CN110458060A - A kind of vehicle image optimization method and system based on confrontation study - Google Patents
A kind of vehicle image optimization method and system based on confrontation study Download PDFInfo
- Publication number
- CN110458060A CN110458060A CN201910694429.8A CN201910694429A CN110458060A CN 110458060 A CN110458060 A CN 110458060A CN 201910694429 A CN201910694429 A CN 201910694429A CN 110458060 A CN110458060 A CN 110458060A
- Authority
- CN
- China
- Prior art keywords
- image
- vehicle image
- vehicle
- optimized model
- loss function
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses it is a kind of based on confrontation study vehicle image optimization method and system, steps of the method are: collect different angle shooting vehicle image, vehicle image is divided into standard scene image and non-standard scene image;Low quality data collection is used as after carrying out image preprocessing to non-standard image;It constructs based on the vehicle image Optimized model for generating confrontation network, model is made of generator, arbiter and feature extractor;Loss function is arranged based on the vehicle image Optimized model for generating confrontation network in training, calculates network weight gradient using backpropagation and updates vehicle image Optimized model parameter;After the completion of the training of vehicle image Optimized model, retains generator as final vehicle image Optimized model, input more scene vehicle images, export the standard scene image of optimization.The present invention realizes that complex scene vehicle image is migrated to standard scene vehicle image, reaches optimization picture quality purpose, promotes vehicle detection recognition accuracy.
Description
Technical field
The present invention relates to technical field of computer vision, and in particular to a kind of vehicle image optimization side based on confrontation study
Method and system.
Background technique
In public transport field, vehicle detection identification technology be widely used in bayonet system, intelligent transportation system,
In the subdomains such as automated driving system, electronic police system, but current vehicle detection identification technology is universal there are one
Problem, requirement of the common vehicle identification algorithm for input data is quite stringent, for example needs visual angle placed in the middle, and picture is clear
It is clear, brightness is suitable etc., otherwise accuracy rate can decline to a great extent, thus under real complicated monitoring scene, such as vehicle shooting angle
Excessive, outdoor illumination variation, the imaging is not clear, has shelter etc., and existing algorithm application there is considerable restraint, be easy to appear mistake
The high problem of discrimination.
And for vehicle image optimisation technique, it is more the image optimization of single scene at present, such as vehicle angles are rectified
Just, currently existing scheme is first to extract vehicle key point, does matrix operation according to key point position and realizes vehicle alignment, based on tradition
Machine learning needs manually to extract feature, and the degree of automation is low, and complex scene is difficult to adaptively, in addition, the party
Method is simply possible to use in vehicle angles correction, other factors for causing recognition accuracy to decline are not accounted for, therefore deposits
In certain limitation.
Summary of the invention
In order to overcome shortcomings and deficiencies of the existing technology, the present invention provides a kind of vehicle image based on confrontation study
Optimization method and system, fighting network by the generation based on confrontation loss, pixel loss and perception loss will be by actual acquisition
To low quality vehicle image be mapped as the vehicle image of high quality, realize complex scene vehicle image to standard scene vehicle figure
As migration, reach optimization picture quality purpose, by providing quality more preferably data, is tied with reducing external complex factor to detection
The influence of fruit promotes vehicle detection recognition accuracy.
In order to achieve the above object, the invention adopts the following technical scheme:
The present invention provides a kind of vehicle image optimization method based on confrontation study, includes the following steps:
S1: the vehicle image of different angle shooting is collected, vehicle image is divided into standard scene image and non-standard field
Scape image;
S2: image preprocessing is carried out to non-standard image, using pretreated image as low quality data collection;
S3: it constructs based on the vehicle image Optimized model for generating confrontation network, the vehicle image Optimized model includes life
It grows up to be a useful person, arbiter and feature extractor;
S4: training is based on the vehicle image Optimized model for generating confrontation network:
S41: input low quality data collection is defeated via layer neuron nonlinear combination each in generator network into generator
The high quality graphic being born;
S42: the high quality graphic of generation, corresponding true high quality graphic are input to arbiter and feature extractor
In, the high quality graphic, the corresponding true high quality graphic that are generated are determined as the probability and characteristics of image square of true picture
Battle array;
S43: setting loss function calculates network weight gradient using backpropagation and updates vehicle image Optimized model ginseng
Number;
S44: circulation executes step S41-S43;
S5: after the completion of the training of vehicle image Optimized model, reservation generator is defeated as final vehicle image Optimized model
Enter more scene vehicle images, exports the standard scene image of optimization.
Image preprocessing is carried out to non-standard image described in step S2 as a preferred technical solution, described image is located in advance
Reason added using image make an uproar, any one or more in luminance transformation, selective erasing or Fuzzy Processing.
Into generator, the generator is set input low quality data collection described in step S41 as a preferred technical solution,
Encoder and decoder are set, the low quality vehicle image feature of encoder study input is simultaneously encoded into eigenmatrix, and decoder will
Input picture characteristic information is decoded into standard scene vehicle image.
Loss function is set described in step S43 as a preferred technical solution, and the loss function includes confrontation loss
Function, L1 loss function and perception loss function.
The confrontation loss function specific formula for calculation as a preferred technical solution, are as follows:
Wherein, y~Y indicates that y submits to the distribution of the vehicle image under standard scene, and x~X indicates that x submits to complex scene
Under vehicle image distribution, E indicates the expectation of each lot sample sheet;
The L1 loss function specific formula for calculation are as follows:
Wherein,Indicate j-th of pixel of the i-th row in true picture,It indicates to generate j-th of picture of the i-th row in image
Element, W and H respectively indicate the length and width of input picture;
The perception loss function specific formula for calculation are as follows:
Wherein,Indicate that the i-th row of true picture feature jth arranges the value of m dimension,It indicates to generate figure
As feature the i-th row jth arranges the value of m dimension, WmAnd HmRespectively indicate the length and width of the eigenmatrix extracted;
The final loss function of generator are as follows:
L=λ1Ladv+λ2Lpixcel+λ3Lper;
Wherein, λ1、λ2、λ3Indicate confrontation loss function, L1 loss function and the accounting weight for perceiving loss function.
The present invention also provides it is a kind of based on confrontation study vehicle image optimization system, comprising:, image pre-processing module,
Vehicle image Optimized model constructs module and vehicle image Optimized model training module;
Described image preprocessing module is used to carry out image preprocessing to non-standard image, obtains low quality data collection;
The vehicle image Optimized model building module includes generator, arbiter and feature extractor, the generator
For by low quality data collection generate high quality graphic, the arbiter for distinguish input picture be true high quality graphic or
Person is the high quality graphic generated, and the feature extractor is used to extract the high quality graphic and true high quality graphic generated
Characteristics of image;
Vehicle image Optimized model of the vehicle image Optimized model training module for training building, by low quality number
According in collection input vehicle image Optimized model, model parameter is updated by loss function and backpropagation, obtains final vehicle
Image optimization model.
The generator includes encoder and decoder as a preferred technical solution, and the encoder is equipped with port number
10 residual blocks that respectively 64,128,256,512 four layers of convolutional layer and port number are 512, the decoder are equipped with channel
Number is respectively 512,256,128,64 four layers of warp lamination.
It is respectively the six of 64,256,512,128,64,1 that the arbiter, which uses port number, as a preferred technical solution,
The full convolution sorter network of layer.
The feature extractor uses deep learning network VGGNet-16 model as a preferred technical solution,.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) present invention, using the competition between generator and arbiter, input vehicle image is turned using confrontation network is generated
Change the vehicle image of standard scene into, the vehicle image of input is constantly drawn close to the transformation of standard scene image, whole image optimization
Process is simple and efficient.
(2) the present invention is based on convolutional neural networks, automatically the learning characteristic from mass data, can effectively mitigate artificial negative
Load.
(3) the present invention is based on confrontation study to optimize to low-quality image, solves that vehicle detection recognition efficiency is low asks
Topic, do not need to increase new hardware device, only need to carry out software upgrading on existing, substantially reduce deployment difficulty and at
This.
(4) by the present invention in that it is complete with the details of the standard scene vehicle image of a variety of loss function common guarantees generation
Property, keep image truer, while a variety of loss functions can also help network more stably to train.
(5) present invention can adapt to the various complex scenes such as vehicle angles are big, vehicle body blocks, picture is unintelligible, compared to
Existing method has better scene adaptability and robustness.
Detailed description of the invention
Fig. 1 is the flow diagram of vehicle image optimization method of the present embodiment based on confrontation study;
Fig. 2 is that the generation of vehicle image optimization method of the present embodiment based on confrontation study fights schematic network structure.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
Embodiment
As shown in Figure 1 and Figure 2, it the present embodiment provides a kind of vehicle image optimization method based on confrontation study, collects first
The vehicle image of different angle simulates several scenes by image translation operation, and then building generates confrontation network model, then
With the data set training pattern divided in pairs by standard scene and non-standard scene, finally retain generator as final vehicle
Image optimization model, it is described that specific step is as follows:
S1: collect different angle shooting vehicle image, divide the image into vehicle angles between two parties and two class not placed in the middle, wherein
Vehicle angles centered image is standard scene image, is switch target generating confrontation network Plays scene image, is low-quality
It measures image optimization and reference, standard scene image namely true high quality graphic is provided;
S2: pre-processing non-vehicle image placed in the middle, will by the complex scene in image processing operations simulation reality
Treated image is target to be optimized in generating confrontation network as low quality data collection;
Image processing operations in step S2 include four kinds, be respectively image add make an uproar, luminance transformation, selective erasing, mould
Paste processing, common, picture noise in these four operation simulations reality, image overexposure are owed to expose, and vehicle body blocks fuzzy with image
Four kinds of situations, making the vehicle image optimization method of embodiment is being to have stronger robustness in face of various complex scenes;
1) image, which adds, makes an uproar: being influenced by factors such as shooting environmental, imaging sensor, transmission channels, digital picture can have one
A little noises, Gaussian noise is added to part training set image using matlab for the present embodiment and salt-pepper noise carrys out analog image noise
Scene;
2) luminance transformation: the light variation in one day will lead to video monitoring image and the case where dark or overexposure occurred, be
Meet this scene, the present embodiment has respectively done part training set image using matlab and lightened and dim processing, with this mould
Intend the light variation of reality;
3) selective erasing: in actual scene, the case where vehicle body is blocked, is relatively conventional, the present embodiment using matlab from
Partial region is randomly choosed in the training set image of part to be filled, and circumstance of occlusion is simulated with this;
4) Fuzzy Processing: the case where focusing inaccuracy and vehicle movement will cause image blur, the present embodiment uses
Matlab does Gauss or mean filter part training set image to obscure caused by simulating both of these case;
S3: constructing based on the vehicle image Optimized model for generating confrontation network, and model is mainly made of three components, respectively
It is generator, arbiter and feature extractor, three components are all made of deep neural network structure, and generator includes encoder
With two parts of decoder, encoder is responsible for the low quality vehicle image feature of study input and is encoded into eigenmatrix, decodes
Device is decoded into vehicle image (the vehicle visual angle of target style image namely standard scene according to input picture characteristic information
Placed in the middle and unobstructed, moderate, image clearly of vehicle image brightness etc.), it is true standard that arbiter effect, which is to discriminate between input picture,
Scene image still passes through the image of generator conversion, and feature extractor, which is then used to extract, generates image and true picture feature;
Wherein, encoder by four layers of convolutional layer and port number that port number is 64,128,256,512 are 512 10 it is residual
Poor block composition;Decoder is made of four warp laminations, and port number is respectively 512,256,128,64;Arbiter is six layers of full volume
Product sorter network, port number is 64,256,512,128,64,1 respectively, and above-mentioned each component all uses for every layer in addition to the last layer
LRelu is as activation primitive;The present embodiment uses deep learning network VGGNet-16 model as feature extractor, in order to make
It carries out pre-training with preferable ability in feature extraction, using BIT vehicle data the set pair analysis model;
S4: as shown in Fig. 2, training inputs a collection of low-quality spirogram based on the vehicle image Optimized model for generating confrontation network
As into generator, the high quality graphic by its conversion is obtained, respectively by the high quality graphic of generation, corresponding true high-quality
Spirogram picture is input in arbiter and feature extractor, and both output is the probability of true picture and the eigenmatrix of the two,
According to model loss function, network weight gradient is found out by the way of backpropagation and is updated, the above process is repeated, directly
Reach demand to effect of optimization of the generator to vehicle image;
In the present embodiment, generator uses neural network model, is equivalent to a nonlinear function, may be implemented to input
To the mapping of output, image array is input to network, can do NONLINEAR CALCULATION with the neuron in network, low layer nerve in network
Member is combined with each other with certain weight, is then output to next layer of neuron, finally output layer output want as a result, this implementation
Example input low quality data collection is into generator, via the height of layer neuron nonlinear combination each in generator network output generation
Quality image;
In the present embodiment, the advantages of arbiter equally uses neural network, neural network exactly can only need to prepare
Good training data and label, network training can learn automatically according to label, and training terminates just to obtain desired as a result, pilot process
It is the forward and reverse communication process of neural network, the effect of arbiter is to learn true picture simultaneously and generate image, then
Them are distinguished as far as possible, its output is 0~1 numerical value, and it is true picture that expression, which is input to this picture of arbiter,
Probability, the probability value and eigenmatrix of output take part in respectively confrontation costing bio disturbance and perception costing bio disturbance, loss function
Effect is that the weight of each neuron in neural network, the i.e. training process of neural network are updated by back-propagation process;
In the present embodiment, encoder and decoder also all use multilayer neural network, distinguish only because function is different
So every layer of neuron number difference, the process of the parameter learning of the present embodiment can be summarized simply as follows two steps
It suddenly, is prediction process from input layer via the Nonlinear Mapping of middle layer to output layer first, referred to as forward-propagating, followed by
According to parameter renewal process of the chain rule from output layer to input layer, this process is backpropagation, passes through neural network
Forward-propagating process obtains high quality graphic, with trained progress, constantly learns to data, ultimately generates high-quality
Picture, the image namely standard scene image of generation.
In the present embodiment, need to be collected simultaneously non-standard scene image and standard scene image when collecting data,
Every non-standard scene image is required to have a corresponding standard scene image simultaneously, such as one is clapped in greasy weather flank angle
The figure for certain vehicle taken the photograph, will there is a same vehicle front clearly standard scene image, and the non-standard scene of the present embodiment is adopted
It is simulated with image transformation;
In order to which the image for generating the Optimized model of the present embodiment has the enough sense of reality, in addition to fighting loss, this implementation
Example also introduces two loss functions, respectively pixel loss and perception loss, the two losses are respectively from pixel scale and spy
Levy rank minimize generate image between true picture at a distance from, it only includes confrontation loss function that existing generations, which fights network,
It is bad to generate image effect, the present embodiment increases pixel loss and perception loss is promoted and generates picture quality, on the one hand can guarantee
Generate image detail consistency and rich, on the other hand also can stabilizing network optimization process, each loss function is as follows:
1) confrontation loss
Generator G in network is generated to be continuously generated the vehicle image of standardization scene and attempt that arbiter D is made to think the figure
Seem true rather than what generator was simulated, and arbiter D then will as possible distinguish input image when really or generate,
By the cross-training of the two, mutual game finally enables generator simulate truthful data distribution, realizes low quality vehicle figure
Conversion as being distributed to the vehicle image distribution of high quality.Generate the objective function of confrontation network are as follows:
Wherein, y~Y indicates that y submits to the distribution of the vehicle image under standard scene, and x~X indicates that x submits to complex scene
Under vehicle image distribution, E indicates the expectation of each lot sample sheet;
In the present embodiment, for arbiter D, its task is to distinguish truthful data and generate data, therefore work as and encounter
Authentic specimen y, it will make the value of D (y) maximum as far as possible, be equivalent to maximize log D (y), and for generating sample G (x), D gives
The lower score out should be the better, namely maximizes log (1-D (G (x))), therefore the objective function of arbiter D is corresponding to above-mentioned
Objective functionPart needs to cheat arbiter D and allows it considers that generating sample belongs to truthful data for generator G,
Therefore G will make the value of D (G (x)) the higher the better, that is, minimize log (1-D (G (x))), due to the present embodiment generator
Training is without using authentic specimen, and the first item of objective function is not related to generator, and the majorized function of generator is V (D, G);
2) L1 loses
L1 loss is also referred to as Pixel-level loss, generates the distance between sample and authentic specimen pixel by optimization, guarantees to generate
Image further stablizes training closer to true picture, and calculation formula is as follows:
Wherein,Indicate j-th of pixel of the i-th row in true picture,It indicates to generate j-th of picture of the i-th row in image
Element, W and H respectively indicate the length and width of input picture;
3) perception loss
Perception loss needs trained vehicle classification model, the feature that it obtains true picture convolution with
It generates the feature that image convolution obtains to make comparisons, using feature extractor, optimizes from the level of characteristics of image and generate image, make it
Better authenticity, perception costing bio disturbance formula are as follows:
Wherein,Indicate that the i-th row of true picture feature jth arranges the value of m dimension,It indicates to generate figure
As feature the i-th row jth arranges the value of m dimension, WmAnd HmRespectively indicate the length and width of the eigenmatrix extracted;
In conclusion the majorized function that generator is final are as follows:
L=λ1Ladv+λ2Lpixcel+λ3Lper;
Wherein, λ1、λ2、λ3It is weight shared by each loss, in the present embodiment, λ1=1, λ2=0.05, λ3=0.05;
S5: after completing entire model training, reservation generator inputs any as final vehicle image Optimized model
Vehicle image under complex scene, standard scene image of the model by output by optimization.
In the present embodiment, a kind of vehicle image optimization system based on confrontation study is also provided, comprising: image preprocessing
Module, vehicle image Optimized model building module and vehicle image Optimized model training module;
Described image preprocessing module is used to carry out image preprocessing to non-standard image, obtains low quality data collection;Institute
Stating vehicle image Optimized model building module includes generator, arbiter and feature extractor, and the generator is used for low-quality
Data set generation high quality graphic is measured, the arbiter is that true high quality graphic either generates for distinguishing input picture
High quality graphic, the feature extractor are used to extract the characteristics of image of the high quality graphic and true high quality graphic that generate;
Vehicle image Optimized model of the vehicle image Optimized model training module for training building, by low quality number
According in collection input vehicle image Optimized model, model parameter is updated by loss function and backpropagation, obtains final vehicle
Image optimization model.
In the present embodiment, the network structure used can learn characteristics of image automatically, keep away for depth convolutional neural networks
Inefficient and high-cost manual features extraction process is exempted from.
In the present embodiment, residual error structure is used in depth convolutional neural networks, enhancing gradient flowing is alleviated due to layer
Gradient disappearance problem caused by number increases.
The above embodiment is a preferred embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, other any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present invention,
It should be equivalent substitute mode, be included within the scope of the present invention.
Claims (9)
1. a kind of vehicle image optimization method based on confrontation study, which is characterized in that include the following steps:
S1: the vehicle image of different angle shooting is collected, vehicle image is divided into standard scene image and non-standard scene figure
Picture;
S2: image preprocessing is carried out to non-standard image, using pretreated image as low quality data collection;
S3: construct based on generate confrontation network vehicle image Optimized model, the vehicle image Optimized model include generator,
Arbiter and feature extractor;
S4: training is based on the vehicle image Optimized model for generating confrontation network:
S41: input low quality data collection is exported via layer neuron nonlinear combination each in generator network and is given birth into generator
At high quality graphic;
S42: the high quality graphic of generation, corresponding true high quality graphic are input in arbiter and feature extractor, obtained
It is determined as the probability and image characteristic matrix of true picture to high quality graphic, the corresponding true high quality graphic of generation;
S43: setting loss function calculates network weight gradient using backpropagation and updates vehicle image Optimized model parameter;
S44: circulation executes step S41-S43;
S5: after the completion of the training of vehicle image Optimized model, retain generator as final vehicle image Optimized model, input more
Scene vehicle image exports the standard scene image of optimization.
2. the vehicle image optimization method according to claim 1 based on confrontation study, which is characterized in that described in step S2
Image preprocessing carried out to non-standard image, described image pretreatment added using image make an uproar, luminance transformation, selective erasing or fuzzy
Any one or more in processing.
3. the vehicle image optimization method according to claim 1 based on confrontation study, which is characterized in that step S41 institute
Input low quality data collection is stated into generator, encoder and decoder are arranged in the generator, and encoder study inputs low
Mass x vehicle characteristics of image is simultaneously encoded into eigenmatrix, and input picture characteristic information is decoded into standard scene vehicle figure by decoder
Picture.
4. the vehicle image optimization method according to claim 1 based on confrontation study, which is characterized in that in step S43
The setting loss function, the loss function include confrontation loss function, L1 loss function and perception loss function.
5. the vehicle image optimization method according to claim 4 based on confrontation study, which is characterized in that
The confrontation loss function specific formula for calculation are as follows:
Wherein, y~Y indicates that y submits to the distribution of the vehicle image under standard scene, and x~X indicates that x is submitted under complex scene
Vehicle image distribution, E indicate the expectation of each lot sample sheet;
The L1 loss function specific formula for calculation are as follows:
Wherein,Indicate j-th of pixel of the i-th row in true picture,It indicates to generate j-th of pixel of the i-th row, W and H in image
Respectively indicate the length and width of input picture;
The perception loss function specific formula for calculation are as follows:
Wherein,Indicate that the i-th row of true picture feature jth arranges the value of m dimension,It indicates to generate image spy
Levy the value that the i-th row jth arranges m dimension, WmAnd HmRespectively indicate the length and width of the eigenmatrix extracted;
The final loss function of generator are as follows:
L=λ1Ladv+λ2Lpixcel+λ3Lper;
Wherein, λ1、λ2、λ3Indicate confrontation loss function, L1 loss function and the accounting weight for perceiving loss function.
6. it is a kind of based on confrontation study vehicle image optimization system characterized by comprising, image pre-processing module, vehicle
Image optimization model construction module and vehicle image Optimized model training module;
Described image preprocessing module is used to carry out image preprocessing to non-standard image, obtains low quality data collection;
The vehicle image Optimized model building module includes generator, arbiter and feature extractor, and the generator is used for
By low quality data collection generate high quality graphic, the arbiter for distinguish input picture be true high quality graphic either
The high quality graphic of generation, the feature extractor are used to extract the image of the high quality graphic and true high quality graphic that generate
Feature;
Vehicle image Optimized model of the vehicle image Optimized model training module for training building, by low quality data collection
It inputs in vehicle image Optimized model, model parameter is updated by loss function and backpropagation, obtains final vehicle image
Optimized model.
7. the vehicle image optimization system according to claim 6 based on confrontation study, which is characterized in that the generator
Including encoder and decoder, the encoder is equipped with four layers of convolutional layer and the channel that port number is respectively 64,128,256,512
10 residual blocks that number is 512, the decoder are equipped with four layers of warp lamination that port number is respectively 512,256,128,64.
8. the vehicle image optimization system according to claim 6 based on confrontation study, which is characterized in that the arbiter
It is respectively 64,256,512,128,64,1 six layers of full convolution sorter network using port number.
9. the vehicle image optimization system according to claim 6 based on confrontation study, which is characterized in that the feature mentions
Take device using deep learning network VGGNet-16 model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910694429.8A CN110458060A (en) | 2019-07-30 | 2019-07-30 | A kind of vehicle image optimization method and system based on confrontation study |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910694429.8A CN110458060A (en) | 2019-07-30 | 2019-07-30 | A kind of vehicle image optimization method and system based on confrontation study |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110458060A true CN110458060A (en) | 2019-11-15 |
Family
ID=68484009
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910694429.8A Pending CN110458060A (en) | 2019-07-30 | 2019-07-30 | A kind of vehicle image optimization method and system based on confrontation study |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110458060A (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111191654A (en) * | 2019-12-30 | 2020-05-22 | 重庆紫光华山智安科技有限公司 | Road data generation method and device, electronic equipment and storage medium |
CN111199550B (en) * | 2020-04-09 | 2020-08-11 | 腾讯科技(深圳)有限公司 | Training method, segmentation method, device and storage medium of image segmentation network |
CN112016506A (en) * | 2020-09-07 | 2020-12-01 | 重庆邮电大学 | Classroom attitude detection model parameter training method capable of rapidly adapting to new scene |
CN112115771A (en) * | 2020-08-05 | 2020-12-22 | 暨南大学 | Gait image synthesis method based on star-shaped generation confrontation network |
CN112116537A (en) * | 2020-08-31 | 2020-12-22 | 中国科学院长春光学精密机械与物理研究所 | Image reflected light elimination method and image reflected light elimination network construction method |
CN112862702A (en) * | 2021-01-18 | 2021-05-28 | 平安科技(深圳)有限公司 | Image enhancement method, device, equipment and storage medium |
CN113628121A (en) * | 2020-05-06 | 2021-11-09 | 阿里巴巴集团控股有限公司 | Method and device for processing data and training multimedia data |
CN113822248A (en) * | 2021-11-23 | 2021-12-21 | 江苏金晓电子信息股份有限公司 | Cross-domain vehicle detection method for generating countermeasure network based on cycleGAN |
CN114492059A (en) * | 2022-02-07 | 2022-05-13 | 清华大学 | Multi-agent confrontation scene situation assessment method and device based on field energy |
CN114979470A (en) * | 2022-05-12 | 2022-08-30 | 咪咕文化科技有限公司 | Camera rotation angle analysis method, device, equipment and storage medium |
CN117237901A (en) * | 2023-11-15 | 2023-12-15 | 深圳市城市交通规划设计研究中心股份有限公司 | Cross-domain self-adaptive automatic driving scene data generation method |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107590774A (en) * | 2017-09-18 | 2018-01-16 | 北京邮电大学 | A kind of car plate clarification method and device based on generation confrontation network |
-
2019
- 2019-07-30 CN CN201910694429.8A patent/CN110458060A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107590774A (en) * | 2017-09-18 | 2018-01-16 | 北京邮电大学 | A kind of car plate clarification method and device based on generation confrontation network |
Non-Patent Citations (2)
Title |
---|
LONGFEI LIU 等: "X-GANs: Image Reconstruction Made Easy for Extreme Cases", 《HTTPS://ARXIV.ORG/PDF/1609.04802V1.PDF》 * |
XINTAO WANG 等: "ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks", 《ECCV 2018》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111191654A (en) * | 2019-12-30 | 2020-05-22 | 重庆紫光华山智安科技有限公司 | Road data generation method and device, electronic equipment and storage medium |
CN111199550B (en) * | 2020-04-09 | 2020-08-11 | 腾讯科技(深圳)有限公司 | Training method, segmentation method, device and storage medium of image segmentation network |
CN113628121A (en) * | 2020-05-06 | 2021-11-09 | 阿里巴巴集团控股有限公司 | Method and device for processing data and training multimedia data |
CN113628121B (en) * | 2020-05-06 | 2023-11-14 | 阿里巴巴集团控股有限公司 | Method and device for processing and training multimedia data |
CN112115771A (en) * | 2020-08-05 | 2020-12-22 | 暨南大学 | Gait image synthesis method based on star-shaped generation confrontation network |
CN112115771B (en) * | 2020-08-05 | 2022-04-01 | 暨南大学 | Gait image synthesis method based on star-shaped generation confrontation network |
CN112116537A (en) * | 2020-08-31 | 2020-12-22 | 中国科学院长春光学精密机械与物理研究所 | Image reflected light elimination method and image reflected light elimination network construction method |
CN112016506B (en) * | 2020-09-07 | 2022-10-11 | 重庆邮电大学 | Classroom attitude detection model parameter training method capable of quickly adapting to new scene |
CN112016506A (en) * | 2020-09-07 | 2020-12-01 | 重庆邮电大学 | Classroom attitude detection model parameter training method capable of rapidly adapting to new scene |
CN112862702A (en) * | 2021-01-18 | 2021-05-28 | 平安科技(深圳)有限公司 | Image enhancement method, device, equipment and storage medium |
CN112862702B (en) * | 2021-01-18 | 2023-10-31 | 平安科技(深圳)有限公司 | Image enhancement method, device, equipment and storage medium |
CN113822248A (en) * | 2021-11-23 | 2021-12-21 | 江苏金晓电子信息股份有限公司 | Cross-domain vehicle detection method for generating countermeasure network based on cycleGAN |
CN114492059A (en) * | 2022-02-07 | 2022-05-13 | 清华大学 | Multi-agent confrontation scene situation assessment method and device based on field energy |
CN114492059B (en) * | 2022-02-07 | 2023-02-28 | 清华大学 | Multi-agent confrontation scene situation assessment method and device based on field energy |
CN114979470A (en) * | 2022-05-12 | 2022-08-30 | 咪咕文化科技有限公司 | Camera rotation angle analysis method, device, equipment and storage medium |
CN117237901A (en) * | 2023-11-15 | 2023-12-15 | 深圳市城市交通规划设计研究中心股份有限公司 | Cross-domain self-adaptive automatic driving scene data generation method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110458060A (en) | A kind of vehicle image optimization method and system based on confrontation study | |
CN109815893B (en) | Color face image illumination domain normalization method based on cyclic generation countermeasure network | |
CN110097131B (en) | Semi-supervised medical image segmentation method based on countermeasure cooperative training | |
CN111292264B (en) | Image high dynamic range reconstruction method based on deep learning | |
CN110473154A (en) | A kind of image de-noising method based on generation confrontation network | |
CN108830252A (en) | A kind of convolutional neural networks human motion recognition method of amalgamation of global space-time characteristic | |
CN110348376A (en) | A kind of pedestrian's real-time detection method neural network based | |
CN108416755A (en) | A kind of image de-noising method and system based on deep learning | |
CN107808132A (en) | A kind of scene image classification method for merging topic model | |
CN110059741A (en) | Image-recognizing method based on semantic capsule converged network | |
CN111681178B (en) | Knowledge distillation-based image defogging method | |
CN113313644B (en) | Underwater image enhancement method based on residual double-attention network | |
CN108648197A (en) | A kind of object candidate area extracting method based on image background mask | |
CN110634108A (en) | Composite degraded live webcast video enhancement method based on element-cycle consistency countermeasure network | |
CN108389192A (en) | Stereo-picture Comfort Evaluation method based on convolutional neural networks | |
CN107346629A (en) | A kind of intelligent blind reading method and intelligent blind reader system | |
CN111161178A (en) | Single low-light image enhancement method based on generation type countermeasure network | |
CN110675335B (en) | Superficial vein enhancement method based on multi-resolution residual error fusion network | |
CN113392711B (en) | Smoke semantic segmentation method and system based on high-level semantics and noise suppression | |
CN109376589A (en) | ROV deformation target and Small object recognition methods based on convolution kernel screening SSD network | |
CN109117877A (en) | A kind of Pelteobagrus fulvidraco and its intercropping kind recognition methods generating confrontation network based on depth convolution | |
CN108537115A (en) | Image-recognizing method, device and electronic equipment | |
CN110020658A (en) | A kind of well-marked target detection method based on multitask deep learning | |
CN113486700A (en) | Facial expression analysis method based on attention mechanism in teaching scene | |
CN115330620A (en) | Image defogging method based on cyclic generation countermeasure network |
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: 20191115 |
|
RJ01 | Rejection of invention patent application after publication |