CN109670508A - A kind of cloud atlas segmentation network and its method intensively connecting full convolutional network based on symmetrical expression - Google Patents
A kind of cloud atlas segmentation network and its method intensively connecting full convolutional network based on symmetrical expression Download PDFInfo
- Publication number
- CN109670508A CN109670508A CN201811569320.3A CN201811569320A CN109670508A CN 109670508 A CN109670508 A CN 109670508A CN 201811569320 A CN201811569320 A CN 201811569320A CN 109670508 A CN109670508 A CN 109670508A
- Authority
- CN
- China
- Prior art keywords
- network
- link block
- sampling module
- full convolutional
- intensive
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
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)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of cloud atlas segmentation networks and its method that full convolutional network is intensively connected based on symmetrical expression, the intensive convolutional network layer of the symmetrical expression for intensively connecting full convolutional network first with symmetrical expression carries out spy with shallow-layer feature to the deep layer of cloud atlas and merges, and then cloud atlas is divided into cloud and non-cloud region using fused characteristic information by the symmetrical full convolutional layer for intensively connecting full convolutional network.The present invention is higher than conventional satellite cloud atlas dividing method accuracy rate, robustness is stronger, without carrying out complicated Feature Engineering to cloud atlas, and under the conditions of same hardware the test sample time also it is faster than most methods very much.
Description
Technical field
The invention belongs to cloud technical field of image processing, in particular to a kind of intensively to connect full convolution net based on symmetrical expression
The cloud atlas segmentation network and its method of network.
Background technique
Extensive research has been obtained as a kind of important meteorological element in cloud in the past few decades, right at present
There are important application, such as the prediction of instant precipitation predicting, cloud cover, ocean face in various industries in the analysis of cloud feature
Color acquisition of information, optical remote sensing application, satellite communication connection optimization.Segmentation is the first step of nephanalysis, due to Yun Tian
Influenced that there is no fixed shapes by air-flow in the air, and over time with the variation of illumination, the shape of cloud atlas is special
Sign can change, therefore the classical image partition method based on shape prior divides field suitable for application in cloud atlas, can
Accurately segmentation cloud atlas is still a challenging task.The main method of cloud atlas segmentation at present has: by comparing RGB
The ratio in the channel R and channel B in color space, then by adjusting the threshold value of ratio, the two of Lai Shengcheng cloud and non-cloud region
It is worth mask.The segmentation of cloud atlas is realized by the super-pixel method of Adaptive Thresholding, characteristics of image.Machine learning algorithm is in cloud atlas
The automatic segmentation in field is applied.In above-mentioned cloud atlas dividing method, traditional cloud atlas based on threshold method and machine learning method
Dividing method is more sensitive to the selection of the parameter of algorithm, lacks certain robustness, and the prospect and background of working as image are more
Relatively good result could be obtained by needing researcher to do a large amount of Feature Engineering when complicated.
In addition cloud atlas dividing method is predominantly based on the full convolutional network of deep learning (FCN).But FCN is last to network
The characteristic pattern of one convolutional layer just makes it restore size identical with original input image, this behaviour only with step up-sampling operation
The result that will lead to segmentation is not fine enough.
Summary of the invention
The purpose of the present invention is overcoming the shortcomings of above-mentioned background technique, one kind is provided based on symmetrical expression and intensively connects full convolution
The cloud atlas dividing method of network, it is inadequate to Cloud-Picture Characteristics utilization rate to overcome traditional algorithm, divides coarse disadvantage.In order to realize
Above-mentioned technical purpose, the technical solution of the present invention is as follows:
One kind intensively connecting full convolutional network cloud atlas segmentation network, including symmetrically arranged feature extraction net based on symmetrical expression
Network and full convolutional network；
The feature extraction network, for carrying out feature extraction, input of the feature vector of extraction as full convolutional network；
It includes several intensive link blocks for realizing feature extraction that this feature, which extracts network, and the intensive link block is consistent by size
Characteristic pattern be attached, and by convolutional neural networks to ground cloud atlas carry out feature extraction；
The full convolutional network, including several up-sampling modules, the up-sampling module are extracted according to feature extraction network
The feature vector arrived obtains convolution output result；Result and formula (1), output probability figure are exported according to convolution；
Wherein, fijThe corresponding probability of i row j column pixel in probability graph is represented, w is convolution nuclear parameter, and T indicates that transposition, b indicate
Bias term.
Further, the quantity of the intensive link block is consistent with up-sampling module number.
Further, the input of each intensive link block is the down-sampling knot of previous intensive link block output
Fruit, input of the output of current intensive link block as the intensive link block of next stage；
In the input of the last one intensive link block exported as first up-sampling module, each up-sampling
The input of module is the output result of previous up-sampling module；
Each intensive gang mould block has corresponding up-sampling module, and the relationship between them is to export
As a result having the same wide and high.The output result of the corresponding up-sampling module of each intensive link block is carried out
Parallel connection, and using the result after parallel connection as the input of next up-sampling module；
After the last one up-samples module, use dimension for 1 full convolutional layer, the input of the full convolutional layer is last
The output of one up-sampling module.
The invention also discloses one kind intensively to connect full convolutional network cloud atlas dividing method, including following step based on symmetrical expression
It is rapid:
S1: cloud atlas segmentation network is trained: utilizing the sample (X markedi, Yi) be trained, after being trained
Network parameter, wherein XiFor the image of a N × N, YiIndicate XiCorresponding cloud and non-cloud region, i represent i-th of sample, i=
1,2,3...p, p are total sample number；
S2: being N × N size by ground cloud atlas image cropping, as the input data of cloud atlas segmentation network, according to formula (1)
Obtain the probability graph of N*N size:
Wherein, fijThe corresponding probability of i row j column pixel in probability graph is represented, w is convolution nuclear parameter, represents bias term, T table
Show that transposition, b indicate bias term；
S3: according to cloud sector domain decision threshold, the judgement in cloud sector domain and non-cloud region is carried out to the probability graph of S2 output, when general
Rate is greater than the pixel of decision threshold as cloud sector domain, otherwise is used as non-cloud region.
Further, the cloud atlas segmentation network includes symmetrically arranged feature extraction network and deconvolution network；
The feature extraction network, for carrying out feature extraction, input of the feature vector of extraction as deconvolution network；
It includes several intensive link blocks for realizing feature extraction that this feature, which extracts network, and the intensive link block is consistent by size
Characteristic pattern be attached, and by convolutional neural networks to ground cloud atlas carry out feature extraction；
The full convolutional network, including several up-sampling modules, the up-sampling module obtain convolution according to feature vector
Export result；It is exported according to convolution as a result, output probability figure.
Further, the quantity of the intensive link block is consistent with up-sampling module number.
Further, the input of each intensive link block is the down-sampling knot of previous intensive link block output
Fruit, input of the output of current intensive link block as the intensive link block of next stage；
In the input of the last one intensive link block exported as first up-sampling module, each up-sampling
The input of module is the output result of previous up-sampling module；
Each intensive gang mould block has corresponding up-sampling module, and the relationship between them is to export
As a result having the same wide and high.The output result of the corresponding up-sampling module of each intensive link block is carried out
Parallel connection, and using the result after parallel connection as the input of next up-sampling module；
After the last one up-samples module, use dimension for 1 full convolutional layer, the input of the full convolutional layer is last
The output of one up-sampling module.
The utility model has the advantages that compared with prior art, the present invention the invention firstly uses symmetrical expressions intensively to connect full convolutional network
The intensive convolutional network layer of symmetrical expression spy carried out with shallow-layer feature to the high level of cloud atlas merge, then symmetrical intensive connection is rolled up entirely
The full convolutional layer of product network divides cloud atlas to cloud layer and non-cloud layer region using fused characteristic information.The present invention is than traditional cloud atlas
Dividing method accuracy rate is higher, robustness is stronger, without carrying out complicated Feature Engineering to cloud atlas, and under the conditions of same hardware
The test sample time also it is faster than most methods very much.
Detailed description of the invention
Fig. 1 is basic flow chart of the invention；
Fig. 2 is the structural schematic diagram that this hair hit symmetrical expression intensively connects the intensive link block of full convolutional network；
Fig. 3 is intensively to connect the structural schematic diagram that full convolutional network carries out cloud Picture with symmetrical expression in the present invention.
Specific embodiment
The present invention is further explained with reference to the accompanying drawings and examples.
One kind as shown in Figure 1 is based on symmetrical expression and intensively connects full convolutional network cloud atlas dividing method, includes the following steps:
S1: symmetrical expression intensively connects the training of full convolution pessimistic concurrency control: setting network model is characterized extractor and full convolution
Two part of layer；Whole network structure is a symmetrical structure, and left side network is characterized extractor in entire symmetrical structure, right
Side is deconvolution network, and the step-length of each layer of deconvolution operation is 2, and it is completely right with left side to can be obtained by characteristic pattern size in this way
Characteristic pattern of the same size, is then attached by the network structure of title, then recycles full convolutional layer output probability figure, finally
Using the sample (Xi, Yi) marked, which is learnt, carries out character representation using symmetrical network structure,
Full convolutional layer carries out supervised learning and output probability figure, the network parameter after being trained, wherein Xi is the figure of a N × N
Picture, Yi indicate the corresponding cloud of Xi and non-cloud region, and N meets 244≤N≤1000, and i represents i-th of sample, i=1,2,3...p, p
For total sample number；
S2: cloud atlas segmentation: be intensively to connect full convolution net as symmetrical for N × N size by ground cloud atlas image cropping
The input data of network, and the characteristic probability figure output of whole network is obtained, finally, decision probability is greater than 0.5 in probability graph
Pixel is the region of cloud, and pixel of the probability less than 0.5 is non-cloud region；
Symmetrical expression intensively connects full convolutional network and includes:
Feature extraction network enhances the character representation of data by symmetrical intensive connection；This feature extracts network packet
Intensive link block is included, the intensive link block is by convolutional neural networks to the two values matrix that label is N*N size
Ground cloud atlas carry out feature extraction, with the convolution kernel window of Wi*Wi size on the picture of N*N size using step-length as sliding, and
The characteristic pattern consistent to size is attached, using the feature of extraction as the input of full convolutional network.
Full convolutional network is used for supervised learning, including up-sampling module, is the feature vector for extracting feature extraction network
As input, full convolutional network exports the probability graph consistent with input picture size；Based on symmetrical intensive connection network knot
Structure carries out feature learning, the network parameter after being trained.
As shown in figure 3, the input of each intensive link block is all previous intensive link block output in network structure
Down-sampling as a result, and the processing result of current intensive link block exported giving next stage intensive link block.
It in the network architecture, is up-sampling identical with intensive link block quantity after the last one intensive link block
Module, the input of each up-sampling module are the output results of previous up-sampling module.
In the network architecture, result of the output of each intensive gang mould block all with up-sampling module is attached.
Each intensive gang mould block has corresponding up-sampling module, and the relationship between them is to export
As a result having the same wide and high.The output result of the corresponding up-sampling module of each intensive link block is carried out
Parallel connection, and using the result after parallel connection as the input of next up-sampling module；
After the last one up-samples module, use dimension for 1 full convolutional layer, the input of full convolutional layer is last
The output of group up-sampling module.
As shown in figure 3, the result that the picture of input is generated by batch standardization, line rectification, convolution, maximum pondization operation
Input as first intensive link block.Then the output result of first intensive link block passes through to criticize and standardizes, is linear whole
Stream, convolution operation are similarly operated and shown in the figure share 4 by average input of the pond layer as second intensive link block
It is secondary.After aforesaid operations, the output result the last one intensive link block is defeated after up-sampling and line rectification function
Result out is in parallel with from the intensive output result of link block of third, is used as on next time and adopts again after convolution twice
The input of sample.Same operation has carried out 4 times as shown in Figure 3.The result of 4th operation passes through S after once up-sampling
Function output just obtains final result.
By multiple intensive link blocks, up-sampling module, the result of full convolution output is inputted sigmoid function, most
The probability graph of output N*N size eventually.Pixel of the probability graph intermediate value greater than 0.5 is judged as cloud, and value is judged as non-less than 0.5 pixel
Cloud obtains segmentation result；
Sigmoid specific algorithm is as follows:
Wherein, fijThe corresponding probability of i row j column pixel in probability graph is represented, w is convolution nuclear parameter, and T indicates that transposition, b indicate
Bias term.
To in probability graph, all probability values set 0 less than 0.5, and probability value sets 1 greater than 0.5, obtains the segmentation knot of cloud atlas
Fruit, including cloud sector domain, non-cloud region, output sample are set as corresponding two-value mask.
The present embodiment, which carries out feature extraction by the symmetrical model structure intensively connected, ensure that network with good
Generalization Capability, the cloud feature being sufficiently extracted in cloud atlas, so that the cloud atlas segmentation result that network model obtains is more accurate.This hair
It is bright to guarantee preferably to merge shallow-layer feature and further feature using the symmetrical network structure intensively connected to realize spy
The diversity of sign.The present invention does not need complicated Feature Engineering, under the premise of accuracy rate improves, the sample under same hardware condition
This test speed and sample decomposition speed all achieves very big raising.The experimental results showed that the obtained result of the present invention compared with
It is good, it is more suitable for subsequent meteorological research work and application.
Table 1 is that the present invention and the experimental result of the other algorithms of tradition compare:
1 present invention of table and the experimental result of the other algorithms of tradition compare
| Method | Accurate rate | Recall rate | F1- score | False Rate |
| Symmetrical expression intensively connects full convolutional network structure | 0.951 | 0.946 | 0.949 | 0.049 |
| FCN32s | 0.923 | 0.922 | 0.923 | 0.082 |
| FCN8s | 0.951 | 0.943 | 0.947 | 0.059 |
| SLIC+DBSCAN | 0.72 | 0.79 | 0.65 | 0.33 |
| GRAY+SVM | 0.87 | 0.56 | 0.64 | 0.33 |
| LBP+SVM | 0.62 | 0.65 | 0.63 | 0.36 |
| ColorHIST+SVM | 0.81 | 0.64 | 0.66 | 0.31 |
| dSIFT+BOW+SVM | 0.65 | 0.88 | 0.72 | 0.28 |
| Texture+BOW+SVM | 0.82 | 0.71 | 0.70 | 0.31 |
| Color-based Segmentation | 0.92 | 0.90 | 0.90 | 0.09 |
To sum up, the method for the present invention is relative to conventional threshold values method, machine learning algorithm and the side FCN based on deep learning
Result acquired by method will be got well, more suitable subsequent meteorological research work and application.
Claims (7)
1. one kind intensively connects full convolutional network cloud atlas segmentation network based on symmetrical expression, it is characterised in that: including symmetrically arranged
Feature extraction network and full convolutional network；
The feature extraction network, for carrying out feature extraction, input of the feature vector of extraction as full convolutional network；The spy
It includes several intensive link blocks for realizing feature extraction, the intensive link block spy that size is consistent that sign, which extracts network,
Sign figure is attached, and carries out feature extraction to ground cloud atlas by convolutional neural networks；
The full convolutional network, including several up-sampling modules, what the up-sampling module was extracted according to feature extraction network
Feature vector obtains convolution output result；Result and formula (1), output probability figure are exported according to convolution；
Wherein, fijThe corresponding probability of i row j column pixel in probability graph is represented, w is convolution nuclear parameter, and T indicates that transposition, b represent biasing
?.
2. a kind of symmetrical expression that is based on according to claim 1 intensively connects full convolutional network cloud atlas segmentation network, feature
Be: the quantity of the intensive link block is consistent with up-sampling module number.
3. a kind of symmetrical expression that is based on according to claim 2 intensively connects full convolutional network cloud atlas segmentation network, feature
Be: the input of each intensive link block is the down-sampling of previous intensive link block output as a result, current intensive
Input of the output of link block as the intensive link block of next stage；
In the input of the last one intensive link block exported as first up-sampling module, each up-sampling module
Input be it is previous up-sampling module output result；
Each intensive link block has corresponding up-sampling module, if the output result of up-sampling module and one close
The output result for collecting link block is having the same wide and high, then the up-sampling module is on the intensive link block is corresponding
Sampling module, by the output result of the corresponding up-sampling module of the result of the output of each intensive link block into
Row is in parallel, and using the parallel connection result as the input of next up-sampling module；
After the last one up-samples module, use dimension for 1 full convolutional layer, the input of the full convolutional layer is the last one
Up-sample the output of module.
4. one kind intensively connects full convolutional network cloud atlas dividing method based on symmetrical expression, it is characterised in that: the following steps are included:
S1: cloud atlas segmentation network is trained: utilizing the sample (X markedi, Yi) be trained, the network after being trained
Parameter, wherein XiFor the image of a N × N, YiIndicate XiCorresponding cloud and non-cloud region, i represent i-th of sample, i=1, and 2,
3...p, p is total sample number；
S2: being N × N size by ground cloud atlas image cropping, as the input data of cloud atlas segmentation network, is obtained according to formula (1)
The probability graph of N*N size:
Wherein, fijThe corresponding probability of i row j column pixel in probability graph is represented, w is convolution nuclear parameter, and T indicates that transposition, b represent biasing
?；
S3: according to cloud sector domain decision threshold, the judgement in cloud sector domain and non-cloud region is carried out to the probability graph of S2 output, when probability is big
In decision threshold pixel as cloud sector domain, otherwise be used as non-cloud region.
5. a kind of symmetrical expression that is based on according to claim 4 intensively connects full convolutional network cloud atlas dividing method, feature
Be: the cloud atlas segmentation network includes symmetrically arranged feature extraction network and deconvolution network；
The feature extraction network, for carrying out feature extraction, input of the feature vector of extraction as deconvolution network；The spy
It includes several intensive link blocks for realizing feature extraction, the intensive link block spy that size is consistent that sign, which extracts network,
Sign figure is attached, and carries out feature extraction to ground cloud atlas by convolutional neural networks；
The full convolutional network, including several up-sampling modules, the up-sampling module obtain convolution output according to feature vector
As a result；It is exported according to convolution as a result, output probability figure.
6. a kind of symmetrical expression that is based on according to claim 5 intensively connects full convolutional network cloud atlas segmentation network, feature
Be: the quantity of the intensive link block is consistent with up-sampling module number.
7. a kind of symmetrical expression that is based on according to claim 6 intensively connects full convolutional network cloud atlas segmentation network, feature
Be: the input of each intensive link block is the down-sampling of previous intensive link block output as a result, current intensive
Input of the output of link block as the intensive link block of next stage；
In the input of the last one intensive link block exported as first up-sampling module, each up-sampling module
Input be it is previous up-sampling module output result；
Each intensive gang mould block has corresponding up-sampling module, and it is right with it that each intensive link block has
The up-sampling module answered, if up-sampling module output result and an intensive link block output result it is having the same width and
Height, then the up-sampling module is the corresponding up-sampling module of the intensive link block, will each intensive link block and
The output result of its corresponding up-sampling module carries out parallel connection, and using the result after parallel connection as the defeated of next up-sampling module
Enter；
After the last one up-samples module, use dimension for 1 full convolutional layer, the input of the full convolutional layer is the last one
Up-sample the output of module.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811569320.3A CN109670508A (en) | 2018-12-21 | 2018-12-21 | A kind of cloud atlas segmentation network and its method intensively connecting full convolutional network based on symmetrical expression |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811569320.3A CN109670508A (en) | 2018-12-21 | 2018-12-21 | A kind of cloud atlas segmentation network and its method intensively connecting full convolutional network based on symmetrical expression |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109670508A true CN109670508A (en) | 2019-04-23 |
Family
ID=66145671
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811569320.3A Pending CN109670508A (en) | 2018-12-21 | 2018-12-21 | A kind of cloud atlas segmentation network and its method intensively connecting full convolutional network based on symmetrical expression |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109670508A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363171A (en) * | 2019-07-22 | 2019-10-22 | 北京百度网讯科技有限公司 | The method of the training method and identification sky areas of sky areas prediction model |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108846474A (en) * | 2018-05-18 | 2018-11-20 | 南京信息工程大学 | The satellite cloud picture cloud amount calculation method of convolutional neural networks is intensively connected based on multidimensional |
-
2018
- 2018-12-21 CN CN201811569320.3A patent/CN109670508A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108846474A (en) * | 2018-05-18 | 2018-11-20 | 南京信息工程大学 | The satellite cloud picture cloud amount calculation method of convolutional neural networks is intensively connected based on multidimensional |
Non-Patent Citations (1)
| Title |
|---|
| GUANZHOU CHEN等: "Symmetrical Dense-Shortcut Deep Fully Convolutional Networks for Semantic Segmentation of Very-High-Resolution Remote Sensing Images", 《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363171A (en) * | 2019-07-22 | 2019-10-22 | 北京百度网讯科技有限公司 | The method of the training method and identification sky areas of sky areas prediction model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106447658B (en) | Conspicuousness object detection method based on global and local convolutional network | |
| CN108960245B (en) | Tire mold character detection and recognition method, device, equipment and storage medium | |
| CN112329658B (en) | Detection algorithm improvement method for YOLOV3 network | |
| CN104834912B (en) | A kind of weather recognition methods and device based on image information detection | |
| US20190228268A1 (en) | Method and system for cell image segmentation using multi-stage convolutional neural networks | |
| CN109635744A (en) | A kind of method for detecting lane lines based on depth segmentation network | |
| CN105701480B (en) | A kind of Video Semantic Analysis method | |
| CN109271960A (en) | A kind of demographic method based on convolutional neural networks | |
| CN109871830A (en) | Spatial-spectral fusion hyperspectral image classification method based on three-dimensional depth residual error network | |
| CN109359681A (en) | A kind of field crop pest and disease disasters recognition methods based on the full convolutional neural networks of improvement | |
| CN108364278B (en) | Rock core fracture extraction method and system | |
| CN109446922B (en) | Real-time robust face detection method | |
| CN111401380B (en) | RGB-D image semantic segmentation method based on depth feature enhancement and edge optimization | |
| CN111723764A (en) | Improved fast RCNN hydroponic vegetable seedling state detection method | |
| CN114037930B (en) | Video action recognition method based on space-time enhanced network | |
| CN110991349B (en) | Lightweight vehicle attribute identification method based on metric learning | |
| CN109741341A (en) | A kind of image partition method based on super-pixel and long memory network in short-term | |
| CN113887517B (en) | Crop remote sensing image semantic segmentation method based on parallel attention mechanism | |
| CN112052772A (en) | Face shielding detection algorithm | |
| CN116052016A (en) | Fine segmentation detection method for remote sensing image cloud and cloud shadow based on deep learning | |
| CN111062381A (en) | License plate position detection method based on deep learning | |
| CN113221913A (en) | Agriculture and forestry disease and pest fine-grained identification method and device based on Gaussian probability decision-level fusion | |
| CN115810191A (en) | Pathological cell classification method based on multi-attention fusion and high-precision segmentation network | |
| CN111563542A (en) | Automatic plant classification method based on convolutional neural network | |
| CN114881286A (en) | Short-time rainfall prediction method based on deep learning |
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 |