CN109934122A - A kind of remote sensing image ship detecting method based on deep learning - Google Patents
A kind of remote sensing image ship detecting method based on deep learning Download PDFInfo
- Publication number
- CN109934122A CN109934122A CN201910128637.1A CN201910128637A CN109934122A CN 109934122 A CN109934122 A CN 109934122A CN 201910128637 A CN201910128637 A CN 201910128637A CN 109934122 A CN109934122 A CN 109934122A
- Authority
- CN
- China
- Prior art keywords
- image
- target
- remote sensing
- ship
- carried out
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The remote sensing image ship detecting method based on deep learning that the invention discloses a kind of, it specifically includes that and 1) first has to pre-process remote sensing image, 2) spatial sampling operation is carried out, obtain a large amount of ship target sample image, 3) it planned network structure and is built, 4) data set of generation is sent into network model, 5) after network model training, careful division is carried out using output result marginal information of the full terms random field to model, 6) using in test sample feeding designed model, export result, and check the ship effect picture detected.The invention patent passes through the analysis to remote sensing images above water craft target, it is made an uproar according to target property to the image place of being filtered, the image procossings such as enhancing, and cloud removing is carried out the interference of eliminating cloud layer to target, good target image is obtained using image partition method simultaneously, and spatial sampling is carried out to the target image of acquisition, to provide more required training sample images.
Description
Technical field
The invention belongs to deep learnings and remote sensing fields, particularly relate to a kind of algorithm of target detection, are used for large scale remote sensing
Image ship test problems.
Background technique
Computer technology, optical remote sensing technology promotion under, the resolution ratio of the remote sensing images of acquisition is being continuously improved,
It can be applied to the target detection of ship on ocean.Although remote sensing image is easy by environmental factors such as cloud and mist stormy waves streams
Influence and cannot for 24 hours, round-the-clock detection, but its high-resolution, sea information collects (ship is identified on ocean), with
And the meaning that secondary development has importance is carried out to information is extracted.
But the template matching method that no matter detects using traditional ship or with the mode of machine learning carry out target
Identification, not only takes a large amount of man power and material, and the study identification structure built has the target identified
Limitation.It is specifically exactly that target similar for a kind of or several category features carries out target with designed identification model
Identification, but to identify that the discrimination of model designed when the bigger target of another category feature gap will be significantly lower,
It even fails, the robustness of method is very poor.
Summary of the invention
In view of above-mentioned technical background, it is an object of the invention to: a kind of remote sensing image ship based on deep learning is provided
Detection method, it is characterized in that the model does not require the input size of data, and the output result of the model is original defeated
The size entered is consistent, and SoftMax function carries out the detection of ship target to output layer;Using full terms random field to the model
The marginal information of output result optimizes.
In order to solve the above technical problems, the present invention provides a kind of remote sensing image ship detecting side based on deep learning
Method, comprising the following steps::
1) it first has to pre-process remote sensing image, including geometric correction, thin cloud and mist removal, image ship targets improvement
Deng, with this come obtain succeeding target identification training sample data collection;
2) spatial sampling operation is carried out, a large amount of ship target sample image is obtained;
3) it planned network structure and is built, is realized using Tensorflow frame;
4) data set of generation is sent into network model, is iterated training, until sorter network is restrained;
5) it after network model training, is carried out using output result marginal information of the full terms random field to model
Careful division;
6) it using in test sample " feeding " designed model, exports as a result, and checking the ship effect detected
Figure.
Preferably, model in VGG-19 network models by using convolution operation to replace traditional full articulamentum
Mode constructs, and carries out the careful optimization in edge using result of the full condition of contact random field to network model.
Preferably, network model includes the convolutional layer extracted for bottom visual signature, high-layer semantic information;For feature
Fusion, the pond layer for reducing dimension;For keeping output result up-sampling layer identical with raw video size;For optimizing
The condition random field of FCN output pixel classification;The input includes original image;The output includes that test picture is corresponding
Tag along sort probability distribution.
Preferably, model does not require the input size of data, and the output result of the model is originally inputted
Size is consistent, and softmax function carries out the detection of ship target to output layer;The model is exported using full terms random field
As a result marginal information optimizes.
Preferably, thin cloud and mist removal, it will usually be handled using homographic filtering method, which is substantially exactly by image
Greyscale transformation and image filtering combine.Since cloud layer is in the high illuminated state of presentation of remote sensing images, nonlinear transformation is carried out to it, is increased
Add contrast, to enhance dark place image detail, and high-pass filtering is carried out to image, asked to solve image overall brightness unevenness
Topic has achieved the effect that remove cloud layer.
Preferably, by the analysis to remote sensing images above water craft target, place is filtered to image according to target property
It makes an uproar, the image procossings such as enhancing, and cloud removing is carried out the interference of eliminating cloud layer to target, while using image partition method
Good target image is obtained, and spatial sampling is carried out to the target image of acquisition, to provide more required training sample figures
Training sample image is sent into designed network model, is iterated training by picture.It will be generated after repetitive exercise
Result figure carried out full condition of contact Random Fields Method, as post-processing with refined image segmentation result.
Compared with prior art, the invention has the following beneficial effects:
In the present invention, by the analysis to remote sensing images above water craft target, image is filtered according to target property
Place makes an uproar, the image procossings such as enhancing, and carries out cloud removing the interference of eliminating cloud layer to target, while using image partition method
To obtain good target image, and spatial sampling is carried out to the target image of acquisition, to provide more required training samples
Training sample image is sent into designed network model, is iterated training, will give birth to after repetitive exercise by image
At result figure carried out full condition of contact Random Fields Method, as post-processing with refined image segmentation result, bring preferably
Prospect of the application.
Detailed description of the invention
Fig. 1 is a kind of flow chart of the remote sensing image ship detecting method based on deep learning of the present invention.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings, and following embodiment is only used for clearly illustrating the present invention
Technical solution, and not intended to limit the protection scope of the present invention.
Embodiment 1
Step S1 first has to pre-process remote sensing image, including geometric correction, thin cloud and mist removal, image ship mesh
Mark enhancing etc., the training sample data collection of succeeding target identification is obtained with this.
Thin cloud and mist removal: would generally be handled using homographic filtering method, which is substantially exactly to become image grayscale
It changes and is combined with image filtering.Since cloud layer is in the high illuminated state of presentation of remote sensing images, nonlinear transformation is carried out to it, increases comparison
Degree to enhance dark place image detail, and carries out high-pass filtering to image and reaches to solve the problems, such as image overall brightness unevenness
The effect of removal cloud layer is arrived.
Assuming that remote sensing image is seen as a two-dimensional function f (x, y), and it is regarded as a two-dimension light source amount of incident letter
Number f (x, y) and a two dimension reflection flow function frThe product of (x, y).It is represented by,
F (x, y)=fi(x, y) × fr(x, y)
Logarithmic transformation carried out to above-mentioned formula, while and Fourier transformation is carried out simultaneously to volume reflection and amount of incident,
F (ln (f (x, y)))=F (ln (fi(x, y)))+F (ln (fr(x, y)))
If F (μ, v)=F (ln (f (x, y))), F (μ, v)=F (ln (f (x, y))), F (μ, v)=F (ln (f (x, y)))
F in this way (μ, v)=I (μ, v)+R (μ, v).
Wherein, amount of incident function frequency represented by I (μ, v) changes, the low-frequency range in frequency domain, and I (μ, v) is represented
Volume reflection function frequency variation, the high band in frequency domain.
It designs homomorphic filter H (λ, v) and carries out high-pass filtering, lower frequency region spectral characteristic is compressed, to reach filtering low
The effect of domain cloud layer,
Wherein, D0For by frequency.
And then be arranged high-pass filtering after function be S (μ, v)=H (μ, v) × F (μ, v)=H (μ, v) × I (μ, v)+H (μ,
V) × R (μ, v)
Above-mentioned formula is subjected to inverse Fourier transform, realizes that function converts back to spatial domain from frequency domain.
S (x, y)=F-1(S(μ, v))=F-1(H (μ, v) × I (μ, v))+F-1(H (μ, v) × R (μ, v))
Enable F-1(H (μ, v) × I(μ, v))=gi(x, y), F-1(H (μ, v) × R (μ, v))=gr(x, y).Image is restored
Intensity value ranges to where original function f (x, y), i.e.,
G (x, y)=exp (S (x, y))=exp (gi(x, y))+exp (gr(x, y))
Eventually by this method, the thin cloud sector domain in remote sensing images is removed, to obtain required remote sensing target figure
Picture.
Ship target enhancing: the interference by similar pixel cluster around, such as green pepper are easy during detecting ship target
The result of the background of salt noise or ocean complexity meeting Interference Detection.Here by being carried out using Top-Hat operator to ship mesh
Target enhancing.
Step S2 due to lacking training sample, can carry out spatial sampling after training sample set generates
Operation, target in each sample is rotated, and translation equiaffine transformation increases training sample set, to avoid training process
The phenomenon that middle model over-fitting.
The experiment sample collection sampled is upset at random, and is sent into designed network model by step S3, network knot
Structure is as shown in table 1.
Wherein, Deconv layers of expression warp lamination, and be up-sampling layer, major function will adopt on the Hot-Map of generation
Sample is the sample-size size for being originally inputted layer, and combines raw video medium-high frequency information.The result figure ultimately generated with it is original
It is consistent to input size, and only there are two channels.
The sample sampled is sent into the network and is iterated training by step S4.Here the number of iterations reaches 10000
Secondary, iteration optimization device is Adam optimizer, and the mini-batch size used is 16, and initial learning rate is 3le-3.
Step S5, after realizing network model training, to realize that the fining at detection zone edge divides, using complete
Condition of contact Random Fields Method.
Formula below this method use indicates the integral energy of output result, in other words chaos degree, referred to as energy
Function, obtain the function value it is smaller illustrate that chaos degree is lower, the classification of each pixel is more accurate.
E (x)=∑i(θi(xi))+∑ij(θij(xi, xj))
Wherein, using θi(xi) indicate to represent the cohesion degree of pixel, referred to as unitary potential function, work as pixel for measuring
The color value of point i is yiWhen, which belongs to class label xiProbability.
θi(xi)=- logP (xi)
Wherein, θij(xi, xj) it is binary potential function, two adjacent pixels, if color value yi, yjClosely, that
The two pixels xi, xjGreat talent couple should be compared by belonging to the other probability of same class;If instead color difference is bigger,
The probability that the result that so we divide is split from the two pixels should compare great talent couple.This energy term precisely in order to
By segmentation result splitting from the place of image border as far as possible, that is, in order to make up the shortcoming of preceding networks model segmentation
Wherein μ (xi, xj) it is mark bit function, wkFor the weight of each Gaussian kernel.Gk(fi, fj) it is k-th of Gaussian kernel letter
Number.
By minimizing CRF energy function above, so that it may realize in CRF to hidden variable X (true tag of pixel)
Reasoning.
The present invention is filtered place to image by the analysis to remote sensing images above water craft target, according to target property
It makes an uproar, the image procossings such as enhancing, and cloud removing is carried out the interference of eliminating cloud layer to target, while using image partition method
Good target image is obtained, and spatial sampling is carried out to the target image of acquisition, to provide more required training sample figures
Picture.Training sample image is sent into designed network model, training is iterated.It will be generated after repetitive exercise
Result figure carried out full condition of contact Random Fields Method, as post-processing with refined image segmentation result.
Embodiment 2
Remote sensing images above water craft object detection method is realized on Tensorflow frame, using stream as shown in Figure 1
Journey, comprising the following steps:
Step S1~step S8 is the same as embodiment 1.
Embodiment 3
The pretreatment of image is carried out in the homologous satellite remote-sensing image of several GF-2, such as thin cloud and mist removes, image enhancement
Deng;The sample space of generation is sampled, about 3000 224 × 224 × 3 patches samples are finally obtained;Sample is sent respectively
Enter maximal possibility estimation (MLE), support vector machines (SVM), convolutional neural networks (CNN) and method of the invention (Ours) four
The method of kind of detection ship, and respectively obtain as a result, it uses recall, precision and F1 Measure Indexes evaluate four kinds
Method.
The processed GF-2 data set experimental result of table 2
In the comparative experiments of the method for the present invention others Classification in Remote Sensing Image model, this algorithm is on processed GF-2 data set
Nicety of grading significantly improves.And realize that high quality is classified in a manner of end to end, while the model of training is to remote sensing image data
With preferable capability of fitting and generalization ability.
The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention.The technology of the industry
Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe this
The principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes
Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its
Equivalent thereof.
Claims (6)
1. a kind of remote sensing image ship detecting method based on deep learning, which is characterized in that mainly comprise the steps that
1) it first has to pre-process remote sensing image, including geometric correction, thin cloud and mist removal, image ship targets improvement etc.,
The training sample data collection of succeeding target identification is obtained with this;
2) spatial sampling operation is carried out, a large amount of ship target sample image is obtained;
3) it planned network structure and is built, is realized using Tensorflow frame;
4) data set of generation is sent into network model, is iterated training, until sorter network is restrained;
5) after network model training, it is careful to be carried out using output result marginal information of the full terms random field to model
It divides;
6) it using in test sample " feeding " designed model, exports as a result, and checking the ship effect picture detected.
2. a kind of remote sensing image ship detecting method based on deep learning according to right 1, it is characterised in that: model is logical
It crosses and is constructed in VGG-19 network models using the mode that convolution operation replaces traditional full articulamentum, and use full connection
Condition random field carries out the careful optimization in edge to the result of network model.
3. a kind of remote sensing image ship detecting method based on deep learning according to right 1, it is characterised in that: network mould
Type includes the convolutional layer extracted for bottom visual signature, high-layer semantic information;For Fusion Features, the pond of reduction dimension
Layer;For keeping output result up-sampling layer identical with raw video size;For optimizing the condition of FCN output pixel classification
Random field;The input includes original image;The output includes the probability distribution of the corresponding tag along sort of test picture.
4. a kind of remote sensing image ship detecting method based on deep learning according to right 1, it is characterised in that: model pair
The input size of data does not require, and the output result of the model is that the size that is originally inputted is consistent, softmax function pair
The detection of output layer progress ship target;It is carried out using marginal information of the full terms random field to model output result excellent
Change.
5. a kind of remote sensing image ship detecting method based on deep learning according to right 1, it is characterised in that: thin cloud and mist
Removal, it will usually be handled using homographic filtering method, which is substantially exactly by image gray-scale transformation and image filtering phase
In conjunction with since cloud layer is in the high illuminated state of presentation of remote sensing images, nonlinear transformation is carried out to it, increases contrast, so that enhancing is dark
Locate image detail, and high-pass filtering is carried out to image, to solve the problems, such as image overall brightness unevenness, reaches removal cloud layer
Effect.
6. a kind of remote sensing image ship detecting method based on deep learning according to right 5, it is characterised in that: by right
The analysis of remote sensing images above water craft target makes an uproar to the image place of being filtered according to target property, and the image procossings such as enhancing are gone forward side by side
Row cloud removing eliminates interference of the cloud layer to target, while good target image is obtained using image partition method, and
Spatial sampling is carried out to the target image of acquisition, to provide more required training sample images, training sample image is sent into
In designed network model, it is iterated training, the result figure of generation has been subjected to full connection after repetitive exercise
Maximum matching method, as post-processing with refined image segmentation result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910128637.1A CN109934122A (en) | 2019-02-21 | 2019-02-21 | A kind of remote sensing image ship detecting method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910128637.1A CN109934122A (en) | 2019-02-21 | 2019-02-21 | A kind of remote sensing image ship detecting method based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109934122A true CN109934122A (en) | 2019-06-25 |
Family
ID=66985687
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910128637.1A Pending CN109934122A (en) | 2019-02-21 | 2019-02-21 | A kind of remote sensing image ship detecting method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109934122A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110867045A (en) * | 2019-11-07 | 2020-03-06 | 武汉多谱多勒科技有限公司 | Infrared image human body detection method and device used in fire fighting site |
CN111784676A (en) * | 2020-07-03 | 2020-10-16 | 湖南大学 | Novel feature extraction and segmentation method for liver CT image |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101359399A (en) * | 2008-09-19 | 2009-02-04 | 常州工学院 | Cloud-removing method for optical image |
CN105740894A (en) * | 2016-01-28 | 2016-07-06 | 北京航空航天大学 | Semantic annotation method for hyperspectral remote sensing image |
CN106897683A (en) * | 2017-02-15 | 2017-06-27 | 武汉喜恩卓科技有限责任公司 | The ground object detecting method and system of a kind of remote sensing images |
CN107516103A (en) * | 2016-06-17 | 2017-12-26 | 北京市商汤科技开发有限公司 | A kind of image classification method and system |
CN107527352A (en) * | 2017-08-09 | 2017-12-29 | 中国电子科技集团公司第五十四研究所 | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks |
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 |
CN108491854A (en) * | 2018-02-05 | 2018-09-04 | 西安电子科技大学 | Remote sensing image object detection method based on SF-RCNN |
CN108830224A (en) * | 2018-06-19 | 2018-11-16 | 武汉大学 | A kind of high-resolution remote sensing image Ship Target Detection method based on deep learning |
CN108830844A (en) * | 2018-06-11 | 2018-11-16 | 北华航天工业学院 | A kind of facilities vegetable extracting method based on multidate high-resolution remote sensing image |
-
2019
- 2019-02-21 CN CN201910128637.1A patent/CN109934122A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101359399A (en) * | 2008-09-19 | 2009-02-04 | 常州工学院 | Cloud-removing method for optical image |
CN105740894A (en) * | 2016-01-28 | 2016-07-06 | 北京航空航天大学 | Semantic annotation method for hyperspectral remote sensing image |
CN107516103A (en) * | 2016-06-17 | 2017-12-26 | 北京市商汤科技开发有限公司 | A kind of image classification method and system |
CN106897683A (en) * | 2017-02-15 | 2017-06-27 | 武汉喜恩卓科技有限责任公司 | The ground object detecting method and system of a kind of remote sensing images |
CN107527352A (en) * | 2017-08-09 | 2017-12-29 | 中国电子科技集团公司第五十四研究所 | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks |
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 |
CN108491854A (en) * | 2018-02-05 | 2018-09-04 | 西安电子科技大学 | Remote sensing image object detection method based on SF-RCNN |
CN108830844A (en) * | 2018-06-11 | 2018-11-16 | 北华航天工业学院 | A kind of facilities vegetable extracting method based on multidate high-resolution remote sensing image |
CN108830224A (en) * | 2018-06-19 | 2018-11-16 | 武汉大学 | A kind of high-resolution remote sensing image Ship Target Detection method based on deep learning |
Non-Patent Citations (4)
Title |
---|
LIANG-CHIEH CHEN 等: "SEMANTIC IMAGE SEGMENTATION WITH DEEP CONVOLUTIONAL NETS AND FULLY CONNECTED CRFS", 《ARXIV》 * |
田萱 等: "基于深度学习的图像语义分割方法综述", 《软件学报》 * |
米禹丰: "基于卫星遥感图像水面船舶目标检测与识别技术研究", 《中国优秀硕士学位论文全文数据库(电子期刊) 信息科技辑》 * |
陈天华 等: "采用改进DeepLab网络的遥感图像分割", 《测控技术》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110867045A (en) * | 2019-11-07 | 2020-03-06 | 武汉多谱多勒科技有限公司 | Infrared image human body detection method and device used in fire fighting site |
CN111784676A (en) * | 2020-07-03 | 2020-10-16 | 湖南大学 | Novel feature extraction and segmentation method for liver CT image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Salman et al. | Automatic fish detection in underwater videos by a deep neural network-based hybrid motion learning system | |
Luo et al. | Automated plankton image analysis using convolutional neural networks | |
Li et al. | Fabric defect detection based on multi-scale wavelet transform and Gaussian mixture model method | |
Mahmood et al. | Deep learning for coral classification | |
CN109376611A (en) | A kind of saliency detection method based on 3D convolutional neural networks | |
CN102096824B (en) | Multi-spectral image ship detection method based on selective visual attention mechanism | |
CN109376591B (en) | Ship target detection method for deep learning feature and visual feature combined training | |
CN109034184A (en) | A kind of grading ring detection recognition method based on deep learning | |
Han et al. | Research on multiple jellyfish classification and detection based on deep learning | |
CN109934122A (en) | A kind of remote sensing image ship detecting method based on deep learning | |
Lima et al. | Application of deep convolutional neural networks for ocean front recognition | |
CN116168240A (en) | Arbitrary-direction dense ship target detection method based on attention enhancement | |
Shantkumari et al. | Grape leaf image classification based on machine learning technique for accurate leaf disease detection | |
CN103185731A (en) | Device for detecting beef tenderness based on color image textural features and method thereof | |
Siddiqi | Fruit-classification model resilience under adversarial attack | |
Yang et al. | Contrastive learning-based image retrieval for automatic recognition of in situ marine plankton images | |
Hussain et al. | Image denoising to enhance character recognition using deep learning | |
Luo et al. | RBD-Net: robust breakage detection algorithm for industrial leather | |
Ju et al. | A novel fully convolutional network based on marker-controlled watershed segmentation algorithm for industrial soot robot target segmentation | |
Shantkumari et al. | Machine learning techniques implementation for detection of grape leaf disease | |
Ran et al. | Sketch-guided spatial adaptive normalization and high-level feature constraints based GAN image synthesis for steel strip defect detection data augmentation | |
CN115311544A (en) | Underwater fish target detection method and device | |
SHAKER et al. | The impact of image enhancement and transfer learning techniques on marine habitat mapping | |
Rimavičius et al. | Automatic benthic imagery recognition using a hierarchical two-stage approach | |
Blais et al. | Deep learning for low altitude coastline segmentation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190625 |