CN111861906B - Pavement crack image virtual augmentation model establishment and image virtual augmentation method - Google Patents
Pavement crack image virtual augmentation model establishment and image virtual augmentation method Download PDFInfo
- Publication number
- CN111861906B CN111861906B CN202010574126.5A CN202010574126A CN111861906B CN 111861906 B CN111861906 B CN 111861906B CN 202010574126 A CN202010574126 A CN 202010574126A CN 111861906 B CN111861906 B CN 111861906B
- Authority
- CN
- China
- Prior art keywords
- network
- model
- image
- crack image
- pavement crack
- 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
Links
- 230000003416 augmentation Effects 0.000 title claims abstract description 33
- 238000000034 method Methods 0.000 title claims abstract description 25
- 230000006870 function Effects 0.000 claims abstract description 21
- 230000004913 activation Effects 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 6
- 230000003190 augmentative effect Effects 0.000 claims description 4
- 238000004042 decolorization Methods 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 2
- 238000012549 training Methods 0.000 abstract description 16
- 238000012545 processing Methods 0.000 abstract description 3
- 238000003709 image segmentation Methods 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 16
- 238000001514 detection method Methods 0.000 description 12
- 238000010586 diagram Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000009916 joint effect Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G06T5/70—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G06T5/90—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- 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
Abstract
The application belongs to the field of pavement crack image processing, and discloses a pavement crack image virtual augmentation model establishment and an image virtual augmentation method. The model building method comprises the following steps: step 1: acquiring pavement crack images, and sequentially carrying out data quality improvement and image segmentation on the pavement crack images to obtain a real pavement crack image set; step 2: establishing a DCGAN generation countermeasure network model, wherein the DCGAN generation countermeasure network model comprises a generator network and a discriminator network, and penalty items are arranged behind loss functions of the generator network and the discriminator network; step 3: and (3) acquiring random noise, inputting the random noise and the real pavement crack image set acquired in the step (1) into the DCGAN generated countermeasure network model acquired in the step (2) for training, wherein the trained model is the pavement crack image virtual augmentation model. The application effectively solves the problem of insufficient crack image data sets, and well realizes the increase of the quantity and diversity of the crack image data sets.
Description
Technical Field
The application belongs to the field of pavement crack image processing, and particularly relates to a pavement crack image virtual augmentation model establishment and an image virtual augmentation method.
Background
Along with the combination of the computer vision algorithm and the road image processing field, the detection and maintenance of the road surface diseases can be possibly assisted by artificial intelligence. In practical application, a large-scale and high-diversity data set is usually required for training a crack detection model with good performance, but acquisition of a crack image is a time-consuming and labor-consuming project, and the data requirement of the current deep learning network model is difficult to meet. Although conventional methods such as overturning and translation are commonly used at the present stage to amplify an image data set, the number of data set samples amplified by the method is limited and the diversity is poor, and if models such as pavement crack detection, classification and segmentation are trained by the data set, the trained network model lacks generalization performance and is easy to generate a fitting phenomenon. Therefore, how to improve the phenomena of poor generalization capability and over-fitting of the network model by improving the image augmentation method becomes a problem to be solved urgently.
Disclosure of Invention
The application aims to provide a pavement crack image virtual augmentation model establishment and image virtual augmentation method which are used for solving the problems that the road crack cannot be greatly and efficiently augmented by the augmentation technology in the prior art.
In order to realize the tasks, the application adopts the following technical scheme:
a pavement crack image virtual augmentation model building method comprises the following steps:
step 1: collecting pavement crack images, and preprocessing the pavement crack images to obtain a real pavement crack image set, wherein pavement cracks are transverse cracks or longitudinal cracks;
step 2: establishing a DCGAN generation countermeasure network model, wherein the DCGAN generation countermeasure network model comprises a generator network and a discriminator network, and penalty items are arranged behind loss functions of the generator network and the discriminator network;
step 3: and (3) acquiring random noise, inputting the random noise and the real pavement crack image set acquired in the step (1) into the DCGAN generated countermeasure network model acquired in the step (2) to train, wherein the trained generator network model is the virtual augmentation model of the pavement crack image.
Further, the generator network has 5 layers in total, the first layer is a fully connected layer, the second to five layers are convolution layers, and the convolution layers all adopt Relu as an activation function.
Further, the arbiter network has 5 layers, the first to fourth layers are convolution layers, the fifth layer is a full connection layer, and the convolution layers all adopt the leakyrelu as an activation function.
Further, the penalty term added loss function isAnd is also provided withWherein X represents the input of the generator network and the arbiter network, y represents the output of the generator network and the arbiter network, w is a weight vector, alpha represents the regularization coefficient ranging from 0 to 1, and w is 1 Is the sum of the absolute values of the weights w in the model.
Further, the preprocessing in step 1 includes the operations of sequentially performing the color removal, the noise removal, the contrast enhancement, the brightness enhancement and the division into uniform sizes.
An image virtual augmentation method, comprising the steps of:
random noise is acquired, the random noise is input into a pavement crack image virtual augmentation model obtained by any pavement crack image virtual augmentation model building method, and a pavement crack image augmentation data set is obtained.
Compared with the prior art, the application has the following technical characteristics:
(1) The application uses the deep convolution generation type countermeasure network to generate the pavement crack image for the first time. The strategy of selecting L1 regularization constraint is used for reducing test errors in order to reduce the learning of noise by a network, enhancing the generalization capability of the model, enabling the model to be smoother and preventing the occurrence of the over-fitting phenomenon.
(2) According to the application, through optimizing each super parameter in the deep convolution generation type countermeasure network, more vivid and high-quality pavement crack images are obtained.
(3) The method provided by the application effectively solves the problem of insufficient crack image data sets, and well realizes the increase of the quantity and diversity of the crack image data sets.
Drawings
FIG. 1 is a general flow chart of the present application;
FIG. 2 is a schematic diagram of a generated countermeasure network;
FIG. 3 is a diagram of a network architecture of a generator in a deep convolution generation countermeasure network;
FIG. 4 is a diagram of a network architecture of a arbiter in a deep convolution generation countermeasure network;
FIG. 5 is a graph of the effect of cross-joint generation of a crack-propagation model; wherein, (a) generating an effect graph for 100 epoch; (b) generating an effect map for 200 epoch; (c) generating an effect map for 300 epoch; (d) generating an effect map for 400 epoch; (e) generating an effect map for 500 epoch; (f) generating an effect map for 600 epoch;
FIG. 6 is a graph of the effect of longitudinal seam generation for a crack-propagation model; wherein, (a) generating an effect graph for 100 epoch; (b) generating an effect map for 200 epoch; (c) generating an effect map for 300 epoch; (d) generating an effect map for 400 epoch; (e) generating an effect map for 500 epoch; (f) generating an effect map for 600 epoch;
FIG. 7 is a graph of the variation in crack detection accuracy before and after data set augmentation.
The details of the application are explained in further detail below with reference to the drawings and examples.
Detailed Description
The following specific embodiments of the present application are provided, and it should be noted that the present application is not limited to the following specific embodiments, and all equivalent changes made on the basis of the technical scheme of the present application fall within the protection scope of the present application.
Example 1
The embodiment discloses a pavement crack image virtual augmentation model building method, which comprises the following steps:
step 1: collecting pavement crack images, and preprocessing the pavement crack images to obtain a real pavement crack image set, wherein pavement cracks are transverse cracks or longitudinal cracks;
step 2: establishing a DCGAN generation countermeasure network model, wherein the DCGAN generation countermeasure network model comprises a generator network and a discriminator network, and penalty items are arranged behind loss functions of the generator network and the discriminator network;
step 3: and (3) acquiring random noise, inputting the random noise and the real pavement crack image set acquired in the step (1) into the DCGAN generated countermeasure network model acquired in the step (2) to train, wherein the trained generator network model is the virtual augmentation model of the pavement crack image.
The generator network is used for inputting random noise and outputting to generate a pavement crack image set, and the discriminator is used for inputting to generate the pavement crack image set and outputting a true probability value by the true pavement crack image set.
Specifically, the random noise is generated by using a TensorFlow.
Specifically, the preprocessing in step 1 includes operations of sequentially performing color removal, denoising, contrast enhancement, brightness enhancement and segmentation into uniform sizes, and the quality of the preprocessed image data is improved.
Specifically, the true probability value in step 3 can characterize the similarity degree of the picture generated by the generator and the true picture, and if the probability value is greater than 0.5, the input of the true probability value can be considered to be derived from the true data set; if the probability value is smaller than 0.5, the input of the probability value is the data newly generated by the generation model;
specifically, the loss function of each training batch of the model is shown in a formula (1),is the true probability value, y, output by the jth training batch of the discriminant network j The model is a predicted probability value, N is the total number of training batches, and the loss function value of the model is obtained by summing the loss function values of each batch and then averaging the sum:
specifically, a gradient descent method is employed herein to update the values of the weights w and offsets b in each layer.
Specifically, the gradient is obtained through a back propagation algorithm, and in addition, training data are divided into small batches (mini-batches) for improving the training speed. In each small batch, a small fraction of training samples was randomly selected for gradient calculation at each step. This has the advantage that the training speed can be increased.
Each cycle is defined as the process of a single complete training of training data, the entire training process requiring a plurality of such cycles. At the end of each cycle, the loss function used on the validation set is evaluated, and the network that minimizes the validation set loss function is selected as the best choice. The generator is trained according to the magnitude of the loss function, and the training direction of the model is guided.
In deep convolutional neural networks, sample data is the basis for network training, and regularization constraints are guidelines for guaranteeing the direction of network learning. In order to reduce test errors in the learning of noise by a network, enhance the generalization capability of a model, enable the model to be smoother and prevent the occurrence of the over-fitting phenomenon, the application selects a strategy using L1 regularization constraint.
Specifically, the penalty term-added loss function isAnd is also provided withWherein X represents the input of the generator network and the arbiter network, y represents the output of the generator network and the arbiter network, w is a weight vector, alpha represents the regularization coefficient ranging from 0 to 1, and w is 1 Is the sum of the absolute values of the weights w in the model.
The gradient can be obtained by solving the above steps:
when w is positive, w becomes smaller after updating; when w is negative, w becomes larger after updating. And the method is as close to 0 as possible, so that the complexity of the network is reduced, and the phenomenon of overfitting is prevented. When w=0, |w| is not conductive, let sign (0) =0, and the case where w is equal to 0 is included in the regularization constraint.
L1 regularization utilizes sparse characteristics to select features, meaningful features are selected from feature subsets, and the feature greatly simplifies deep learning feature extraction problem, so that the generating effect of the generator is better.
Specifically, the generator network has a total of 5 layers, with 4 convolutional layers and 1 fully connected layer, and specifically operates to take as the input to the generator a 100-dimensional random variable subject to (0, 1) uniform distribution. Firstly, the image is passed through a full-connection layer to obtain a 4×4×1024 image, then passed through a transposed convolution layer of 4 convolution kernels and adopting 5×5 so as to continuously enlarge the image, and finally obtain the crack image output by the generator. In the whole generator structure, the Relu activation function is used by other layers in the network except the first layer which does not use the activation function.
Specifically, the input of the discriminator is the generated crack image and the real crack image, and the output is a probability value for judging the authenticity, namely the value can discriminate the authenticity of the input image. The arbiter network employed herein has a total of 5 layers, with 4 convolutional layers and 1 fully connected layer. An image of 64×64×3 at the input of the discriminator. Firstly, the image is sampled continuously for 4 times, the step length is 2, the size of the convolution kernel is 5 multiplied by 5, so that the image can be gradually reduced, and finally, a probability value representing the authenticity of the input image is obtained through the full connection layer. The entire arbiter structure uses the leakyrelu function except for the last layer, which does not use the activation function.
Preferably, the optimizers of the generator and the arbiter are set as Adam optimizers, and the learning rates are all set to 0.0002, and the convolution kernel sizes are all 5*5.
The network parameter settings of the deep convolution generation countermeasure network constructed by the application are shown in table 1;
table 1 DCGAN network parameter settings
As shown in FIG. 2, the network is composed of two parts, namely a generator and a discriminator
As shown in fig. 3, a generator network configuration diagram. The network structure is deconvolution (deconvolution), which is also commonly called transpose convolution (Transposed Convolution), i.e. the feature map size is gradually increased by learning. The input for generating the model in the training process is generally random noise Z, and new data G (Z) consistent with the real data distribution as much as possible is directionally or non-directionally output through a plurality of layers of transposed convolution calculation. Whether a generating network can generate high quality results depends on how the structure and parameters of the generating network are optimized.
Fig. 4 shows a network structure diagram of the arbiter. The network structure is similar to the traditional cyclic convolutional neural network, and the characteristic of the input image is extracted mainly through a plurality of convolutions, so that the size of the characteristic diagram of the hidden layer is gradually reduced. The method comprises the steps of simultaneously inputting a real data set X and new data G (z) generated by a generation model, obtaining an output probability value, judging whether the input of the probability value is from the real data set, and if the probability value is greater than 0.5, determining that the input of the probability value is from the real data set; if the probability value is less than 0.5, it is indicated that the data newly generated for generating the model is input.
The method selects a Faster R-CNN detection model to evaluate the quality of the generated crack image and the reliability of the enhanced crack image data set.
Example 2
The embodiment discloses an image virtual augmentation method, which comprises the following steps:
random noise is acquired, the random noise is input into a pavement crack image virtual augmentation model obtained by any pavement crack image virtual augmentation model building method, and a pavement crack image augmentation data set is obtained.
Fig. 5 is a graph of the effect of the road surface transverse seam generated by the final optimized augmented model. It can be seen from fig. 5 that as the number of iterations increases, the depth convolution produces a more realistic map of the road surface transverse seam created against the network model.
Fig. 6 is a graph of the road longitudinal joint effect generated by the final optimized augmented model. It can be seen from fig. 6 that as the number of iterations increases, the depth convolution produces a more realistic map of the road longitudinal joint effect against the network model.
FIG. 7 is a graph of the variation in crack detection accuracy before and after data set augmentation. The X-axis represents the virtual crack image generated by the generated type countermeasure network, the Y-axis represents the average accuracy of crack detection, and the orange curve represents the crack detection accuracy change curve of the Faster R-CNN detection model after the 1000 real crack image data sets are amplified. It can be seen from the graph that the crack detection accuracy is continuously improved as the number of virtual crack images is gradually increased. Experiments show that under the condition that the number of the real crack images is certain, after the original training set is amplified by using the generated type countermeasure network, the crack detection accuracy of the crack detection model can be improved to a certain extent.
Claims (3)
1. The method for establishing the virtual augmentation model of the pavement crack image is characterized by comprising the following steps of:
step 1: collecting pavement crack images, and preprocessing the pavement crack images to obtain a real pavement crack image set, wherein pavement cracks are transverse cracks or longitudinal cracks;
step 2: establishing a DCGAN generation countermeasure network model, wherein the DCGAN generation countermeasure network model comprises a generator network and a discriminator network, and penalty items are arranged behind loss functions of the generator network and the discriminator network;
the generator network has 5 layers, the first layer is a full-connection layer, the second to the fifth layers are convolution layers, and the convolution layers all adopt Relu as an activation function; the input of the generator is a random variable which is 100-dimensional and is subject to (0, 1) uniform distribution, an image obtained through a full-connection layer passes through a transposed convolution layer with the size of 5 multiplied by 5 of 4 convolution kernels, and finally a crack image output by the generator is obtained;
the arbiter network has 5 layers, wherein the first layer to the fourth layer are convolution layers, the fifth layer is a full-connection layer, and the convolution layers all adopt the releasyrlu as an activation function; the input of the discriminator is a generated crack image and a real crack image, the image is subjected to 4 times of continuous downsampling, the step length is 2, the size of the convolution kernel is 5 multiplied by 5, and finally a probability value representing the authenticity of the input image is obtained through a full connection layer;
the penalty term added loss function isAnd->Wherein X represents the input of the generator network and the arbiter network, y represents the output of the generator network and the arbiter network, w is a weight vector, alpha represents the regularization coefficient ranging from 0 to 1, and w is 1 Is the sum of the absolute values of the weights w in the model;
step 3: and (3) acquiring random noise, inputting the random noise and the real pavement crack image set acquired in the step (1) into the DCGAN generated countermeasure network model acquired in the step (2) to train, wherein the trained generator network model is the virtual augmentation model of the pavement crack image.
2. The method for creating a virtual augmented model of a pavement crack image according to claim 1, wherein the preprocessing in step 1 includes the operations of sequentially performing the color removal, the noise removal, the contrast enhancement, the brightness enhancement, and the division into uniform sizes.
3. An image virtual augmentation method, comprising the steps of:
obtaining random noise, inputting the random noise into the road surface crack image virtual augmentation model obtained by the road surface crack image virtual augmentation model establishing method according to claim 1 or 2, and obtaining a road surface crack image augmentation data set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010574126.5A CN111861906B (en) | 2020-06-22 | 2020-06-22 | Pavement crack image virtual augmentation model establishment and image virtual augmentation method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010574126.5A CN111861906B (en) | 2020-06-22 | 2020-06-22 | Pavement crack image virtual augmentation model establishment and image virtual augmentation method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111861906A CN111861906A (en) | 2020-10-30 |
CN111861906B true CN111861906B (en) | 2023-10-31 |
Family
ID=72987142
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010574126.5A Active CN111861906B (en) | 2020-06-22 | 2020-06-22 | Pavement crack image virtual augmentation model establishment and image virtual augmentation method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111861906B (en) |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112415514B (en) * | 2020-11-16 | 2023-05-02 | 北京环境特性研究所 | Target SAR image generation method and device |
CN112396110B (en) * | 2020-11-20 | 2024-02-02 | 南京大学 | Method for generating augmented image of countermeasure cascade network |
CN112862706A (en) * | 2021-01-26 | 2021-05-28 | 北京邮电大学 | Pavement crack image preprocessing method and device, electronic equipment and storage medium |
CN113009447B (en) * | 2021-03-05 | 2023-07-25 | 长安大学 | Road underground cavity detection and early warning method based on deep learning and ground penetrating radar |
CN113222114B (en) * | 2021-04-22 | 2023-08-15 | 北京科技大学 | Image data augmentation method and device |
CN113592000A (en) * | 2021-08-03 | 2021-11-02 | 成都理工大学 | Convolution-based crack identification technology for antagonistic neural network |
CN113822880A (en) * | 2021-11-22 | 2021-12-21 | 中南大学 | Crack identification method based on deep learning |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018028255A1 (en) * | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
CN108876780A (en) * | 2018-06-26 | 2018-11-23 | 陕西师范大学 | Bridge Crack image crack detection method under a kind of complex background |
CN109345507A (en) * | 2018-08-24 | 2019-02-15 | 河海大学 | A kind of dam image crack detection method based on transfer learning |
CN110956591A (en) * | 2019-11-06 | 2020-04-03 | 河海大学 | Dam crack image data enhancement method based on depth convolution generation countermeasure network |
-
2020
- 2020-06-22 CN CN202010574126.5A patent/CN111861906B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018028255A1 (en) * | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
CN108876780A (en) * | 2018-06-26 | 2018-11-23 | 陕西师范大学 | Bridge Crack image crack detection method under a kind of complex background |
CN109345507A (en) * | 2018-08-24 | 2019-02-15 | 河海大学 | A kind of dam image crack detection method based on transfer learning |
CN110956591A (en) * | 2019-11-06 | 2020-04-03 | 河海大学 | Dam crack image data enhancement method based on depth convolution generation countermeasure network |
Non-Patent Citations (2)
Title |
---|
基于生成式对抗网络的裂缝图像修复方法;胡敏;李良福;;计算机应用与软件(第06期);全文 * |
复杂背景下基于图像处理的桥梁裂缝检测算法;李良福;孙瑞;;激光与光电子学进展(第06期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN111861906A (en) | 2020-10-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111861906B (en) | Pavement crack image virtual augmentation model establishment and image virtual augmentation method | |
CN109035149B (en) | License plate image motion blur removing method based on deep learning | |
CN110188685B (en) | Target counting method and system based on double-attention multi-scale cascade network | |
CN108596258B (en) | Image classification method based on convolutional neural network random pooling | |
CN108985317B (en) | Image classification method based on separable convolution and attention mechanism | |
CN109345507B (en) | Dam image crack detection method based on transfer learning | |
CN113313657B (en) | Unsupervised learning method and system for low-illumination image enhancement | |
CN110827213A (en) | Super-resolution image restoration method based on generation type countermeasure network | |
CN108847223B (en) | Voice recognition method based on deep residual error neural network | |
CN108596818B (en) | Image steganalysis method based on multitask learning convolutional neural network | |
CN108764250B (en) | Method for extracting essential image by using convolutional neural network | |
CN106709877A (en) | Image deblurring method based on multi-parameter regular optimization model | |
CN112489164B (en) | Image coloring method based on improved depth separable convolutional neural network | |
CN112365514A (en) | Semantic segmentation method based on improved PSPNet | |
CN111915522A (en) | Image restoration method based on attention mechanism | |
CN111046917B (en) | Object-based enhanced target detection method based on deep neural network | |
CN110223231A (en) | A kind of rapid super-resolution algorithm for reconstructing of noisy image | |
CN112884668A (en) | Lightweight low-light image enhancement method based on multiple scales | |
Jin et al. | Defect identification of adhesive structure based on DCGAN and YOLOv5 | |
CN111382684B (en) | Angle robust personalized facial expression recognition method based on antagonistic learning | |
Xiang et al. | Optical flow estimation using spatial-channel combinational attention-based pyramid networks | |
CN114723049A (en) | Class activation mapping method and device based on gradient optimization | |
CN114066744A (en) | Artistic image restoration method and system based on style constraint | |
CN113657290A (en) | Snail collection and fine classification recognition system | |
CN115457611B (en) | Vein recognition method based on characteristic decoupling 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |