data augmentation | |
CN111199255A (en) | Small target detection network model and detection method based on dark net53 network | |
CN111476226B (en) | Text positioning method and device and model training method | |
Yamashita et al. | Cost-alleviative learning for deep convolutional neural network-based facial part labeling | |
Rao et al. | Roads detection of aerial image with FCN-CRF model | |
CN115953743A (en) | Parking space state identification method based on improved YOLO model | |
CN116052149A (en) | CS-ABCNet-based electric power tower plate detection and identification 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 |