CN111985324B - Road detection method combining full convolution regression neural network and conditional random field - Google Patents

Road detection method combining full convolution regression neural network and conditional random field Download PDF

Info

Publication number
CN111985324B
CN111985324B CN202010677001.5A CN202010677001A CN111985324B CN 111985324 B CN111985324 B CN 111985324B CN 202010677001 A CN202010677001 A CN 202010677001A CN 111985324 B CN111985324 B CN 111985324B
Authority
CN
China
Prior art keywords
neural network
road
convolution
regression neural
image
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.)
Active
Application number
CN202010677001.5A
Other languages
Chinese (zh)
Other versions
CN111985324A (en
Inventor
陈雪云
黄夐翾
金鑫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangxi University
Original Assignee
Guangxi University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Guangxi University filed Critical Guangxi University
Priority to CN202010677001.5A priority Critical patent/CN111985324B/en
Publication of CN111985324A publication Critical patent/CN111985324A/en
Application granted granted Critical
Publication of CN111985324B publication Critical patent/CN111985324B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/50Context or environment of the image
    • G06V20/56Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/588Recognition of the road, e.g. of lane markings; Recognition of the vehicle driving pattern in relation to the road
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention discloses a road detection method combining a full convolution regression neural network and a conditional random field, which solves the technical problem that the existing detection method causes unsmooth holes and edges in a road detection area due to extreme illumination conditions, and comprises the following steps: (1) Acquiring image information of a vehicle event data recorder, and designing a preprocessing program; (2) designing a full convolution regression neural network; (3) designing a conditional random field based on a convolution kernel; (4) network training and testing; through the steps, an end-to-end full convolution regression neural network model can be finally obtained; (5) And detecting the road image by using the trained full convolution regression neural network model. The method of the invention can effectively solve the problem of unsmooth holes and edges caused by extreme lighting conditions, thereby improving the accuracy of road detection.

Description

Road detection method combining full convolution regression neural network and conditional random field
Technical Field
The invention belongs to the technical field of deep learning of semantic segmentation, and particularly relates to a road detection method combining a full convolution regression neural network and a conditional random field.
Background
Nowadays, a driving assistance system has been the focus of research for a long time, and road detection is the core problem of the driving assistance system, and is more important; at present, the deep learning technology is already used for road detection, and a good detection result can be obtained, but when the road detection area faces extreme illumination conditions, the road detection area has the phenomena of holes, unsmooth edges and the like;
at present, although the full convolution recurrent neural network has high detection performance, the full convolution recurrent neural network lacks consideration on target space probability distribution among pixels, and the energy optimization function of the conditional random field is not really combined with the full convolution recurrent neural network by the existing method. The invention designs a conditional random field based on a convolution kernel, and combines an energy optimization function of the conditional random field with a full convolution regression neural network to eliminate error detection results such as uneven holes and edges caused by extreme illumination conditions.
Disclosure of Invention
The invention aims to overcome the problem of unsmooth holes and edges caused by extreme illumination conditions, and provides a road detection method combining a full convolution regression neural network and a conditional random field.
In order to achieve the purpose, the invention adopts the technical scheme that:
a road detection method combining a full convolution regression neural network and a conditional random field comprises the following steps:
s1, collecting a video of a driving recorder, randomly editing to obtain images of the driving recorder, wherein the resolution ratio is 720p or 1080p, manually calibrating a road area and a non-road area in each image to obtain calibration pictures corresponding to the road images to form a road detection data set, designing an image preprocessing program, and preprocessing each image by using the image preprocessing program before inputting the image into a full convolution regression neural network model;
s2, constructing a full convolution regression neural network, and setting initial training parameters and a loss function;
s3, constructing a conditional random field based on a convolution kernel, and placing the conditional random field of the convolution kernel at the output of a full convolution regression neural network for calculating energy distribution;
s4, step-by-step training is carried out on the full convolution regression neural network model by using the calibrated road detection data set;
and S5, carrying out road detection by using the trained full convolution regression neural network model.
The preprocessing procedure in step S1 includes the steps of:
s11, adjusting the road image resolution to 1024 × 512;
s12, randomly selecting a point as a rotation central point and rotating at any random angle within the range from 206 to 306 pixels on the Y axis at 412 to 612 pixels on the X axis of the image by taking the top left vertex of the image as an original point, the horizontal direction as the X axis and the vertical direction as the Y axis;
s13, horizontally shifting the image to any value of-200 to 200 pixels, and vertically shifting the image to any value of-100 to 100 pixels;
s14, randomly selecting any symmetric transformation to process the image, wherein the symmetric transformation mode comprises horizontal direction symmetric transformation, vertical direction symmetric transformation and diagonal direction symmetric transformation;
s15, carrying out normalization processing on the road image, wherein the normalization processing form is as follows:
Figure BDA0002584412940000021
wherein I represents a road image, m and n represent a row value and a column value of the road image matrix respectively, and I represents the road image after normalization.
The step S2 includes the steps of:
s21, constructing a full convolution regression neural network, wherein the network structure is divided into two parts, the first part consists of eight convolution modules, the convolution modules consist of a convolution layer, a normalization layer, a maximum pooling layer and a Relu layer, the second part consists of eight deconvolution modules, and the deconvolution modules consist of a deconvolution layer, a normalization layer and a Relu layer;
and S22, determining the training parameters of the full convolution regression neural network, and setting a loss function.
4. The method for road detection with the combination of the full convolution regression neural network and the conditional random field as claimed in claim 3, wherein in the step S22, the full convolution regression neural network uses a regression loss function, and the form of the regression loss function is as follows:
Loss pix (x label ,x out )=||x label -x out || 2
wherein x label Representative road data set S l Sample of (1), x out Representing the output of the full convolution recurrent neural network.
The step S3 includes the steps of:
s31, constructing a convolution kernel conditional random field, taking a detection result of the full convolution regression neural network as input, wherein the convolution kernel conditional random field is composed of a convolution layer and a Relu layer, and the convolution layer uses a specially designed convolution kernel;
and S32, setting an energy optimization function of the convolution kernel conditional random field, and optimizing energy distribution output by the convolution kernel conditional random field.
The specially designed convolution kernel form in step S31 is as follows:
Figure BDA0002584412940000022
wherein x is i Representing the central pixel of a particular convolution kernel, x j Represents x i The peripheral pixels, σ, represent the center pixel x of the particular convolution kernel i And in the peripheral pixel neighborhood, the special convolution kernel does not participate in weight value updating.
The energy optimization function in step S32 is expressed as follows:
Figure BDA0002584412940000031
wherein x out Representing the output of a full convolution recurrent neural network, P e () Representing the energy distribution calculated by the convolution kernel conditional random field, sum () representing the summation operation, and n representing the number of positive values in the energy distribution matrix.
The step S4 includes the steps of:
s41, performing first-stage training on the full convolution regression neural network, inputting two road images serving as a training batch into a full convolution regression neural network model for training, and iterating the whole data set for 10 times by only using a regression loss function as a loss function;
and S42, performing second-stage training on the fully-convolution regression neural network, inputting two road images serving as a training batch into a fully-convolution regression neural network model for training, using the sum of a regression loss function and an energy optimization function of a convolution kernel conditional random field as a loss function, and iterating the whole data set for 90 times.
The step S5 includes the steps of:
s51, after normalization processing is carried out on the new road image, the new road image is input into a trained full convolution regression neural network, target space probability distribution of a road is output, the numerical range of each pixel is 0-1, and the probability that the pixel is a road area is represented;
s52, taking the threshold value as 0.75, and carrying out threshold segmentation on the target space probability distribution output by the full convolution regression neural network to obtain a final road detection result.
Compared with the prior art, the invention has the beneficial effects that: aiming at the problem of unsmooth holes and edges caused by extreme illumination conditions, the conditional random field and the full convolution regression neural network are really combined together, and an energy optimization function of the conditional random field is introduced, so that the problem that the probability relation between adjacent pixels is not considered by the neural network is solved, the problem of unsmooth holes and edges caused by the extreme illumination conditions is effectively solved, and the road detection precision is improved.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of a full convolution recurrent neural network according to the present invention;
FIG. 3 is a schematic diagram of the structure of the conditional random field of the convolution kernel of the present invention;
FIG. 4 is a graph of the effect of the application experiment of the present invention;
FIG. 5 is a comparative graph of the application experiment of the present invention.
Detailed Description
The invention is further described with reference to the following figures and detailed description:
referring to fig. 1-3, a road detection method combining a full convolution regression neural network and a conditional random field includes the following steps:
(1) Collecting the videos of the automobile data recorder, randomly editing to obtain images of the automobile data recorder, manually calibrating a road area and a non-road area in each image to obtain calibration pictures corresponding to the road images, forming a road detection data set, designing an image preprocessing program, and preprocessing each image by using the image preprocessing program before each image is input into the full convolution regression neural network model. The method specifically comprises the following steps:
(11) Collecting a video of a driving recorder, randomly editing to obtain images of the driving recorder, wherein the resolution ratio is 720p or 1080p, manually calibrating a road area and a non-road area in each image to obtain calibration pictures corresponding to the road images, and forming a road detection data set;
(12) Designing an image preprocessing program, and preprocessing each image by using the image preprocessing program before inputting the image into the fully-convolution regression neural network;
(13) The image preprocessing program adjusts the resolution of the road image to 1024 × 512 by using bilinear interpolation;
(14) The image preprocessing program takes the top left vertex of the image as an origin, the horizontal direction is an X axis, the vertical direction is a Y axis, and a point is randomly selected as a rotation center point within the range from 412 to 612 pixels on the X axis of the image and from 206 to 306 pixels on the Y axis, and is randomly rotated by a random angle;
(15) The image preprocessing procedure carries out horizontal offset on the image, the offset degree is any value between-200 and 200 pixels, and the image is vertically offset to any value between-100 and 100 pixels;
(16) The image preprocessing program randomly selects any symmetric transformation to process the image, and the symmetric transformation mode comprises horizontal direction symmetric transformation, vertical direction symmetric transformation and diagonal direction symmetric transformation;
(17) The image preprocessing program is used for carrying out normalization processing on the road image, and the normalization processing form is as follows:
Figure BDA0002584412940000041
wherein I represents a road image, m and n represent a row value and a column value of the road image matrix respectively, and I represents the road image after normalization.
(2) And constructing a full convolution regression neural network, and setting initial training parameters and a loss function. The method specifically comprises the following steps:
(21) Constructing a full convolution regression neural network, wherein the network structure is divided into two parts, the first part consists of eight convolution modules, the convolution modules consist of a convolution layer, a normalization layer, a maximum pooling layer and a Relu layer, the second part consists of eight deconvolution modules, and the deconvolution modules consist of a deconvolution layer, a normalization layer and a Relu layer;
(22) Determining the training parameters of the full convolution regression neural network, and setting a loss function form as follows:
Loss pix (x label ,x out )=||x label -x out || 2
wherein x is label Representative road data set S l Sample of (1), x out Representing the output of the full convolution recurrent neural network.
(3) And constructing a conditional random field based on a convolution kernel, and placing the conditional random field based on the convolution kernel at the output of the full convolution regression neural network for calculating energy distribution. The method specifically comprises the following steps:
(31) And constructing a convolution kernel conditional random field, taking the detection result of the full convolution regression neural network as input, wherein the convolution kernel conditional random field is composed of a convolution layer and a Relu layer, and the convolution layer uses a specially designed convolution kernel. The specially designed convolution kernel form is as follows:
Figure BDA0002584412940000051
wherein x is i Center pixel, x, representing a particular convolution kernel j Represents x i Peripheral pixels, σ, representing a special convolution kernelCenter pixel x i A peripheral pixel neighborhood. The convolution kernel can be instantiated as follows:
0.0625 0.125 0.0625
0.125 -0.75 0.125
0.0625 0.125 0.0625
(32) And setting an energy optimization function of the convolution kernel conditional random field, and optimizing the energy distribution output by the convolution kernel conditional random field.
The energy optimization function is of the form:
Figure BDA0002584412940000052
wherein x is out Representing the output of a full convolution recurrent neural network, P e () Represents the energy distribution calculated by the convolution kernel conditional random field, sum () represents the summation operation, and n represents the number of positive values in the energy distribution matrix.
(4) And step-by-step training the full convolution regression neural network model by using the calibrated road detection data set. The method specifically comprises the following steps:
(41) Performing first-stage training on the fully-convolutional regression neural network, wherein only a regression loss function is used as a loss function, and the whole data set is iterated for 10 times;
(42) The full convolution regression neural network performs the second stage training, the sum of the regression loss function and the energy optimization function of the convolution kernel conditional random field is used as the loss function, and the whole data set iterates for 90 times.
(5) And carrying out road detection by using the trained full convolution regression neural network model. The method specifically comprises the following steps:
(51) After the new road image is subjected to normalization processing, inputting the new road image into a trained full convolution regression neural network, and outputting target space probability distribution of a road, wherein the numerical range of each pixel is between 0 and 1 and represents the probability that the pixel is a road area;
(52) And taking the threshold value as 0.75, and performing threshold value segmentation on the target space probability distribution output by the full convolution regression neural network to obtain a final road detection result.
The invention relates to an application experiment of a road detection method combining a full convolution regression neural network and a conditional random field, which comprises the following steps: to verify the effectiveness of the present invention, two data sets were used for model training and comparative verification. The two data sets are respectively a remote sensing road data set and an automobile data recorder data set, a deep learning framework Tensorflow is adopted for training and testing, and a U-net network is used as a network model.
Regarding other parameters of the experiment, for the remote sensing road data set, the data set is divided into a training set and a testing set according to the ratio of 6 to 4. And in the training stage, setting Batchsize to be 20, performing 10000 times of training, setting the initial learning rate to be 0.001, and reducing the learning rate to be 0.9 before every 1000 times of training by using Adam as an optimizer. Training is carried out by randomly cutting small samples, and in the testing stage, 1000 small samples are randomly cut from a testing set to serve as testing samples to be evaluated in detection accuracy. For the tachograph data set, the data set is divided into a training set and a test set on a 7 to 3 scale. At the time of training, since the size of the input image is 1024 × 512, the Batchsize is set to 2, the initial learning rate is 0.0001, and after each 5000 training, the learning rate decreases to 0.5 before, using Adam as an optimizer.
The experiment uses Pixel-Accuracy (Pixel-Accuracy), intersection-over-Union (IoU) as a measure of Accuracy. The pixel accuracy rate represents the proportion of the correctly classified pixel points to all the pixel points, and the intersection ratio represents the ratio of the intersection and the union of a certain type of prediction result and a true value.
Here, two data sets are used for the experiment, and the comparison experiment result can be seen in fig. 4, wherein a small graph a in fig. 4 shows the experiment result of the remote sensing road data set, and a small graph b shows the experiment result of the automobile data recorder data set. Fig. 5 shows the detection accuracy results. The invention can solve the problem of unsmooth holes and edges generated by extreme lighting conditions and effectively improve the road detection precision.

Claims (6)

1. A road detection method combining a full convolution regression neural network and a conditional random field is characterized by comprising the following steps of:
s1, collecting a video of a driving recorder, randomly editing to obtain images of the driving recorder, wherein the resolution ratio is 720p or 1080p, manually calibrating a road area and a non-road area in each image to obtain calibration pictures corresponding to the road images to form a road detection data set, designing an image preprocessing program, and preprocessing each image by using the image preprocessing program before inputting the image into a full convolution regression neural network model;
s2, constructing a full convolution regression neural network, and setting initial training parameters and a loss function;
s3, constructing a conditional random field based on a convolution kernel, and placing the conditional random field based on the convolution kernel at the output of the full convolution regression neural network for calculating energy distribution;
s4, step-by-step training is carried out on the full convolution regression neural network model by using the calibrated road detection data set;
s5, carrying out road detection by using the trained full convolution regression neural network model;
the step S3 includes the steps of:
s31, constructing a convolution kernel conditional random field, taking a detection result of the full convolution regression neural network as an input, wherein the convolution kernel conditional random field is composed of a convolution layer and a Relu layer, and the convolution layer uses a convolution kernel specially designed;
s32, setting an energy optimization function of the convolution kernel conditional random field, and optimizing energy distribution output by the convolution kernel conditional random field;
the specially designed convolution kernel form in step S31 is as follows:
Figure FDA0003745276000000011
wherein x is i Representing the central pixel of a particular convolution kernel, x j Represents x i Peripheral pixels, σ, representing the central pixel x of the particular convolution kernel i In the peripheral pixel neighborhood, the special convolution kernel does not participate in weight updating;
the energy optimization function in step S32 is expressed as follows:
Figure FDA0003745276000000012
wherein x is out Representing the output of a full convolution recurrent neural network, P e () Represents the energy distribution calculated by the convolution kernel conditional random field, sum () represents the summation operation, and n represents the number of positive values in the energy distribution matrix.
2. The method for detecting a road by combining a full convolution regression neural network and a conditional random field according to claim 1, wherein the preprocessing procedure in the step S1 includes the following steps:
s11, adjusting the road image resolution to 1024 × 512;
s12, randomly selecting a point as a rotation central point and randomly rotating by a random angle in a range from 412 to 612 pixels on the X axis of the image and from 206 to 306 pixels on the Y axis by taking the top left vertex of the image as an origin, taking the horizontal direction as the X axis and the vertical direction as the Y axis;
s13, horizontally shifting the image to any value of-200 to 200 pixels, and vertically shifting the image to any value of-100 to 100 pixels;
s14, randomly selecting any symmetric transformation to process the image, wherein the symmetric transformation mode comprises horizontal direction symmetric transformation, vertical direction symmetric transformation and diagonal direction symmetric transformation;
s15, carrying out normalization processing on the road image, wherein the normalization processing form is as follows:
Figure FDA0003745276000000021
0<i<m,0<j<n
wherein I represents a road image, m and n represent the row and column values of the road image matrix, respectively, I * Representing the normalized road image.
3. The method for detecting a road by combining a full convolution regression neural network and a conditional random field according to claim 1, wherein the step S2 comprises the following steps:
s21, constructing a full convolution regression neural network, wherein the network structure is divided into two parts, the first part consists of eight convolution modules, each convolution module consists of a convolution layer, a normalization layer, a maximum pooling layer and a Relu layer, the second part consists of eight deconvolution modules, and each deconvolution module consists of a deconvolution layer, a normalization layer and a Relu layer;
and S22, determining training parameters of the full convolution regression neural network, and setting a loss function.
4. The method for detecting a road according to claim 3, wherein in the step S22, the regression loss function is used in the full convolution regression neural network, and the form of the regression loss function is as follows:
Loss pix (x label ,x out )=||x label -x out || 2
wherein x label Representative road data set S l Sample of (1), x out Representing the output of the full convolution recurrent neural network.
5. The method for detecting a road by combining a full convolution regression neural network and a conditional random field according to claim 1, wherein the step S4 comprises the steps of:
s41, performing first-stage training on the full convolution regression neural network, inputting two road images serving as a training batch into a full convolution regression neural network model for training, and iterating the whole data set for 10 times by only using a regression loss function as a loss function;
and S42, performing second-stage training on the fully-convolution regression neural network, inputting two road images serving as a training batch into a fully-convolution regression neural network model for training, using the sum of a regression loss function and an energy optimization function of a convolution kernel conditional random field as a loss function, and iterating the whole data set for 90 times.
6. The method for detecting a road by combining a full convolution regression neural network and a conditional random field according to claim 1, wherein the step S5 comprises the following steps:
s51, after normalization processing is carried out on the new road image, the new road image is input into a trained full convolution regression neural network, and target space probability distribution of a road is output, wherein the numerical range of each pixel is between 0 and 1 and represents the probability that the pixel is a road area;
s52, taking the threshold value as 0.75, and carrying out threshold segmentation on the target space probability distribution output by the full convolution regression neural network to obtain a final road detection result.
CN202010677001.5A 2020-07-14 2020-07-14 Road detection method combining full convolution regression neural network and conditional random field Active CN111985324B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010677001.5A CN111985324B (en) 2020-07-14 2020-07-14 Road detection method combining full convolution regression neural network and conditional random field

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010677001.5A CN111985324B (en) 2020-07-14 2020-07-14 Road detection method combining full convolution regression neural network and conditional random field

Publications (2)

Publication Number Publication Date
CN111985324A CN111985324A (en) 2020-11-24
CN111985324B true CN111985324B (en) 2022-10-28

Family

ID=73437762

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010677001.5A Active CN111985324B (en) 2020-07-14 2020-07-14 Road detection method combining full convolution regression neural network and conditional random field

Country Status (1)

Country Link
CN (1) CN111985324B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108876796A (en) * 2018-06-08 2018-11-23 长安大学 A kind of lane segmentation system and method based on full convolutional neural networks and condition random field
CN109461177A (en) * 2018-09-29 2019-03-12 浙江科技学院 A kind of monocular image depth prediction approach neural network based
CN109636905A (en) * 2018-12-07 2019-04-16 东北大学 Environment semanteme based on depth convolutional neural networks builds drawing method
CN109993082A (en) * 2019-03-20 2019-07-09 上海理工大学 The classification of convolutional neural networks road scene and lane segmentation method
CN110414670A (en) * 2019-07-03 2019-11-05 南京信息工程大学 A kind of image mosaic tampering location method based on full convolutional neural networks

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108876796A (en) * 2018-06-08 2018-11-23 长安大学 A kind of lane segmentation system and method based on full convolutional neural networks and condition random field
CN109461177A (en) * 2018-09-29 2019-03-12 浙江科技学院 A kind of monocular image depth prediction approach neural network based
CN109636905A (en) * 2018-12-07 2019-04-16 东北大学 Environment semanteme based on depth convolutional neural networks builds drawing method
CN109993082A (en) * 2019-03-20 2019-07-09 上海理工大学 The classification of convolutional neural networks road scene and lane segmentation method
CN110414670A (en) * 2019-07-03 2019-11-05 南京信息工程大学 A kind of image mosaic tampering location method based on full convolutional neural networks

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
全卷积神经网络研究综述;章琳等;《计算机工程与应用》;20191108(第01期);31-43 *

Also Published As

Publication number Publication date
CN111985324A (en) 2020-11-24

Similar Documents

Publication Publication Date Title
CN111709265A (en) Camera monitoring state classification method based on attention mechanism residual error network
CN110287777B (en) Golden monkey body segmentation algorithm in natural scene
Chen et al. Remote sensing image quality evaluation based on deep support value learning networks
CN112884712B (en) Method and related device for classifying defects of display panel
CN111242026B (en) Remote sensing image target detection method based on spatial hierarchy perception module and metric learning
CN112489054A (en) Remote sensing image semantic segmentation method based on deep learning
CN112712488B (en) Remote sensing image super-resolution reconstruction method based on self-attention fusion
CN110751644B (en) Road surface crack detection method
CN114155210B (en) Crowd counting method based on attention mechanism and standardized dense cavity space multi-scale fusion network
CN111209858A (en) Real-time license plate detection method based on deep convolutional neural network
CN115063786A (en) High-order distant view fuzzy license plate detection method
CN110689060A (en) Heterogeneous image matching method based on aggregation feature difference learning network
CN114663392A (en) Knowledge distillation-based industrial image defect detection method
CN116051382A (en) Data enhancement method based on deep reinforcement learning generation type antagonistic neural network and super-resolution reconstruction
CN115240259A (en) Face detection method and face detection system based on YOLO deep network in classroom environment
CN114943894A (en) ConvCRF-based high-resolution remote sensing image building extraction optimization method
CN113888442A (en) Infrared image non-uniformity correction method and system based on fpga
CN111985324B (en) Road detection method combining full convolution regression neural network and conditional random field
CN104123723A (en) Structure compensation based image quality evaluation method
CN113971764B (en) Remote sensing image small target detection method based on improvement YOLOv3
CN113256528B (en) Low-illumination video enhancement method based on multi-scale cascade depth residual error network
US20240070808A1 (en) Enhancement of coarse resolution imagery using super-resolution and convolution neural networks to be used for fit-for-purpose systematic land titling and land valuation
CN112861869A (en) Sorghum lodging image segmentation method based on lightweight convolutional neural network
CN113344006A (en) Polarization image analysis method adopting learnable parameter fusion network
CN110889461A (en) Image multi-feature extraction and classification method for chassis assembly quality detection

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
GR01 Patent grant
GR01 Patent grant