CN110647802A - Remote sensing image ship target detection method based on deep learning - Google Patents
Remote sensing image ship target detection method based on deep learning Download PDFInfo
- Publication number
- CN110647802A CN110647802A CN201910726359.XA CN201910726359A CN110647802A CN 110647802 A CN110647802 A CN 110647802A CN 201910726359 A CN201910726359 A CN 201910726359A CN 110647802 A CN110647802 A CN 110647802A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- network
- sensing image
- target
- calibration frame
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—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
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB
Abstract
The invention discloses a remote sensing image ship target detection method based on deep learning, which comprises the following steps: acquiring a remote sensing image data set; extracting the features of the images in the remote sensing image data set by adopting a residual error network to generate a feature mapping chart; generating a prediction calibration frame on the feature mapping chart by adopting a regional generation network, judging whether the prediction calibration frame is a target calibration frame or not, and performing regression analysis to generate an accurate calibration frame; fusing the accurate calibration frame and the feature mapping chart to generate a fixed feature mapping chart for target positioning; and judging the category of each fixed feature mapping image to realize target detection. The invention can realize end-to-end detection and identification of the ship target in the remote sensing image, has better detection accuracy and robustness to complex scenes, can reach 5 frames per second at the detection speed, and can realize real-time detection of the remote sensing target.
Description
Technical Field
The invention relates to the technical field of image recognition. More specifically, the invention relates to a remote sensing image ship target detection method based on deep learning.
Background
The high-resolution remote sensing image ship target detection is a research hotspot in the field of remote sensing image processing, and is widely applied in the fields of military investigation, civil monitoring and the like. The high-resolution satellite remote sensing is used as a main earth observation means, the precision of the satellite remote sensing can reach the sub-meter level, and targets such as ships, ports and the like can be manually interpreted from remote sensing images. However, the manual visual interpretation takes a lot of time and is inefficient. A large number of remote sensing image detection and identification algorithms are provided for researchers. The traditional remote sensing target detection algorithm mainly extracts the characteristics of texture, gradient, gray level and the like of an image to manufacture a feature descriptor (FeatureDescriptor), and then trains a classifier by using the feature descriptor to complete the detection and classification of the target. Scholars Zhu et al segment sea and land by using shape and texture features of images, then extract Local Multiple Patterns (LMP) features and train a Support Vector Machine (SVM) classifier to classify the features, thereby realizing detection and identification of ship targets. The learner Wang Huili and the like improve the directional gradient histogram characteristics according to the inherent characteristics of the ship target, provide the edge-directional gradient histogram characteristics to describe the ship target, and then train an AdaBoost classifier by constructing a training library to complete the final detection of the target. eCoginization software developed by Definens Imaging company in Germany comprehensively judges the characteristics of spectral information, geometric shapes, textures and the like of images by using a fuzzy classification algorithm, and then realizes the detection and classification of remote sensing targets by using a decision system. In the traditional method, the manual design characteristics need to have rich prior knowledge, parameters need to be adjusted manually in training a classifier, and the process is complex and tedious. The method based on the deep neural network can obtain the high-level abstract characteristics of the remote sensing image through the deep network automatic learning, and has high identification precision and working efficiency. With the improvement of computer performance and the improvement of the accelerated computing capability of the GPU in deep learning. The detection and identification method based on the deep neural network far exceeds the traditional detection algorithm in the aspects of detection precision and speed. In 2012, Krizhevsky proposed an image classification algorithm based on Deep Convolutional Neural Network (DCNN), which improves the target detection accuracy and the image classification accuracy. Erhan performs regression prediction on the target using DCNN and gives confidence for related and unrelated objects. Sermantet uses a sliding window for an input image, detects window categories by using a network model, predicts a target boundary box for the window categories, and finally merges candidate boundary boxes according to classification scores to obtain a final detection result. Girshick proposed a target detection network model R-CNN in 2014, which adopts a selective search strategy to improve detection efficiency. Girshick et al propose a Fast R-CNN network model by improving R-CNN by adopting the idea of Spatial Pyramid Pooling (SPP), wherein the network structure greatly improves the operation speed of the model, and the detection speed is about 100 times of that of R-CNN. However, the Fast R-CNN algorithm still has a bottleneck in speed.
Disclosure of Invention
The invention aims to provide a remote sensing image ship target detection method based on deep learning, which can realize end-to-end detection and identification of a ship target in a remote sensing image, has better detection accuracy and robustness to complex scenes, can reach 5 frames per second at the detection speed, and can realize real-time detection of the remote sensing target.
In order to achieve the objects and other advantages according to the present invention, there is provided a remote sensing image ship target detection method based on deep learning, comprising: acquiring a remote sensing image data set; extracting the features of the images in the remote sensing image data set by adopting a residual error network to generate a feature mapping chart; generating a prediction calibration frame on the feature mapping chart by adopting a regional generation network, judging whether the prediction calibration frame is a target calibration frame or not, and performing regression analysis to generate an accurate calibration frame; fusing the accurate calibration frame and the feature mapping chart to generate a fixed feature mapping chart for target positioning; and judging the category of each fixed feature mapping image to realize target detection.
Preferably, the method for detecting the target of the remote sensing image ship based on the deep learning further comprises the steps of pre-training a basic residual error network by using an Imagenet data set before extracting the characteristics of the image in the remote sensing image data set by using the residual error network, migrating a full connection layer and all network layers above the full connection layer in the pre-trained basic residual error network into the residual error network, freezing node parameters of the full connection layer and all network layers above the full connection layer in the basic residual error network, sequentially adding two convolutional layers and one full connection layer in the migrated residual error network, and then training the residual error network by using a training data set contained in the remote sensing image data set.
Preferably, the remote sensing image ship target detection method based on deep learning further comprises the step of sequentially cutting and expanding the images before feature extraction is carried out on the images in the remote sensing image data set by adopting a residual error network.
Preferably, the remote sensing image ship target detection method based on deep learning is expanded by randomly rotating the image by 90 degrees or 180 degrees so as to realize doubling of the number of images.
Preferably, in the method for detecting the target of the remote sensing image ship based on the deep learning, the aspect ratio of the forecast calibration frame is (1:1,1:2,3: 1).
Preferably, the method for detecting the target of the remote sensing image ship based on the deep learning comprises the steps of judging whether the prediction calibration frame is the target calibration frame or not, comparing the intersection ratio of the prediction calibration frame and the manual marking frame with a preset threshold value, and if the intersection ratio is larger than the preset threshold value, judging that the prediction calibration frame is the target calibration frame.
The invention has the beneficial effects that:
the invention improves the fast R-CNN detection algorithm and applies the fast R-CNN detection algorithm to the detection of the remote sensing image of the ship, expands the data set by a data enhancement technology, then uses a depth residual error Network ResNet to replace a traditional flat Network VGG-16 to extract the characteristics of the data set, modifies the length-width ratio of a calibration frame aiming at the characteristic of high length-width ratio of the ship target in a Region generation Network (RPN), performs transfer learning on a ResNet Network model to obtain better training effect, can perform rapid and accurate detection on the ship target in the remote sensing image, has the accuracy rate of 92.3 percent and the detection speed of 5 frames per second, and can also accurately identify the remote sensing image even when the problems of complex background and local shielding of the target are faced. Compared with the traditional ENVI and eCogination classification means, the method provided by the invention improves the target detection effect and the identification efficiency, and improves the detection precision compared with the fast R-CNN detection algorithm.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.
Drawings
FIG. 1 is a schematic flow diagram according to one embodiment of the present invention;
FIG. 2 is a schematic diagram comparing a residual error network with a conventional flat network according to the above embodiment of the present invention;
FIG. 3 is a graph of training errors in the network model training process according to the above-described embodiment of the present invention;
FIG. 4 is a comparison graph of the residual extraction network and the conventional flat network training according to the above embodiment of the present invention;
FIG. 5 is a result of target detection according to the above-described embodiment of the present invention;
FIG. 6 shows the results of conventional detection methods.
Detailed Description
The present invention will be described in further detail with reference to the following examples and the accompanying drawings so that those skilled in the art can practice the invention with reference to the description.
It is to be understood that the terms "having," "including," and "comprising," as used herein, do not preclude the presence or addition of one or more other elements or groups thereof.
As shown in fig. 1, the invention provides a remote sensing image ship target detection method based on deep learning, which comprises the following steps:
and S1, acquiring a remote sensing image data set.
And S2, extracting the features of the images in the remote sensing image data set by adopting a residual error network, and generating a feature mapping chart.
In the method, the residual error network ResNet is used for replacing traditional flat networks such as VGG-16 used in the original Faster R-CNN, the number of layers of the deep neural network is increased, the complexity of network training is reduced, and the problem of gradient disappearance is prevented. The core of the residual network is a residual learning module as shown in fig. 2, and the representation of the residual unit is shown in formulas (1), (2) and (3).
y1＝h(xl)+F(xl，Wl) (1)
xl+1＝f(y1) (2)
Wherein xlAnd xl+1Respectively representing the input and output of the ith residual unit, F is a residual function representing the learned residual, and h (x)l) Representing an identity mapping and f representing a relu activation function. x is the number ofLRepresenting features learned from the shallow layer L to the deep layer L.
Because the ResNet network structure introduces a design similar to a short-circuit type, the data output of the previous layers directly skips a multi-layer network and the input part of the data layer behind is introduced, the loss and compression of the data can be avoided as much as possible in the learning process of the network, the network can learn richer contents, and the problems of overfitting, gradient disappearance and the like in network training are also avoided.
And S3, generating a prediction calibration frame on the feature mapping chart by adopting a regional generation network, judging whether the prediction calibration frame is a target calibration frame or not, performing regression analysis, and generating an accurate calibration frame.
Here, the quality of the prediction scaling box generated by the region generation network RPN directly affects the accuracy of the target detection task. The RPN is a shallow full convolution neural network, which is composed of 1 convolution layer with convolution kernel size of 1 × 1 and 2 full connection layers. Firstly, a sliding window with a specific size is generated, the size of the sliding window is 3 multiplied by 3 in the invention, then the sliding window slides on each central point on an input feature mapping chart, and 9 prediction calibration frames (anchors) are generated at each central point position, in order to solve the problem that the sizes of the targets to be detected are different in proportion, the anchors are endowed with different areas and different length-width ratios, the length-width ratio in the original Faster R-CNN is (1:1,1:2, 2: 1), and the calibration targets can be well predicted. However, the aspect ratio of the ship remote sensing target real calibration frame is counted, the aspect ratio of the ship target is found to be large, so that the aspect ratio in the original RPN is not suitable, and the aspect ratio of the prediction calibration frame is optimized to be (1:1,1:2 and 3:1), so that the method is more suitable for detecting the ship target. And then, calculating the prediction calibration frames through the convolution layer, obtaining a 256-dimensional vector for each prediction calibration frame, judging whether the prediction calibration frame contains a target through the full connection layer, judging whether the target exists in the prediction calibration frame according to the intersection ratio (IOU) of the prediction calibration frame and a manual marking frame (group Truth), and if the IOU of the prediction calibration frame and the manual marking frame is more than 0.7, determining that the target is contained and the sample is a positive sample. If IOU is less than 0.3, it is considered as containing no target, a negative sample. And finally, obtaining candidate marking frames from the prediction marking frames containing the targets, and performing regression calculation on the candidate marking frames according to a regression function to generate the accurate marking frames.
And S4, fusing the accurate calibration frame and the feature map to generate a fixed feature map for target positioning.
And S5, judging the category of each fixed feature mapping image to realize target detection.
Here, the target location and target detection are performed with reference to the original fast R-CNN algorithm, and are not described in detail.
The method for detecting the remote sensing image ship target based on the deep learning further comprises the steps of pre-training a basic residual error network by using an Imagenet data set before extracting the characteristics of the images in the remote sensing image data set by using the residual error network, migrating a full connection layer and all network layers above the full connection layer in the pre-trained basic residual error network into the residual error network, freezing node parameters of the full connection layer and all network layers above the full connection layer in the basic residual error network, sequentially adding two convolutional layers and one full connection layer in the migrated residual error network, and then training the residual error network by using a training data set contained in the remote sensing image data set. The transfer learning can transfer the knowledge learned in a certain homogeneous problem to the problem of the migration learning, and a part of the knowledge (node parameters in the network) of the trained model is directly applied to another type of model. Because the convolutional neural network used by the invention has deeper structural layers and more parameters, a large number of data samples are needed during training. And the data set of the user only comprises 2000 remote sensing images, and overfitting of network training can be caused if the network is trained by directly using the data set of the remote sensing images, so that the problem that a large number of training data samples are needed by a deep network is solved by using a transfer learning method.
Further, the remote sensing image ship target detection method based on deep learning further comprises the step of sequentially cutting and expanding the images before feature extraction is carried out on the images in the remote sensing image data set by adopting a residual error network, wherein the expansion is that the images are randomly rotated by 90 degrees or 180 degrees so as to realize doubling of the number of the images and expansion of data volume.
Naval vessel target detection experiment
1. Selection of a deep learning framework:
tensorflow is selected and developed and maintained by Google, has the advantages of simplicity and convenience in calling, high code quality and high transportability, has a powerful visual component TensorBoard, can visualize a network structure and a training process, and is very helpful for observing the network structure and monitoring long-time and large-scale training.
2. Data preprocessing and data set generation
A marine ship target is selected as a research sample, and two types of ships are mainly detected and classified, wherein the types of the ships are respectively a U.S. naval destroyer (DDG) and an endangered battle ship (LCS). Then cutting the size to 397 × 397, randomly selecting 60% of 1000 images as a training set, 20% as a verification set and 20% as a test set, and finally performing data enhancement by randomly rotating each image by 90 degrees or 180 degrees to obtain 2000 remote sensing images. The details of the data set are shown in table 1. The specific format of the data set adopts the format of a Pascol VOC data set, and the content of the data format is divided into three types, namely JPEGlmages, Annotation and ImageSets. JPEGlmages are used for storing jpg images, Annotation is used for storing XML tags corresponding to each image, and ImageSets mainly store the name of each image.
TABLE 1 Ship image data set statistics
3. Experimental platform and parameters
The software and hardware used in the experimental platform for training the network model are shown in tables 2 and 3. The training is divided into 10 batches, each round comprises 7000 epochs, a Ckpt model file and a node file are generated in each round, the Ckpt model file stores the parameter weight in the network structure, and the node file records the training process. The loss values during training are shown in fig. 3. With the continuous increase of training batches, the training error is gradually reduced, and the network fitting effect is better and better.
TABLE 2 hardware of the Experimental platform
TABLE 3 software for the Experimental platform
4. Network improved experimental contrast
4.1 data enhancement Experimental contrast
The invention adopts the data enhancement technology to rotate the obtained remote sensing data set by 90 degrees and 180 degrees to expand the data set, 1000 previously obtained ship remote sensing image data sets are expanded to 2000, the data set containing 1000 images and the data set containing 2000 images are respectively input into the improved network based on the Faster R-CNN, the training is carried out by using the same parameters, and the final result shows that the detection accuracy of the network to the ship target is increased by 1.36 percent and the recall rate is increased by 3.12 percent after the number of the data sets is increased from 1000 to 2000. Therefore, the detection precision of the network for the remote sensing data can be effectively improved by properly expanding the data set.
4.2 comparison of residual error network with Flat-layer network experiment
In order to better extract the high-dimensional characteristics of the remote sensing image and solve the problem that the identification precision is reduced due to the disappearance of the gradient in the network training process, the invention uses a residual error network ResNet to replace a flat network VGG-16 used by the original Faster R-CNN network. In order to verify that the replacement of the invention is reasonable and effective, a data set containing 2000 remote sensing images is respectively input into an original Faster R-CNN network using VGG-16 to extract features and a Faster R-CNN improved network using ResNet to extract features, results obtained by the two networks are compared, a loss value comparison graph of the two networks in a training process is shown in FIG. 4, and the detection accuracy of the networks is shown in Table 4. As shown in FIG. 4, with the continuous increase of network training batches, the loss value of ResNet network is obviously lower than that of VGG-16 network, so that the problems of gradient disappearance and the like can be better solved in network training by using ResNet instead of VGG-16, and better fitting of the model is realized. As shown in Table 4, the residual error network ResNet is used for replacing the flat-layer network VGG-16, so that the accuracy and recall rate of the network for identifying the ship remote sensing images can be improved. It follows that it is effective to use a ResNet network instead of a VGG-16 network.
TABLE 4 comparison of flat-layer network and residual extraction network experiments
5. The result of the detection
5.1 test results of the invention
The detection results of the remote sensing image ship target are shown in fig. 5, and all the detection results of the data test set are counted in table 5, wherein the detection results of the ship target are given by the calibration of the calibration frame, the confidence coefficient of the target identification as DDG or LCS is displayed above the calibration frame, the value range of the confidence coefficient is 0-1, the closer the confidence coefficient is to 1, the better the detection effect of the network is, and otherwise, the worse the detection effect is. As can be seen from FIG. 5, the confidence of ship detection is very high, and most of the confidence is above 0.95, which indicates that the detection method of the present invention has very high identification precision for DDG or LCS ships. Meanwhile, in the scene with a complex background environment such as a port and the like, the detection method of the invention has fewer false alarms and false alarms, which shows that the detection method of the invention has stronger Robang property for the interference of the complex background. The method has high detection precision, and experiments prove that the method has high detection speed, the detection time of a single image is 0.2 second, and real-time detection can be realized. As can be seen from Table 5, the detection method has high detection accuracy and recall rate for DDG and LCG.
TABLE 5 statistics of test results
5.2 comparison with the results of ENVI, eCogination and Faster R-CNN
In order to show the advantages of the detection algorithm of the invention in ship target detection, the detection method of the invention is compared with the traditional ENVI and eCoginization detection methods. The detection results are shown in fig. 6, and the elapsed time and the division accuracy are shown in table 6.
When ENVI is used for detecting the ship target, a supervision classification method is used for achieving a good classification effect, firstly, ROI (region of interest) of DDG (data transmission graphics) ships and LCS (liquid Crystal display) ships are selected, then separability is calculated for samples of the DDG ships and LCS ships, and the calculation result meets requirements. The classifier uses a Support Vector Machine (SVM), and the detection result is shown in FIG. 6, so that a large amount of missing detection and error detection exist. The reason is that in the ROI selection, the ship target sample is small, the texture of two ships is similar, and the classification error occurs in the ENVI detection, and even the land with similar color (both land and ship are gray) is sometimes classified as a ship. And the detection method based on deep learning can obtain deeper features through deep learning, so that the detection precision is higher.
A large amount of time is spent in detection using the eCognition, mainly because the eCognition is a process of dividing an entire image and is mainly calculated by a CPU. The detection algorithm generates the prediction area through the RPN convolutional neural network, and uses the GPU to carry out calculation acceleration, thereby greatly improving the detection speed. The detection algorithm of the invention realizes end-to-end automatic detection and can be completely automated. While eCognition requires cumbersome manual operations.
When the original fast R-CNN is used for detection, the detection method is found to be higher in detection accuracy and recall rate, mainly because the method uses a residual error network, the problem of gradient disappearance is solved while the depth of a convolution network is increased, and the aspect ratio of a calibration frame is predicted by the RPN network so that the calibration frame is more suitable for a ship target.
TABLE 6 comparison of detection accuracy with elapsed time
The invention provides a remote sensing image detection method based on deep learning, aiming at the problems of low detection precision, incapability of automation and the like in the traditional remote sensing detection algorithm. In order to train the deep network model structure, 1 data set containing 1000 remote sensing images is manually marked, and the data set not only contains ship targets, but also contains a plurality of complex scenes such as seaports and the like. In the training process, parameters do not need to be manually adjusted, the finally obtained network model obtains 92.6% of accuracy and 94.4% of recall rate in the concentrated detection of the ship target, and a better detection result is obtained. The method has better detection accuracy and robustness to complex scenes, can reach 5 frames per second at the detection speed, and can realize real-time detection of the remote sensing target. The method can be applied to interpretation, processing and other works in the remote sensing image.
While embodiments of the invention have been described above, it is not limited to the applications set forth in the description and the embodiments, which are fully applicable in various fields of endeavor to which the invention pertains, and further modifications may readily be made by those skilled in the art, it being understood that the invention is not limited to the details shown and described herein without departing from the general concept defined by the appended claims and their equivalents.
Claims (6)
1. The remote sensing image ship target detection method based on deep learning is characterized by comprising the following steps:
acquiring a remote sensing image data set;
extracting the features of the images in the remote sensing image data set by adopting a residual error network to generate a feature mapping chart;
generating a prediction calibration frame on the feature mapping chart by adopting a regional generation network, judging whether the prediction calibration frame is a target calibration frame or not, and performing regression analysis to generate an accurate calibration frame;
fusing the accurate calibration frame and the feature mapping chart to generate a fixed feature mapping chart for target positioning;
and judging the category of each fixed feature mapping image to realize target detection.
2. The method for detecting the target of the remote sensing image ship based on the deep learning of claim 1, further comprising, before extracting the features of the images in the remote sensing image data set by using a residual network, pre-training a basic residual network by using an Imagenet data set, migrating a full link layer and all network layers above the full link layer in the pre-trained basic residual network into the residual network, freezing the node parameters of the full link layer and all network layers above the full link layer in the basic residual network, then sequentially adding two convolutional layers and one full link layer in the migrated residual network, and then training the residual network by using the training data set contained in the remote sensing image data set.
3. The method for detecting the remote sensing image ship target based on the deep learning of claim 1, further comprising sequentially cutting and expanding the image before extracting the features of the image in the remote sensing image data set by using a residual error network.
4. The remote sensing image ship target detection method based on deep learning of claim 3, wherein the expansion is that the image is randomly rotated by 90 ° or 180 ° to realize doubling of the number of images.
5. The method for detecting the ship target based on the deep learning remote sensing image as claimed in claim 1, wherein the aspect ratio of the prediction calibration frame is (1:1,1:2,3: 1).
6. The method for detecting the target of the remote sensing image ship based on the deep learning as claimed in claim 1, wherein the step of judging whether the predicted calibration frame is the target calibration frame comprises the steps of comparing the intersection ratio of the predicted calibration frame and the manual calibration frame with a preset threshold value, and if the intersection ratio is greater than the preset threshold value, judging that the predicted calibration frame is the target calibration frame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910726359.XA CN110647802A (en) | 2019-08-07 | 2019-08-07 | Remote sensing image ship target detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910726359.XA CN110647802A (en) | 2019-08-07 | 2019-08-07 | Remote sensing image ship target detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110647802A true CN110647802A (en) | 2020-01-03 |
Family
ID=68990047
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910726359.XA Pending CN110647802A (en) | 2019-08-07 | 2019-08-07 | Remote sensing image ship target detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110647802A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111232200A (en) * | 2020-02-10 | 2020-06-05 | 北京建筑大学 | Target detection method based on micro aircraft |
| CN111611925A (en) * | 2020-05-21 | 2020-09-01 | 重庆现代建筑产业发展研究院 | Building detection and identification method and device |
| CN111931572A (en) * | 2020-07-07 | 2020-11-13 | 广东工业大学 | Target detection method of remote sensing image |
| CN111985376A (en) * | 2020-08-13 | 2020-11-24 | 湖北富瑞尔科技有限公司 | Remote sensing image ship contour extraction method based on deep learning |
| CN112270326A (en) * | 2020-11-18 | 2021-01-26 | 珠海大横琴科技发展有限公司 | Detection optimization method and device for ship sheltering and electronic equipment |
| CN112288819A (en) * | 2020-11-20 | 2021-01-29 | 中国地质大学（武汉） | Multi-source data fusion vision-guided robot grabbing and classifying system and method |
| CN112365540A (en) * | 2020-11-18 | 2021-02-12 | 北京观微科技有限公司 | Ship target positioning detection method and system suitable for multiple scales |
| CN112949614A (en) * | 2021-04-29 | 2021-06-11 | 成都市威虎科技有限公司 | Face detection method and device for automatically allocating candidate areas and electronic equipment |
| US20210226583A1 (en) * | 2020-01-21 | 2021-07-22 | Arizona Board Of Regents On Behalf Of Arizona State University | Autonomous solar field and receiver inspections based on polarimetric-enhanced imaging |
| CN113870284A (en) * | 2021-09-29 | 2021-12-31 | 柏意慧心(杭州)网络科技有限公司 | Method, apparatus, and medium for segmenting medical images |
-
2019
- 2019-08-07 CN CN201910726359.XA patent/CN110647802A/en active Pending
Non-Patent Citations (2)
| Title |
|---|
| 张倩: "《基于深度学习的安全目标检测与识别研究》", 《中国优秀博硕士学位论文全文数据库（硕士） 信息科技辑》 * |
| 李全杰: "《基于迁移学习和深度卷积神经网络的车标识别方法研究》", 《中国优秀博硕士学位论文全文数据库（硕士） 信息科技辑》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11843350B2 (en) * | 2020-01-21 | 2023-12-12 | Arizona Board Of Regents On Behalf Of Arizona State University | Autonomous solar field and receiver inspections based on polarimetric-enhanced imaging |
| US20210226583A1 (en) * | 2020-01-21 | 2021-07-22 | Arizona Board Of Regents On Behalf Of Arizona State University | Autonomous solar field and receiver inspections based on polarimetric-enhanced imaging |
| CN111232200B (en) * | 2020-02-10 | 2021-07-16 | 北京建筑大学 | Target detection method based on micro aircraft |
| CN111232200A (en) * | 2020-02-10 | 2020-06-05 | 北京建筑大学 | Target detection method based on micro aircraft |
| CN111611925A (en) * | 2020-05-21 | 2020-09-01 | 重庆现代建筑产业发展研究院 | Building detection and identification method and device |
| CN111931572A (en) * | 2020-07-07 | 2020-11-13 | 广东工业大学 | Target detection method of remote sensing image |
| CN111931572B (en) * | 2020-07-07 | 2024-01-09 | 广东工业大学 | Target detection method for remote sensing image |
| CN111985376A (en) * | 2020-08-13 | 2020-11-24 | 湖北富瑞尔科技有限公司 | Remote sensing image ship contour extraction method based on deep learning |
| CN112365540A (en) * | 2020-11-18 | 2021-02-12 | 北京观微科技有限公司 | Ship target positioning detection method and system suitable for multiple scales |
| CN112270326B (en) * | 2020-11-18 | 2022-03-22 | 珠海大横琴科技发展有限公司 | Detection optimization method and device for ship sheltering and electronic equipment |
| CN112270326A (en) * | 2020-11-18 | 2021-01-26 | 珠海大横琴科技发展有限公司 | Detection optimization method and device for ship sheltering and electronic equipment |
| CN112288819A (en) * | 2020-11-20 | 2021-01-29 | 中国地质大学（武汉） | Multi-source data fusion vision-guided robot grabbing and classifying system and method |
| CN112949614A (en) * | 2021-04-29 | 2021-06-11 | 成都市威虎科技有限公司 | Face detection method and device for automatically allocating candidate areas and electronic equipment |
| CN113870284A (en) * | 2021-09-29 | 2021-12-31 | 柏意慧心(杭州)网络科技有限公司 | Method, apparatus, and medium for segmenting medical images |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110647802A (en) | Remote sensing image ship target detection method based on deep learning | |
| CN110598029B (en) | Fine-grained image classification method based on attention transfer mechanism | |
| CN109101897A (en) | Object detection method, system and the relevant device of underwater robot | |
| CN112395987B (en) | SAR image target detection method based on unsupervised domain adaptive CNN | |
| CN111079739B (en) | Multi-scale attention feature detection method | |
| CN113569667B (en) | Inland ship target identification method and system based on lightweight neural network model | |
| CN110569738A (en) | natural scene text detection method, equipment and medium based on dense connection network | |
| CN110598693A (en) | Ship plate identification method based on fast-RCNN | |
| CN112348758B (en) | Optical remote sensing image data enhancement method and target identification method | |
| Yang et al. | Visual tracking with long-short term based correlation filter | |
| CN112926486A (en) | Improved RFBnet target detection algorithm for ship small target | |
| CN114781514A (en) | Floater target detection method and system integrating attention mechanism | |
| CN116168240A (en) | Arbitrary-direction dense ship target detection method based on attention enhancement | |
| CN111968124A (en) | Shoulder musculoskeletal ultrasonic structure segmentation method based on semi-supervised semantic segmentation | |
| CN112084860A (en) | Target object detection method and device and thermal power plant detection method and device | |
| Fan et al. | A novel sonar target detection and classification algorithm | |
| CN112149664A (en) | Target detection method for optimizing classification and positioning tasks | |
| CN114565824A (en) | Single-stage rotating ship detection method based on full convolution network | |
| Yulin et al. | Wreckage target recognition in side-scan sonar images based on an improved faster r-cnn model | |
| CN113128564B (en) | Typical target detection method and system based on deep learning under complex background | |
| CN112991281B (en) | Visual detection method, system, electronic equipment and medium | |
| CN115272856B (en) | Ship target fine-grained identification method and equipment | |
| CN113076876B (en) | Face spoofing detection method and system based on three-dimensional structure supervision and confidence weighting | |
| CN115861595A (en) | Multi-scale domain self-adaptive heterogeneous image matching method based on deep learning | |
| CN116912670A (en) | Deep sea fish identification method based on improved YOLO model |
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 |