CN108629789A - A kind of well-marked target detection method based on VggNet - Google Patents
A kind of well-marked target detection method based on VggNet Download PDFInfo
- Publication number
- CN108629789A CN108629789A CN201810457552.3A CN201810457552A CN108629789A CN 108629789 A CN108629789 A CN 108629789A CN 201810457552 A CN201810457552 A CN 201810457552A CN 108629789 A CN108629789 A CN 108629789A
- Authority
- CN
- China
- Prior art keywords
- image
- convolution
- result
- size
- vggnet
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4007—Interpolation-based scaling, e.g. bilinear interpolation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20172—Image enhancement details
- G06T2207/20192—Edge enhancement; Edge preservation
Abstract
The well-marked target detection method based on VggNet that the invention discloses a kind of, this method are to be characterized as that main target, extraction high-level semantics features are the second target to retain image bottom edges.By the way that image bottom edges feature and high-level semantics features are carried out convolution, and then obtain target significant image.While ensuring to effectively improve Detection accuracy, network training speed is significantly promoted by reducing full articulamentum and increasing normalization layer.By Image Adjusting it is fixed size by linear interpolation, to realize that different size of image can be handled effectively for the RGB image of arbitrary size.The method of the present invention rapid extraction high-level semantics features and can retain image bottom edges information, to effectively solve the problems, such as that the network structure of simple extraction feature class can not obtain image border.
Description
Technical field
The present invention relates to the technical fields of Digital Image Processing, refer in particular to a kind of well-marked target detection based on VggNet
Method.
Background technology
Convolutional neural networks are proposed based on artificial neural network.The nervous system of artificial Neural Network Simulation people, by
A certain number of neurons are constituted.In a supervised learning problem, there is one group of training data (xi,yi), x is sample, and y is class
It does not mark, they is inputted artificial neural network, a nonlinear Optimal Separating Hyperplane h can be obtainedw,b(x), super flat by this
Face can classify all image datas of input.
One neuron is an arithmetic element in neural network, it is substantially exactly a function.As shown in Figure 1,
For the schematic diagram of a neuron:
There are 3 input x1、x2、x3,+1 is a bias, outputF is sharp
Function living, w are the shared proportion of each input, and b is bias, and activation primitive here is sigmoid functions:
Artificial neural network is combined by multiple above-mentioned neurons, as shown in Fig. 2, small-sized artificial for one
The schematic diagram of neural network:
In convolutional neural networks in figure, input is exactly a series of image, and weight w is exactly convolution mask, It is finally the neural network connected entirely for the weight of different neurons, usually convolutional layer and down-sampling layer alternating,
The artificial neural network of namely above-mentioned classics.As shown in figure 3, being a simple convolutional neural networks schematic diagram:
In figure, C is convolutional layer, and S is down-sampling layer, and the output of S4 is pulled into a vector by Stretch layers, inputs tradition god
Through in network N N, then being exported.
VggNet is one kind of convolutional neural networks, it obtain ILSVRC tournament sorting projects in 2014 second place and
The first place of positioning project.Its main contribution is that the depth for showing network is the key component of algorithm excellent performance.
VggNet includes five kinds of structures, wherein best network structure contains 16 convolution/full articulamentum.Network structure very one
It causes, all uses the down-sampling layer of 3 × 3 convolution kernel 2 × 2 from the beginning to the end.Up to the present, VggNet is still extensive
Ground is used for extracting characteristics of image.
Multilayer convolution blur in VggNet network structures image bottom edges information, full articulamentum considerably increase ginseng
Several quantity is unfavorable for carrying out target detection work.In view of the above problems, the present invention proposes that a kind of well-marked target that is suitable for detects
Solution.
Invention content
The shortcomings that it is an object of the invention to overcome the prior art and deficiency, it is proposed that one kind can rapid extraction high level language
Adopted feature, and the well-marked target detection method based on VggNet of image bottom edges information can be retained, it can effectively solve simple
The problem of network structure of extraction feature class can not obtain image border.
To achieve the above object, technical solution provided by the present invention is:A kind of well-marked target detection based on VggNet
Method, this method are characterized as main target to retain image bottom edges, and extraction high-level semantics features are the second target, pass through by
Image bottom edges feature carries out convolution with high-level semantics features, and then obtains target significant image, is ensuring to effectively improve detection
While accuracy rate, network training speed is significantly promoted by reducing full articulamentum and increasing normalization layer, for arbitrarily large
Small RGB image, by linear interpolation by Image Adjusting be fixed size, to realize that different size of image can obtain
Effective processing;It includes the following steps:
1) input picture is pre-processed by linear interpolation, all images is made to meet network inputs requirement;
2) characteristics of image is extracted by level by convolution operation, multiple channels is obtained according to convolution kernel size and step-length respectively
Convolution results;
3) convolution results are handled by down-sampling operation, reduces data manipulation time and space consuming;
4) the case where layer normalized image data being normalized by batch, preventing gradient disperse or gradient explosion;
5) deconvolution operation is carried out to convolution results and image array is restored to original image size, prepare for notable figure;
6) result after deconvolution is subjected to convolution with bottom edges information, obtains final goal notable figure.
In step 1), the RGB image of arbitrary size is adjusted to the image of 224 × 224 sizes using linear interpolation,
Core concept is to carry out once linear interpolation respectively in two directions, and linear interpolation operating process and formula are as follows:
There are four point (x on known image matrix data0,y0)、(x0,y1)、(x1,y0)、(x1,y1),f(x0,y0)、f
(x0,y1)、f(x1,y0)、f(x1,y1) it is respectively the corresponding value of four points;
It is x to abscissa0Y-axis direction carry out linear interpolation calculation formula it is as follows:
In formula, Z1Indicate that result of calculation, v indicate the point for calculating interpolation result from coordinate (x0,y0) y-axis direction distance;
It is x to abscissa1Y-axis direction carry out linear interpolation calculation formula it is as follows:
In formula, Z2For result of calculation, v indicates the point for calculating interpolation result from coordinate (x1,y0) y-axis direction distance;
The calculation formula that linear interpolation is carried out to x-axis direction is as follows:
In formula, Z is final interpolation result, Z2With Z1Result of calculation before respectively, u indicate calculating difference result
Point is from x0X-axis direction distance.
In step 2), the convolution kernel size is divided into two kinds:The convolution kernel of 3 × 3 sizes can capture up and down and
The size of central concept can obtain neighbor information, can also play the effect of nonlinear adjustment data structure, 1 × 1 size
Convolution kernel can realize dimensionality reduction operation as auxiliary convolution kernel by controlling the quantity of convolution kernel.
In step 3), the down-sampling operation is divided into mean value sampling and maximum value sampling, and the mean value sampling is to neighborhood
Interior characteristic point is averaged, and can be reduced estimated value variance caused by Size of Neighborhood is limited and be increased problem;The maximum value sampling
To the characteristic point maximizing in neighborhood, the offset problem that convolutional layer parameter error causes estimation mean value can be solved.
In step 4), the normalized specific effect of batch is to conclude the statistical distribution of unified samples so that institute
There is its mean value of the input signal of sample to be less than setting value close to 0 or compared with its mean square deviation.
In step 5), the deconvolution is the inverse process of convolution operation, the characteristic pattern that deconvolution is obtained with convolution operation
It for input, is calculated, obtains deconvolution as a result, to verify the characteristic pattern for showing that each layer obtains.
In step 6), the meaning that deconvolution result and bottom edges information are carried out to convolution is that multiple convolution is fuzzy
The edge feature of image, the bottom convolution copy advanced for compensate for the deficiency of deconvolution result so that the positioning of target is more
It is accurate to add.
Compared with prior art, the present invention having the following advantages that and advantageous effect:
1, the copy that image edge information is remained in convolution early stage, is conducive to carry out convolution behaviour again when exporting result
Make acquisition target significant image.
2, the full articulamentum in convolutional neural networks is replaced with into convolutional layer, is reducing training parameter, is accelerating the training time
While be more effectively extracted the high-level semantics features of image.
3, addition batch normalization layer solves the problems, such as that gradient disappears and gradient is exploded, and further shortens convergence time.
Description of the drawings
Fig. 1 is the schematic diagram of a neuron in background technology.
Fig. 2 is the schematic diagram of a small-sized artificial neural network in background technology.
Fig. 3 is a simple convolutional neural networks schematic diagram in background technology.
Fig. 4 is schematic network structure used in the method for the present invention.
Fig. 5 is the flow chart of the method for the present invention.
Fig. 6 is the schematic diagram of convolutional calculation process.
Fig. 7 is the input and output case of the method for the present invention.
Specific implementation mode
The present invention is further explained in the light of specific embodiments.
Before introducing the present invention, need that VggNet network structures are introduced.The VggNet network knots of normal mode
Structure includes convolutional layer, down-sampling layer and full articulamentum.The size of convolution kernel is based on 3 × 3 while realizing local receptor field
Also can guarantee effectively reduces network architecture parameters quantity.
The network structure being illustrated in figure 4 used in the method for the present invention, the structure is relative to normal mode VggNet's
Difference lies in the last full articulamentum of network structure is replaced with convolutional layer, batch normalization layer and warp lamination.Cancellation connects entirely
The reason of connecing layer is as follows:The high-level semantics features of image can further be extracted by replacing full articulamentum using convolutional layer, improve inspection
Survey the accuracy of result;Addition batch normalization layer can accelerate network convergence rate, reduce training time and hardware cost;Add
Add warp lamination that intermediate data size is reduced to original image size, is conducive to the generation of final detection result.
The well-marked target detection method based on VggNet that this implementation is provided, it is specifically special to retain image bottom edges
Sign is main target, and extraction high-level semantics features are the second target, by by image bottom edges feature and high-level semantics features
Carry out convolution, and then obtain target significant image, while ensuring to effectively improve Detection accuracy, by reduce full articulamentum and
Increase normalization layer and significantly promote network training speed, for the RGB image of arbitrary size, by linear interpolation by image tune
Whole is fixed size, to realize that different size of image can be handled effectively;It includes the following steps:
1) input picture is pre-processed by linear interpolation, all images is made to meet network inputs requirement;Wherein, sharp
The RGB image of arbitrary size is adjusted to the image of 224 × 224 sizes with linear interpolation, core concept is in both direction
Upper to carry out once linear interpolation respectively, linear interpolation operating process and formula are as follows:
There are four point (x on known image matrix data0,y0)、(x0,y1)、(x1,y0)、(x1,y1),f(x0,y0)、f
(x0,y1)、f(x1,y0)、f(x1,y1) it is respectively the corresponding value of four points;
It is x to abscissa0Y-axis direction carry out linear interpolation calculation formula it is as follows:
In formula, Z1Indicate that result of calculation, v indicate the point for calculating interpolation result from coordinate (x0,y0) y-axis direction distance;
It is x to abscissa1Y-axis direction carry out linear interpolation calculation formula it is as follows:
In formula, Z2For result of calculation, v indicates the point for calculating interpolation result from coordinate (x1,y0) y-axis direction distance;
The calculation formula that linear interpolation is carried out to x-axis direction is as follows:
In formula, Z is final interpolation result, Z2With Z1Result of calculation before respectively, u indicate calculating difference result
Point is from x0X-axis direction distance.
2) characteristics of image is extracted by level by convolution operation, multiple channels is obtained according to convolution kernel size and step-length respectively
Convolution results;Wherein, the convolution kernel size is divided into two kinds:The convolution kernel of 3 × 3 sizes can capture up and down and center
The size of concept can obtain neighbor information, can also play the effect of nonlinear adjustment data structure, the convolution of 1 × 1 size
Core can realize dimensionality reduction operation as auxiliary convolution kernel by controlling the quantity of convolution kernel.
3) convolution results are handled by down-sampling operation, reduces data manipulation time and space consuming;Wherein, institute
It states down-sampling operation and is divided into mean value sampling and maximum value sampling, the mean value sampling averages to characteristic point in neighborhood, can
It reduces estimated value variance caused by Size of Neighborhood is limited and increases problem;The maximum value sampling asks maximum to the characteristic point in neighborhood
Value can solve the offset problem that convolutional layer parameter error causes estimation mean value.
4) the case where layer normalized image data being normalized by batch, preventing gradient disperse or gradient explosion;Its
In, the normalized specific effect of batch is to conclude the statistical distribution of unified samples so that the input signal of all samples
Its mean value is close to 0 or the very little compared with its mean square deviation.
5) deconvolution operation is carried out to convolution results and image array is restored to original image size, prepare for notable figure;
Wherein, the deconvolution is the inverse process of convolution operation, and deconvolution is input with the characteristic pattern that convolution operation obtains, and is counted
It calculates, obtains deconvolution as a result, to verify the characteristic pattern for showing that each layer obtains.
6) result after deconvolution is subjected to convolution with bottom edges information, obtains final goal notable figure;It wherein, will be anti-
The meaning that convolution results carry out convolution with bottom edges information is that multiple convolution has obscured the edge feature of image, advances for
Bottom convolution copy compensate for the deficiency of deconvolution result so that the positioning of target is more accurate.
Concrete operations flow as shown in Figure 5.Training process is the result that propagated forward is combined with backpropagation.Forward direction passes
It is the process that convolutional calculation is carried out using convolution collecting image to broadcast, and during convolutional calculation, convolution kernel KernalW is covered in defeated
Enter on figure InputX, corresponding position quadrature, which sums to obtain one again, is worth and is assigned to the corresponding positions output figure OutputY.Per secondary volume
Product core mobile position on InputX from left to right obtains output matrix after overlapped coverage one time from top to bottom
OutputY (as shown in Figure 6).If the input figure InputX of convolution kernel is m × n sizes, convolution kernel is w × w sizes, and step-length is
1,2 are filled with, then output figure OutputY is (m-w+1) × (n-w+1) sizes.
In error signal back-propagation process, first obtained in the grader of tail portion according to the wrong anti-pass mode of neural network
The error signal of each neuron, then feature extractor propagation of the error signal by grader forwardly.Error signal is adopted from down
The characteristic pattern of sample layer forward propagate and to be completed by primary full convolution process by the characteristic pattern of convolutional layer.
Convolutional neural networks are actually a flight data recorder, it would be desirable to according to the result of network output to network middle layer
Parameter is adjusted to obtain better feature extraction or classifying quality.Therefore training set and test set are just produced, instructed
The effect for practicing collection is for training neural network to carry out feature extraction, can simply be interpreted as finding and predicting potential relationship
Mass data (being typically a large amount of pictures).The effect of test set is for carrying out an accuracy to trained neural network
Test and measurement, test set are independently of training data but follow one group of data of probability distribution identical with training data.
Common test method includes to reserve method and cross-validation method.
As its name suggests, the operating process for reserving method is to select p element in complete or collected works X as testing, then remaining n-p
A element is as training set.It can be obtained according to theorem mathematically, the selection method of p element hasIt is a,
Middle n!Indicate the factorial of n.Under this meaning, it is also very high to stay the time complexity that p is verified.When p=1,1 is stayed
The complexity of verification is precisely n.
Next, by taking k-fold as an example, cross-validation method is simply introduced.In k-fold cross validations, complete or collected works' X quilts
The random set A for being divided into k same sizes1,...,Ak, and | A1|=...=| Ak|.Here | Ai| it refers to gathering
AiElement number, that is, the gesture gathered.I is needed to be traversed for this whens from 1 to k, AiGather as test, other set
Gather as training.According to the test statistics of model, A can be obtainediThe fruiting quantities n of test errors in seti.If complete or collected works X
Gesture be n if, can obtain the error rate of the model isIn order to improve the accuracy of model, k-fold can be handed over
The above-mentioned steps of fork verification repeat t times, are all random division complete or collected works X each time.In being tested at t times, the mistake of t model can be obtained
Accidentally rate E1,...,Et.The error rate of the model in this way is exactly:
Wherein, it inputs shown in Figure 7 with the image effect of output.
In conclusion the present invention be substantially generated by following three kinds of modes it is a kind of for carrying out the new of well-marked target detection
Type network:
1) copy that image edge information is remained in convolution early stage, is conducive to carry out convolution behaviour again when exporting result
Make acquisition target significant image.
2) the full articulamentum in convolutional neural networks is replaced with into convolutional layer, is reducing training parameter, is accelerating the training time
While be more effectively extracted the high-level semantics features of image.
3) addition batch normalization layer solves the problems, such as that gradient disappears and gradient is exploded, and further shortens convergence time.
The examples of implementation of the above are only the preferred embodiments of the invention, and the implementation model of the present invention is not limited with this
It encloses, therefore changes made by all shapes according to the present invention, principle, should all cover within the scope of the present invention.
Claims (7)
1. a kind of well-marked target detection method based on VggNet, it is characterised in that:This method is special to retain image bottom edges
Sign is main target, and extraction high-level semantics features are the second target, by by image bottom edges feature and high-level semantics features
Carry out convolution, and then obtain target significant image, while ensuring to effectively improve Detection accuracy, by reduce full articulamentum and
Increase normalization layer and significantly promote network training speed, for the RGB image of arbitrary size, by linear interpolation by image tune
Whole is fixed size, to realize that different size of image can be handled effectively;It includes the following steps:
1) input picture is pre-processed by linear interpolation, all images is made to meet network inputs requirement;
2) characteristics of image is extracted by level by convolution operation, obtains the volume in multiple channels respectively according to convolution kernel size and step-length
Product result;
3) convolution results are handled by down-sampling operation, reduces data manipulation time and space consuming;
4) the case where layer normalized image data being normalized by batch, preventing gradient disperse or gradient explosion;
5) deconvolution operation is carried out to convolution results and image array is restored to original image size, prepare for notable figure;
6) result after deconvolution is subjected to convolution with bottom edges information, obtains final goal notable figure.
2. a kind of well-marked target detection method based on VggNet according to claim 1, it is characterised in that:In step 1)
In, the RGB image of arbitrary size is adjusted to the image of 224 × 224 sizes using linear interpolation, core concept is at two
Once linear interpolation is carried out on direction respectively, linear interpolation operating process and formula are as follows:
There are four point (x on known image matrix data0,y0)、(x0,y1)、(x1,y0)、(x1,y1),f(x0,y0)、f(x0,
y1)、f(x1,y0)、f(x1,y1) it is respectively the corresponding value of four points;
It is x to abscissa0Y-axis direction carry out linear interpolation calculation formula it is as follows:
In formula, Z1Indicate that result of calculation, v indicate the point for calculating interpolation result from coordinate (x0,y0) y-axis direction distance;
It is x to abscissa1Y-axis direction carry out linear interpolation calculation formula it is as follows:
In formula, Z2For result of calculation, v indicates the point for calculating interpolation result from coordinate (x1,y0) y-axis direction distance;
The calculation formula that linear interpolation is carried out to x-axis direction is as follows:
In formula, Z is final interpolation result, Z2With Z1Result of calculation before respectively, u indicate calculating difference result point from
x0X-axis direction distance.
3. a kind of well-marked target detection method based on VggNet according to claim 1, it is characterised in that:In step 2)
In, the convolution kernel size is divided into two kinds:The convolution kernel of 3 × 3 sizes can capture the size with central concept up and down, energy
Neighbor information is enough obtained, can also play the effect of nonlinear adjustment data structure, the convolution kernel of 1 × 1 size is as secondary volume
Product core can realize dimensionality reduction operation by controlling the quantity of convolution kernel.
4. a kind of well-marked target detection method based on VggNet according to claim 1, it is characterised in that:In step 3)
In, the down-sampling operation is divided into mean value sampling and maximum value sampling, and the mean value sampling averages to characteristic point in neighborhood,
Estimated value variance caused by Size of Neighborhood is limited can be reduced and increase problem;The characteristic point in neighborhood is sought in the maximum value sampling
Maximum value can solve the offset problem that convolutional layer parameter error causes estimation mean value.
5. a kind of well-marked target detection method based on VggNet according to claim 1, it is characterised in that:In step 4)
In, the normalized specific effect of batch is to conclude the statistical distribution of unified samples so that the input signal of all samples
Its mean value is less than setting value close to 0 or compared with its mean square deviation.
6. a kind of well-marked target detection method based on VggNet according to claim 1, it is characterised in that:In step 5)
In, the deconvolution is the inverse process of convolution operation, and deconvolution is input with the characteristic pattern that convolution operation obtains, and is calculated,
Deconvolution is obtained as a result, to verify the characteristic pattern for showing that each layer obtains.
7. a kind of well-marked target detection method based on VggNet according to claim 1, it is characterised in that:In step 6)
In, the meaning that deconvolution result and bottom edges information are carried out to convolution is that multiple convolution has obscured the edge feature of image,
The bottom convolution copy advanced for compensates for the deficiency of deconvolution result so that the positioning of target is more accurate.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810457552.3A CN108629789A (en) | 2018-05-14 | 2018-05-14 | A kind of well-marked target detection method based on VggNet |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810457552.3A CN108629789A (en) | 2018-05-14 | 2018-05-14 | A kind of well-marked target detection method based on VggNet |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108629789A true CN108629789A (en) | 2018-10-09 |
Family
ID=63693062
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810457552.3A Pending CN108629789A (en) | 2018-05-14 | 2018-05-14 | A kind of well-marked target detection method based on VggNet |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108629789A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109919059A (en) * | 2019-02-26 | 2019-06-21 | 四川大学 | Conspicuousness object detecting method based on depth network layerization and multitask training |
CN110097110A (en) * | 2019-04-26 | 2019-08-06 | 华南理工大学 | A kind of semantic image restorative procedure based on objective optimization |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103914675A (en) * | 2014-03-17 | 2014-07-09 | 东华大学 | Garment QD code recognition method |
CN104834898A (en) * | 2015-04-09 | 2015-08-12 | 华南理工大学 | Quality classification method for portrait photography image |
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN106447658A (en) * | 2016-09-26 | 2017-02-22 | 西北工业大学 | Significant target detection method based on FCN (fully convolutional network) and CNN (convolutional neural network) |
CN106599883A (en) * | 2017-03-08 | 2017-04-26 | 王华锋 | Face recognition method capable of extracting multi-level image semantics based on CNN (convolutional neural network) |
CN107016409A (en) * | 2017-03-20 | 2017-08-04 | 华中科技大学 | A kind of image classification method and system based on salient region of image |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
-
2018
- 2018-05-14 CN CN201810457552.3A patent/CN108629789A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103914675A (en) * | 2014-03-17 | 2014-07-09 | 东华大学 | Garment QD code recognition method |
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN104834898A (en) * | 2015-04-09 | 2015-08-12 | 华南理工大学 | Quality classification method for portrait photography image |
CN106447658A (en) * | 2016-09-26 | 2017-02-22 | 西北工业大学 | Significant target detection method based on FCN (fully convolutional network) and CNN (convolutional neural network) |
CN106599883A (en) * | 2017-03-08 | 2017-04-26 | 王华锋 | Face recognition method capable of extracting multi-level image semantics based on CNN (convolutional neural network) |
CN107016409A (en) * | 2017-03-20 | 2017-08-04 | 华中科技大学 | A kind of image classification method and system based on salient region of image |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
Non-Patent Citations (2)
Title |
---|
JONATHAN LONG: "Fully Convolutional Networks for Semantic Segmentation", 《HTTPS://WWW.CV-FOUNDATION.ORG/.../LONG_FULLY_CONVOLUTIONAL_NETWORKS_2015_CVPR_PAPER.PDF》 * |
郭炜强等: "基于贝叶斯模型的专利分类", 《计算机工程与设计》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109919059A (en) * | 2019-02-26 | 2019-06-21 | 四川大学 | Conspicuousness object detecting method based on depth network layerization and multitask training |
CN110097110A (en) * | 2019-04-26 | 2019-08-06 | 华南理工大学 | A kind of semantic image restorative procedure based on objective optimization |
CN110097110B (en) * | 2019-04-26 | 2021-07-20 | 华南理工大学 | Semantic image restoration method based on target optimization |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111368896B (en) | Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network | |
CN110443818B (en) | Graffiti-based weak supervision semantic segmentation method and system | |
CN105528638B (en) | The method that gray relative analysis method determines convolutional neural networks hidden layer characteristic pattern number | |
CN109886286A (en) | Object detection method, target detection model and system based on cascade detectors | |
CN104517122A (en) | Image target recognition method based on optimized convolution architecture | |
CN109145713A (en) | A kind of Small object semantic segmentation method of combining target detection | |
CN106709511A (en) | Urban rail transit panoramic monitoring video fault detection method based on depth learning | |
CN107871119A (en) | A kind of object detection method learnt based on object space knowledge and two-stage forecasting | |
CN109949255A (en) | Image rebuilding method and equipment | |
CN109934115A (en) | Construction method, face identification method and the electronic equipment of human face recognition model | |
CN107909015A (en) | Hyperspectral image classification method based on convolutional neural networks and empty spectrum information fusion | |
CN107529650A (en) | The structure and closed loop detection method of network model, related device and computer equipment | |
CN107092870A (en) | A kind of high resolution image semantics information extracting method and system | |
CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
CN114937151A (en) | Lightweight target detection method based on multi-receptive-field and attention feature pyramid | |
CN107423747A (en) | A kind of conspicuousness object detection method based on depth convolutional network | |
Xu et al. | Robust self-ensembling network for hyperspectral image classification | |
US11941865B2 (en) | Hyperspectral image classification method based on context-rich networks | |
CN113807188A (en) | Unmanned aerial vehicle target tracking method based on anchor frame matching and Simese network | |
CN110334584A (en) | A kind of gesture identification method based on the full convolutional network in region | |
CN108629789A (en) | A kind of well-marked target detection method based on VggNet | |
CN115810157A (en) | Unmanned aerial vehicle target detection method based on lightweight feature fusion | |
CN107564007A (en) | The scene cut modification method and system of amalgamation of global information | |
CN114863266A (en) | Land use classification method based on deep space-time mode interactive network | |
CN109558803B (en) | SAR target identification method based on convolutional neural network and NP criterion |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181009 |
|
WD01 | Invention patent application deemed withdrawn after publication |