CN108921196A - A kind of semantic segmentation method for improving full convolutional neural networks - Google Patents
A kind of semantic segmentation method for improving full convolutional neural networks Download PDFInfo
- Publication number
- CN108921196A CN108921196A CN201810558048.2A CN201810558048A CN108921196A CN 108921196 A CN108921196 A CN 108921196A CN 201810558048 A CN201810558048 A CN 201810558048A CN 108921196 A CN108921196 A CN 108921196A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- training
- convolutional neural
- full convolutional
- semantic segmentation
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the semantic segmentation method for improving full convolutional neural networks, including step:Obtain training image data;Training image data are inputted into porous full convolutional neural networks, Standard convolution pond layer is first passed through and obtains the characteristic pattern of size reduction;Extract denser feature while maintaining characteristic pattern size by porous convolutional layer again;Prediction pixel-by-pixel finally is carried out to characteristic pattern and obtains segmentation result;And the parameter in porous full convolutional neural networks is trained using stochastic gradient descent method SGD in training;It obtains and the image of semantic segmentation is needed to input the porous full convolutional neural networks after training, obtain corresponding semantic segmentation result.The present invention, which can improve, finally up-samples the problem of characteristic pattern of recovery loses the details sensibility to image in full convolutional network, while under the premise of not increasing number of parameters and calculation amount, effectively expanding the receptive field of filter.
Description
Technical field
The present invention relates to a kind of semantic segmentation methods based on porous full convolutional neural networks, belong to the neck of computer vision
Domain.
Background technique
Image, semantic segmentation is the key technique of image understanding, is widely used in automated driving system (specially streetscape
Identification and understand), unmanned plane application (landing point judgement) and wearable device in.Image, semantic segmentation is realized to institute in picture
There is the classification of pixel.Before deep learning is applied to image, semantic segmentation, there are threshold method, the base of simplest pixel scale
In a variety of methods such as the split plot designs that the split plot design and figure of pixel cluster divide.Shi etc. is proposed based on figure division
The connection weight of the different piece of segmentation and full figure node is taken into account to reach by Normalized cut (N-cut) method
To the purpose for considering global information.Rother etc. proposes the Grab cut method equally divided based on figure, this is a kind of interaction
The semantic segmentation method of formula, using the texture information and boundary information in image, so that only needing a small amount of user interactive operation
Preferable contexts segmentation result just can be obtained.These methods mostly be according to the low order visual information of image pixel itself come into
Row image segmentation.Due to there is no the algorithm training stage, although computation complexity is not high, there is higher segmentation error rate.
In recent years, the fast development of deep learning has greatly pushed the progress of semantic segmentation.Dan etc. is based on deep learning
Image block classification method is proposed, i.e., independent classification is carried out to each pixel using the image block around pixel, at that time
Convolutional network end is usually using full articulamentum, it is therefore desirable to which fixed-size image carries out segmentation pixel-by-pixel.Long etc.
It proposes full convolutional neural networks and carries out pixel scale semantic segmentation end to end, also pole while receiving different sized images
The earth improves the efficiency of segmentation, but segmentation effect is still not fine enough, while being difficult to carry out various sizes of same object
Correctly segmentation.Papandreou etc. proposes the method that multi-scale image input network is finally integrated output feature, improves
To the adaptability of multi-scale image.
In conclusion complying with development trend, characteristics of image is extracted using convolutional neural networks.However wherein there are still some
Problem values must be studied, and take into account network weight and feature extraction effect such as the design of convolutional neural networks structure, loss function
Design preferably to carry out semantic segmentation task.
Summary of the invention
It is a kind of based on the full convolution of improvement technical problem to be solved by the present invention lies in overcoming the deficiencies of the prior art and provide
The semantic segmentation method of neural network solves in existing full convolutional network since the maximum pondization of continuous several times and down-sampling operate
It causes feature resolution to drastically reduce, makes finally to up-sample the problem of characteristic pattern restored loses the details sensibility to image.
The present invention realizes that the technical solution of above-mentioned purpose is:A kind of semantic segmentation side improving full convolutional neural networks
Method, it is characterised in that include the following steps:
Step 1 obtains training image data;
The porous full convolutional neural networks that the input reconstruction of training image data obtains are trained by step 2;
Step 3, acquisition need the image data of semantic segmentation, and are input to the porous full convolutional neural networks after training,
Obtain corresponding semantic segmentation result.
Further, as a preferred technical solution of the present invention:The process of the training is:First pass around three marks
Quasi- convolution pond module carries out the denser feature of porous convolution extraction later and obtains the shot chart that size becomes smaller, then to
Component bilinear interpolation obtains original image size characteristic pattern and makees enhanced processing, is finally predicted to obtain segmentation knot pixel-by-pixel
Fruit;And utilize stochastic gradient descent method SGD to the parameter training in full convolutional neural networks FCN in training.
Further, as a preferred technical solution of the present invention:To gained original image size in training process
Characteristic pattern enhanced processing is followed successively by convolution, batch normalization and deconvolution processing.
Further, as a preferred technical solution of the present invention:Stochastic gradient descent method is utilized in training process
The loss function established needed for SGD training:
Wherein, L (x) represents loss function target value;It is described Semantic segmentation for network output predicts knot
Fruit, y are the legitimate reading in training image data set;Parameter c value isI is in image
Pixel coordinate, and x ∈ (- c, c).
Further, as a preferred technical solution of the present invention:Step 1 further includes using data extending method to instruction
Practice image sample data to expand.
Further, as a preferred technical solution of the present invention:The data extending method, which is at least, to be rotated, contracts
It puts, Random Level overturning or translation are handled.
Further, as a preferred technical solution of the present invention:Step 1 further includes to acquired training image data
Subtract the pre-treatment step of itself pixel average.
Semantic segmentation method provided by the invention based on the full convolutional neural networks of improvement substitutes full volume using porous convolution
Standard convolution in product network, so that convolutional network can accurately control the resolution ratio of image when calculating characteristic response,
Simultaneously under the premise of not increasing number of parameters and calculation amount, the receptive field of filter is effectively expanded, makes the image include
Into more contexts, characteristic pattern details is enriched, improves the precision of semantic segmentation.
Detailed description of the invention
Fig. 1 is that the present invention is based on the schematic illustrations of the semantic segmentation method of porous full convolutional network.
Specific embodiment
Just attached drawing in conjunction with the embodiments below, the embodiment of the present invention is described in further detail, so that of the invention
Technical solution is more readily understood, grasps, and relatively sharp defining and supporting to make to protection scope of the present invention.
As shown in Figure 1, the present invention devises a kind of semantic segmentation method for improving full convolutional network, it is based on full convolutional Neural
Network is improved to obtain porous full convolutional network and be trained end to end to it.This method specifically includes following steps:
Step 1 obtains training image data.
Since network layer is more deep, the parameter amount of required training is more, so the amount of training data needs for needing to prepare
Reach the requirement of a certain amount grade.2012 data set of PASCAL VOC is selected, data set is divided into raw data set and enhancing data set
Two parts all include 20 foreground object classifications and a background classification.The application is used for the training stage for data set is enhanced,
11355 pictures are contained, 8498 therein for training;There are 17125 pictures under raw data set, chooses segmentation classification
Under 762 pictures be used as experiment verifying.
It is described that training image sample data is trained process is specific as follows:Maintenance data extending method, i.e. small range
Rotation, scaling, Random Level overturning processing.The method of data extending can expand image pattern amount and increase the multiplicity of image
Property, so that the network model that training obtains has stronger robustness.It preferably, can also include preprocess method for sample graph
As data set subtracts its pixel average.
Step 2, firstly, establish the frame structure based on porous full convolutional network, basic network be ImageNet contest in
Last three therein full articulamentums are changed to Standard convolution layer and constitute full convolutional network by preceding 13 layers of convolutional layer of VGG network, and
Part of Standard convolution is changed to porous convolution.Whole network shares 5 pond layers, passes through several convolutional layers between the layer of pond
With active coating, discuss that Fig. 1 eliminates activation primitive for convenience.It include convolutional layer, batch normalization layer and anti-during fusion
Convolution.
Secondly, the porous full convolutional network that acquired training image data input is established is trained, process is such as
Under:
Training image data are inputted porous full convolutional network by step 21, successively obtain the defeated of each layer by each pond layer
Characteristic image out, i.e.,:Using pond layer as boundary, the pond layer that each step-length is 2 exports characteristic image having a size of its previous pond layer
The 1/2 of characteristic image size is exported, characteristic image size is exported after three pond layers and is reduced to the 1/ of input image size
8.Most latter two pond layer step-length, which is set as 1, prevents characteristic image resolution ratio from further reducing.First four convolution pond module
Port number is respectively 64,128,256 and 512.Convolution kernel size in network in convolutional layer is all made of 3 × 3 size, active coating
Use ReLu activation primitive.
Standard convolution in step 22, a convolutional layer by the 5th convolution pond module and later is changed to porous volume
Product improves pond step-length with this and is changed to the problem of 1 bring filter receptive field reduces.Porous convolution output signal y [i] definition
It is as follows:
Wherein, x [i] is one-dimensional input signal, and w [k] is filter, and K is its length, and proportionality coefficient r has corresponded to input letter
Number sampling step length, as r=1, i.e. Standard convolution.If the sampling step length of the two porous convolution is respectively 2 and 4, output is logical
Road number is 512 and 4096.
Step 23, later image pass through most latter two convolutional layer, and output channel number is 4096 and 21.The two convolutional layers it
It is preceding to have one Dropout layers respectively.The Dropout layers of weight for allowing the certain hidden layer nodes of network at random in model training not work
Make, those idle nodes temporarily not think be network structure a part, but retain its weight for next sample input
When resume work and prepare, effectively prevent model overfitting problem.Finally original image size is obtained by the deconvolution that step-length is 8
Segmentation result.
Step 24 simultaneously utilizes stochastic gradient descent method SGD to the parameter training in porous full convolutional neural networks.Network
Training is completed by stochastic gradient descent method SGD.Data set is ready for completing by step 1, in training by entire data
Collection subtracts its average value and inputs network access network again.The initialization of progress network weight, base net network fractional weight number are needed before training
Value is initialized by VGG16 network, and is jumped layer and initialized by random number, and it is 0 that probability distribution, which obeys mean value, side
The normal distribution that difference is 0.01.Warp lamination is initialized as bilinear interpolation.The loss function used required for training is as follows:
Wherein in training, L (x) represents loss function target value;In the loss function It is defeated for network
Semantic segmentation prediction result out, y are the legitimate reading in training image data set, and loss function measures network and exports prognostic chart
The difference of the standard picture provided when as with for training, that is, need L (x) gradually to restrain during training;Parameter c value
ForI is pixel coordinate in image, that is, acts on a collection of image pixel in entire training, and x ∈
(- c, c).As x ∈ (- c, c), loss function is equivalent to L1 form (hinge loss), conversely, loss function is equivalent to L2 shape
Formula (squared hinge loss).
Batch size is set as 16 and 20 periods of training when training.All layers of initial learning rate is set as in network
0.01, it is gradually reduced after 6 to 8 periods of training later, for example become 0.1 times of a preceding learning rate, until training is complete
20 periods.
Step 3, acquisition need the image of semantic segmentation to input the porous full convolutional neural networks after training, obtain corresponding
Semantic segmentation prediction result, the porous full convolutional neural networks after input training, it preferably includes picture is carried out to subtract mean value
Operation just inputs in network access network later;Finally, obtain and export corresponding semantic segmentation prediction result.
To sum up, the present invention is for the feature due to caused by duplicate maximum pondization and down-sampling operation in full convolutional network
The problem of resolution ratio drastically reduces substitutes the part of standards convolution in full convolutional network using porous convolution, so that convolution
Network can accurately control the resolution ratio of image when calculating characteristic response, while not increase number of parameters and calculation amount
Under the premise of, the receptive field of filter is effectively expanded, making image includes to enrich characteristic pattern into more contexts
Details improves segmentation effect.
Embodiments of the present invention are explained in detail above in conjunction with attached drawing, but the present invention is not limited to above-mentioned implementations
Mode within the knowledge of a person skilled in the art can also be without departing from the purpose of the present invention
It makes a variety of changes.
Claims (7)
1. a kind of semantic segmentation method for improving full convolutional neural networks, it is characterised in that include the following steps:
Step 1 obtains training image data;
The porous full convolutional neural networks that the input reconstruction of training image data obtains are trained by step 2;
Step 3, acquisition need the image data of semantic segmentation, and are input to the porous full convolutional neural networks after training, obtain
Corresponding semantic segmentation result.
2. improving the semantic segmentation method of full convolutional neural networks according to claim 1, it is characterised in that:The training
Process is:
Three Standard convolution pond modules are first passed around, carry out the denser feature of porous convolution extraction later and obtain size becoming smaller
Shot chart, original image size characteristic pattern then is obtained to shot chart bilinear interpolation and makees enhanced processing, finally carry out by
Pixel prediction obtains segmentation result;And utilize stochastic gradient descent method SGD to the parameter in full convolutional neural networks FCN in training
Training.
3. improving the semantic segmentation method of full convolutional neural networks according to claim 2, it is characterised in that:In training process
Convolution, batch normalization and deconvolution processing are followed successively by gained original image size characteristic pattern enhanced processing.
4. improving the semantic segmentation method of full convolutional neural networks according to claim 2, it is characterised in that:In training process
Utilize the loss function established needed for stochastic gradient descent method SGD training:
Wherein, L (x) represents loss function target value;It is described For the semantic segmentation prediction result of network output, y is
Legitimate reading in training image data set;Parameter c value isI is that pixel is sat in image
Mark, and x ∈ (- c, c).
5. improving the semantic segmentation method of full convolutional neural networks according to claim 1, it is characterised in that:Step 1 is also wrapped
It includes and training image sample data is expanded using data extending method.
6. improving the semantic segmentation method of full convolutional neural networks according to claim 5, it is characterised in that:The data expand
It fills method and is at least rotation, scaling, Random Level overturning or translation processing.
7. improving the semantic segmentation method of full convolutional neural networks according to claim 1, it is characterised in that:Step 1 is also wrapped
Include the pre-treatment step that acquired training image data are subtracted with itself pixel average.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810558048.2A CN108921196A (en) | 2018-06-01 | 2018-06-01 | A kind of semantic segmentation method for improving full convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810558048.2A CN108921196A (en) | 2018-06-01 | 2018-06-01 | A kind of semantic segmentation method for improving full convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108921196A true CN108921196A (en) | 2018-11-30 |
Family
ID=64420202
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810558048.2A Pending CN108921196A (en) | 2018-06-01 | 2018-06-01 | A kind of semantic segmentation method for improving full convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108921196A (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109754362A (en) * | 2018-12-24 | 2019-05-14 | 哈尔滨工程大学 | A method of sea cucumber object detection results are marked with rotatable bounding box |
CN109784476A (en) * | 2019-01-12 | 2019-05-21 | 福州大学 | A method of improving DSOD network |
CN109840914A (en) * | 2019-02-28 | 2019-06-04 | 华南理工大学 | A kind of Texture Segmentation Methods based on user's interactive mode |
CN109903303A (en) * | 2019-02-25 | 2019-06-18 | 秦皇岛燕大滨沅科技发展有限公司 | A kind of drauht line drawing method based on convolutional neural networks |
CN109948517A (en) * | 2019-03-18 | 2019-06-28 | 长沙理工大学 | A kind of high-resolution remote sensing image semantic segmentation method based on intensive full convolutional network |
CN110070091A (en) * | 2019-04-30 | 2019-07-30 | 福州大学 | The semantic segmentation method and system rebuild based on dynamic interpolation understood for streetscape |
CN110458841A (en) * | 2019-06-20 | 2019-11-15 | 浙江工业大学 | A method of improving image segmentation operating rate |
CN111080650A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting looseness and loss faults of small part bearing blocking key nut of railway wagon |
CN111161340A (en) * | 2019-12-02 | 2020-05-15 | 上海交通大学 | Image redirection method and system based on depth feature extraction and terminal |
CN111489364A (en) * | 2020-04-08 | 2020-08-04 | 重庆邮电大学 | Medical image segmentation method based on lightweight full convolution neural network |
CN111488887A (en) * | 2020-04-09 | 2020-08-04 | 腾讯科技(深圳)有限公司 | Image processing method and device based on artificial intelligence |
CN111626305A (en) * | 2019-02-28 | 2020-09-04 | 阿里巴巴集团控股有限公司 | Target detection method, device and equipment |
CN111858058A (en) * | 2020-07-24 | 2020-10-30 | 成都成信高科信息技术有限公司 | SGD load balancing method and device based on parallel computing and storage medium |
CN112418228A (en) * | 2020-11-02 | 2021-02-26 | 暨南大学 | Image semantic segmentation method based on multi-feature fusion |
CN112465826A (en) * | 2019-09-06 | 2021-03-09 | 上海高德威智能交通系统有限公司 | Video semantic segmentation method and device |
CN113449743A (en) * | 2021-07-12 | 2021-09-28 | 西安科技大学 | Coal dust particle feature extraction method |
CN113705326A (en) * | 2021-07-02 | 2021-11-26 | 重庆交通大学 | Urban construction land identification method based on full convolution neural network |
CN115019038A (en) * | 2022-05-23 | 2022-09-06 | 杭州缦图摄影有限公司 | Similar image pixel-level semantic matching method |
CN115223154A (en) * | 2022-07-20 | 2022-10-21 | 中国科学院长春光学精密机械与物理研究所 | Target detection system and detection method thereof |
CN115019038B (en) * | 2022-05-23 | 2024-04-30 | 杭州海马体摄影有限公司 | Similar image pixel level semantic matching method |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107527069A (en) * | 2017-08-22 | 2017-12-29 | 京东方科技集团股份有限公司 | Image processing method, device, electronic equipment and computer-readable medium |
CN107578436A (en) * | 2017-08-02 | 2018-01-12 | 南京邮电大学 | A kind of monocular image depth estimation method based on full convolutional neural networks FCN |
CN107688783A (en) * | 2017-08-23 | 2018-02-13 | 京东方科技集团股份有限公司 | 3D rendering detection method, device, electronic equipment and computer-readable medium |
CN107784654A (en) * | 2016-08-26 | 2018-03-09 | 杭州海康威视数字技术股份有限公司 | Image partition method, device and full convolutional network system |
US9953236B1 (en) * | 2017-03-10 | 2018-04-24 | TuSimple | System and method for semantic segmentation using dense upsampling convolution (DUC) |
-
2018
- 2018-06-01 CN CN201810558048.2A patent/CN108921196A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107784654A (en) * | 2016-08-26 | 2018-03-09 | 杭州海康威视数字技术股份有限公司 | Image partition method, device and full convolutional network system |
US9953236B1 (en) * | 2017-03-10 | 2018-04-24 | TuSimple | System and method for semantic segmentation using dense upsampling convolution (DUC) |
CN107578436A (en) * | 2017-08-02 | 2018-01-12 | 南京邮电大学 | A kind of monocular image depth estimation method based on full convolutional neural networks FCN |
CN107527069A (en) * | 2017-08-22 | 2017-12-29 | 京东方科技集团股份有限公司 | Image processing method, device, electronic equipment and computer-readable medium |
CN107688783A (en) * | 2017-08-23 | 2018-02-13 | 京东方科技集团股份有限公司 | 3D rendering detection method, device, electronic equipment and computer-readable medium |
Non-Patent Citations (1)
Title |
---|
代具亭: "基于深度学习的语义分割网络", 《红外》 * |
Cited By (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109754362A (en) * | 2018-12-24 | 2019-05-14 | 哈尔滨工程大学 | A method of sea cucumber object detection results are marked with rotatable bounding box |
CN109784476A (en) * | 2019-01-12 | 2019-05-21 | 福州大学 | A method of improving DSOD network |
CN109784476B (en) * | 2019-01-12 | 2022-08-16 | 福州大学 | Method for improving DSOD network |
CN109903303A (en) * | 2019-02-25 | 2019-06-18 | 秦皇岛燕大滨沅科技发展有限公司 | A kind of drauht line drawing method based on convolutional neural networks |
CN111626305A (en) * | 2019-02-28 | 2020-09-04 | 阿里巴巴集团控股有限公司 | Target detection method, device and equipment |
CN109840914A (en) * | 2019-02-28 | 2019-06-04 | 华南理工大学 | A kind of Texture Segmentation Methods based on user's interactive mode |
CN109840914B (en) * | 2019-02-28 | 2022-12-16 | 华南理工大学 | Texture segmentation method based on user interaction |
CN111626305B (en) * | 2019-02-28 | 2023-04-18 | 阿里巴巴集团控股有限公司 | Target detection method, device and equipment |
CN109948517A (en) * | 2019-03-18 | 2019-06-28 | 长沙理工大学 | A kind of high-resolution remote sensing image semantic segmentation method based on intensive full convolutional network |
CN110070091A (en) * | 2019-04-30 | 2019-07-30 | 福州大学 | The semantic segmentation method and system rebuild based on dynamic interpolation understood for streetscape |
CN110070091B (en) * | 2019-04-30 | 2022-05-24 | 福州大学 | Semantic segmentation method and system based on dynamic interpolation reconstruction and used for street view understanding |
CN110458841A (en) * | 2019-06-20 | 2019-11-15 | 浙江工业大学 | A method of improving image segmentation operating rate |
CN112465826B (en) * | 2019-09-06 | 2023-05-16 | 上海高德威智能交通系统有限公司 | Video semantic segmentation method and device |
CN112465826A (en) * | 2019-09-06 | 2021-03-09 | 上海高德威智能交通系统有限公司 | Video semantic segmentation method and device |
CN111161340B (en) * | 2019-12-02 | 2023-06-02 | 上海交通大学 | Image redirection method and system based on depth feature extraction and terminal |
CN111161340A (en) * | 2019-12-02 | 2020-05-15 | 上海交通大学 | Image redirection method and system based on depth feature extraction and terminal |
CN111080650A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting looseness and loss faults of small part bearing blocking key nut of railway wagon |
CN111489364B (en) * | 2020-04-08 | 2022-05-03 | 重庆邮电大学 | Medical image segmentation method based on lightweight full convolution neural network |
CN111489364A (en) * | 2020-04-08 | 2020-08-04 | 重庆邮电大学 | Medical image segmentation method based on lightweight full convolution neural network |
CN111488887A (en) * | 2020-04-09 | 2020-08-04 | 腾讯科技(深圳)有限公司 | Image processing method and device based on artificial intelligence |
CN111488887B (en) * | 2020-04-09 | 2023-04-18 | 腾讯科技(深圳)有限公司 | Image processing method and device based on artificial intelligence |
CN111858058A (en) * | 2020-07-24 | 2020-10-30 | 成都成信高科信息技术有限公司 | SGD load balancing method and device based on parallel computing and storage medium |
CN112418228A (en) * | 2020-11-02 | 2021-02-26 | 暨南大学 | Image semantic segmentation method based on multi-feature fusion |
CN112418228B (en) * | 2020-11-02 | 2023-07-21 | 暨南大学 | Image semantic segmentation method based on multi-feature fusion |
CN113705326A (en) * | 2021-07-02 | 2021-11-26 | 重庆交通大学 | Urban construction land identification method based on full convolution neural network |
CN113705326B (en) * | 2021-07-02 | 2023-12-15 | 重庆交通大学 | Urban construction land identification method based on full convolution neural network |
CN113449743B (en) * | 2021-07-12 | 2022-12-09 | 西安科技大学 | Coal dust particle feature extraction method |
CN113449743A (en) * | 2021-07-12 | 2021-09-28 | 西安科技大学 | Coal dust particle feature extraction method |
CN115019038A (en) * | 2022-05-23 | 2022-09-06 | 杭州缦图摄影有限公司 | Similar image pixel-level semantic matching method |
CN115019038B (en) * | 2022-05-23 | 2024-04-30 | 杭州海马体摄影有限公司 | Similar image pixel level semantic matching method |
CN115223154A (en) * | 2022-07-20 | 2022-10-21 | 中国科学院长春光学精密机械与物理研究所 | Target detection system and detection method thereof |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108921196A (en) | A kind of semantic segmentation method for improving full convolutional neural networks | |
CN107748895B (en) | Unmanned aerial vehicle landing landform image classification method based on DCT-CNN model | |
CN105657402B (en) | A kind of depth map restoration methods | |
CN106157319B (en) | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks | |
CN107358242B (en) | Target area color identification method and device and monitoring terminal | |
CN108717569A (en) | It is a kind of to expand full convolutional neural networks and its construction method | |
CN108108751B (en) | Scene recognition method based on convolution multi-feature and deep random forest | |
CN109035149A (en) | A kind of license plate image based on deep learning goes motion blur method | |
CN111291696B (en) | Handwriting Dongba character recognition method based on convolutional neural network | |
CN109671070B (en) | Target detection method based on feature weighting and feature correlation fusion | |
CN109002755B (en) | Age estimation model construction method and estimation method based on face image | |
CN106845529A (en) | Image feature recognition methods based on many visual field convolutional neural networks | |
CN108961675A (en) | Fall detection method based on convolutional neural networks | |
CN108062543A (en) | A kind of face recognition method and device | |
CN106203363A (en) | Human skeleton motion sequence Activity recognition method | |
CN107644221A (en) | Convolutional neural networks traffic sign recognition method based on compression of parameters | |
CN109543548A (en) | A kind of face identification method, device and storage medium | |
CN108364278B (en) | Rock core fracture extraction method and system | |
CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
CN107273894A (en) | Recognition methods, device, storage medium and the processor of car plate | |
CN106295613A (en) | A kind of unmanned plane target localization method and system | |
CN112950477A (en) | High-resolution saliency target detection method based on dual-path processing | |
CN107704797B (en) | Real-time detection method, system and equipment based on pedestrians and vehicles in security video | |
CN109740451A (en) | Road scene image semantic segmentation method based on importance weighting | |
CN107564009A (en) | Outdoor scene Segmentation of Multi-target method based on depth convolutional neural networks |
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 |
Application publication date: 20181130 |
|
RJ01 | Rejection of invention patent application after publication |