CN110443763A - A kind of Image shadow removal method based on convolutional neural networks - Google Patents
A kind of Image shadow removal method based on convolutional neural networks Download PDFInfo
- Publication number
- CN110443763A CN110443763A CN201910705551.0A CN201910705551A CN110443763A CN 110443763 A CN110443763 A CN 110443763A CN 201910705551 A CN201910705551 A CN 201910705551A CN 110443763 A CN110443763 A CN 110443763A
- Authority
- CN
- China
- Prior art keywords
- shadow
- image
- neural network
- data set
- convolutional neural
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 36
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 33
- 238000012549 training Methods 0.000 claims abstract description 44
- 238000012360 testing method Methods 0.000 claims abstract description 10
- 238000013528 artificial neural network Methods 0.000 claims description 29
- 230000006870 function Effects 0.000 claims description 12
- 238000011176 pooling Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 238000005286 illumination Methods 0.000 claims description 3
- 230000007547 defect Effects 0.000 claims description 2
- 238000000059 patterning Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 20
- 238000001514 detection method Methods 0.000 description 8
- 238000012545 processing Methods 0.000 description 8
- 230000004927 fusion Effects 0.000 description 4
- 230000000873 masking effect Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 230000015556 catabolic process Effects 0.000 description 2
- 239000003086 colorant Substances 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The Image shadow removal method based on convolutional neural networks that the invention discloses a kind of constitutes Image shadow removal data set method includes the following steps: acquiring and collecting the shadow image under real scene and shadow-free image；Image shadow removal data set is pre-processed；The convolutional neural networks structure of forming end-to-end；Randomly select training set needed for the shadow image in data set constitutes network training；By training set using the convolutional neural networks end to end of training in the way of diversified；Using true picture and randomly select the shadow image composition test set in data set；Shadow removal is carried out using the convolutional neural networks end to end that training is completed using test set, obtains the shadow-free image of high quality.Method of the invention carries out Image shadow removal using full automatic method end to end, obtains the more visible shadow-free image consistent with original image color, texture, treatment of details effect is preferable.
Description
Technical Field
The invention belongs to the technical field of image processing, relates to a shadow removing method, and particularly relates to an image shadow removing method based on a convolutional neural network. Background
When the image is taken as multimedia information, the image is easily influenced by various conditions, so that the image generally has the phenomenon of quality reduction. Shadow is one of the phenomena, which is a quality degradation phenomenon caused by imaging conditions, and can cause the information amount reflected by a target to be defective or interfered, reduce the interpretation precision of an image, and seriously affect various quantitative analyses and applications of the image.
Shadow detection and removal is one of the most fundamental but challenging problems in the computer graphics and computer vision fields, and shadow removal of images is an important preprocessing stage for computer vision and image enhancement. The existence of the shadow not only affects the visual interpretation effect of the image, but also affects the analysis and subsequent processing results of the image. Therefore, it is necessary to detect and analyze the shadow of the image, so as to eliminate or reduce the influence of the shadow of the image, and increase the visual reality and physical reality of the image editing and processing.
Shadows are produced by different lighting conditions, and shadowless images are availableProduct of and shadow proportion(multiplication at pixel level) To represent shadow imagesThe following formula is shown.
（1）
The shadow removal is intended to generate a high-quality shadow-free image given a single shadow image, so that the texture, color, etc. of the original shadow image region are restored to a condition consistent with the shadow-free image. Existing methods of removing shadow areas typically include two steps: shadow detection and shadow removal. The methods firstly position the shadow area through shadow detection or manually mark the shadow area by a user, and then build a model to reconstruct the shadow area and the shadow area, thereby realizing the removal of the shadow.
However, shadow detection itself is an extremely challenging task. Traditional physics-based methods can only be applied to high quality images, while statistical learning-based methods rely on features that are manually labeled by the user. With the development of neural networks, Convolutional Neural Networks (CNN) learn the features of shadow detection, which overcomes the disadvantages of high quality image and manual labeling feature in the conventional method, but they are still limited to small network architecture due to less training data.
Likewise, even if the shadow region is known, it can still be challenging to remove it again. This is because the shadow detection effect can seriously affect the shadow removal result, and if the shadow detection effect is not good, the subsequent shadow removal effect is not possible to obtain a high-quality shadow-free image.
The interactive shadow removal method based on statistics adopts a rough manual labeling mode to detect the shadow, sacrifices the complete autonomy of finer shadow and wide and simpler user input; the shadow removal algorithm based on the Poisson equation does not consider the influence of ambient light and object material change, so that the texture recovery effect of the shadow area is poor. The image shadow removing algorithm based on the gradient domain solves part of defects based on the Poisson equation, but has poor processing effect on discontinuous or smaller shadow areas.
It can be seen from the analysis summary that the current shadow removal method cannot effectively recover the texture of the shadow area, or does not consider the influence of the environment and the material of the object, so as to maintain the visual consistency, and most methods are interactive rather than fully automatic methods, which greatly reduces the use efficiency.
Disclosure of Invention
The invention aims to obtain a high-quality shadow-free image, and provides an end-to-end deep convolutional neural network for removing image shadow, which can be used for removing shadow of images of intelligent traffic systems, medicine and the like.
In order to achieve the purpose, the invention adopts the following technical scheme:
an image shadow removing method based on a convolutional neural network comprises the steps of firstly collecting an image shadow removing data set, preprocessing the image shadow removing data set, then training and learning by using two-layer network structures of a shallow neural network and a deep neural network, and then inputting an original image (shadow image) through the trained network structures so as to achieve full-automatic shadow removing, and finally obtaining a high-quality shadow-free image.
The method comprises the following specific steps:
1) acquiring shadow images and shadow-free images in a real scene to form an image shadow removal data set;
2) preprocessing the image shadow removal data set;
3) constructing an end-to-end convolutional neural network structure;
4) randomly selecting shadow images in the data set to form a training set required by network training;
5) training an end-to-end convolutional neural network by using a training set in a diversified manner;
6) forming a test set by using the real image and the shadow image in the randomly selected data set;
7) and (4) carrying out shadow removal by using the trained end-to-end convolutional neural network by utilizing the test set to obtain a high-quality shadow-free image.
In the step 1), shadow images and shadow-free images in a real scene are collected to obtain a data set for removing image shadows:
in order to ensure the diversity of the image shadow removal data set, shadow images and shadow-free images projected by different objects are shot by using a fixed camera under different conditions of illumination intensity, scenes and the like, and the image shadow removal data set is constructed. Specifically, a plurality of scenes such as grasslands, campuses, streets and the like can be selected, shadow images projected by different objects are shot respectively in different weathers at the same time and different times of the same weather, meanwhile, shadow-free images corresponding to the original images, namely the shadow images, are shot to form shadow images and shadow-free image pairs, and the collection of the image shadow removal data set is completed.
In the step 2), the image shadow removal data set is preprocessed:
2-1) classifying and sorting the acquired image shadow data set according to soft and hard shadows and scene characteristics, forming an image pair by the shadow image and a corresponding shadow-free image, and expanding the data set of the same scene by means of cutting, rotating and the like;
2-2) arranging the images with different sizes and different pixels into images with a plurality of layers of pixels, wherein the pixel level isE.g. of、Etc., where n may take a positive integer.
The step 3) comprises a shallow neural network and a deep neural network, wherein the shallow neural network and the deep neural network have two network structures:
and 3-1) the shallow neural network is used for extracting rough image features and global semantic scene information, and obtaining image shadow mask factors from a fine to coarse mode. The shallow neural network is constructed based on a VGG16 network, the original network structure is finely adjusted, the shadow mask factor is obtained, all connection layers in the network are replaced by coiled layers, sub-sampling layers are not used at the same time, and the shadow mask factor is addedThe prediction layer of (1). Finally, the shallow neural network includes 16 convolutional layers, 5 max pooling layers, and 1 prediction layer.
3-2) the deep neural network is used for fusing a multi-context mechanism for local detail correction with the previous shallow network, so that the result is further improved, the prediction result of the whole network structure is more accurate, the edge processing effect is more precise, and a rough-to-fine mode is used for obtaining the shadow mask factor. In order to avoid increasing the network training burden, the invention defines a small network structure as a deep neural network, which comprises 5 convolutional layers, 2 pooling layers and 1 prediction layer.
3-3) to alleviate the occurrence of the over-fitting problem, and to some extent to achieve the regularization effect, drop is used after each convolution layer of the network, and the activation function used in the present invention is a Rectified linear unit (ReLU), which is defined as follows:
（2）
in the step 4), shadow images in the data set are randomly selected to form a training set required by network training;
in the step 5), training an end-to-end convolutional neural network by using a training set in a diversified manner:
5-1) training the network by using a diversified mode such as grading, layering and the like instead of a single mode, and finally realizing the rapid convergence of the network and effectively preventing overfitting;
5-2) shadow imageAnd its shadow maskThe relation between the shadow masks is given by formula (1), and during training, a real shadow mask is calculated according to a given shadow-free image pairThen, an attempt is made to learn a mapping function to create shadow imagesAnd a shadow maskThe relationship between;
5-3) during training, the network is constrained by the following overall loss, and the overall loss function comprises two loss functions and the fusion mode thereof as follows:
the first loss is calledPredicting the loss, approximating it as a loss function using the following formula:
（3）
wherein,representing true values at pixel q, if it is a shaded areaOtherwise, 0 is adopted;representing network structure based on parametersThe prediction obtained at pixel q.
The second loss, called composition loss, is the real shadow image RGB color and the shadow masking factor output by the unshaded image and the prediction layerThe difference between the RGB colors of the synthesized predicted shadow image is approximated using the following formula:
（4）
wherein,Nrepresenting the total number of training samples in the batch,representing shadow masking factors by predictionThe resulting predicted shadow image RGB channel,representing the actual shadow image RGB channel.
The overall loss experienced by the network is a linear fusion of the first two losses, following the equation:
（5）
here will beSet to 0.5.
In the step 6), a test set is formed by using the real image and the shadow image in the randomly selected data set;
in the step 7), shadow removal is performed by using the trained end-to-end convolutional neural network by using the test set, so that a high-quality shadow-free image is obtained.
The invention has the beneficial effects that:
(1) the quality reduction phenomenon caused by the shadow is eliminated to a great extent, and a satisfactory shadow removing effect is achieved;
(2) because the double-layer network is adopted as the whole grid structure, the network obtains the whole semantic information of the image by adopting two modes of from fine to coarse and from coarse to fine, the prediction result of the whole network structure is more accurate, the detail processing is better, and the quality of the obtained shadow-free image is higher;
(3) the invention inputs shadow images and outputs shadow-free images, thereby realizing a full-automatic end-to-end shadow removing method without interactively obtaining shadow masks through shadow detection or inputting shadow masks by users.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a diagram of the network architecture of the present invention;
FIG. 3 is the result of the image shadow removal applied to a simple scene;
fig. 4 shows the result of the image shadow removal applied to a complex scene.
Detailed Description
The invention is further described with reference to the following figures and examples.
As shown in fig. 1, the method comprises the following steps:
1) acquiring shadow images and shadow-free images in a real scene to obtain a data set with image shadow removed, and comprising the following steps:
1-1) selecting different scenes and different illumination intensities to collect data sets according to the characteristics of diversification of the data sets, specifically, selecting scenes such as a place, a road, a campus and the like, and collecting image shadow removal data sets at the same time in different weathers or at different times in the same weather respectively;
1-2) fixing a camera at a designated position by using a tripod according to a selected scene, and setting fixed parameters such as exposure compensation and focal length, wherein the used focal length is 4 mm, and the exposure compensation is 0 step;
1-3) utilizing a schoolbag, an umbrella, a human body and the like to project shadows to a designated area, and using a Bluetooth remote controller to connect a camera to shoot shadow images to obtain multi-shape shadow images and ensure the shape diversity characteristics of an image shadow removal data set;
1-4) withdrawing the projected object, removing the projection of the object, and using a Bluetooth remote controller to connect a camera to shoot a corresponding background image of the shadow image, namely a shadow-free image, so as to form an image shadow removal data set.
2) Preprocessing the image shadow removal dataset:
2-1) classifying and sorting the acquired image shadow data set according to soft and hard shadows and scene characteristics, forming an image pair by the shadow image and a corresponding shadow-free image, and expanding the data set of the same scene by means of cutting, rotating and the like;
2-2) sorting the images with different sizes and different pixels into a plurality of layered imagesImage of pixels at the pixel level ofE.g. of、Etc., where n may take a positive integer.
3) Constructing an end-to-end convolutional neural network structure, wherein the end-to-end convolutional neural network structure comprises a shallow neural network and a deep neural network:
and 3-1) the shallow neural network is used for extracting rough image features and global semantic scene information, and obtaining image shadow mask factors from a fine to coarse mode. The shallow neural network is constructed based on a VGG16 network, the original network structure is finely adjusted, the shadow mask factor is obtained, all the full-connection layers in the network are replaced by the lapping layers, the subsampling layers are not used, and a prediction layer of the shadow mask factor is added. Finally, the shallow neural network includes 16 convolutional layers, 5 max pooling layers, and 1 prediction layer.
3-2) the deep neural network is used for fusing a multi-context mechanism for local detail correction with the previous shallow network, so that the result is further improved, the prediction result of the whole network structure is more accurate, the edge processing effect is more precise, and a rough-to-fine mode is used for obtaining the shadow mask factor. In order to avoid increasing the network training burden, the invention defines a small network structure as a deep neural network, which comprises 5 convolutional layers, 2 pooling layers and 1 prediction layer.
3-3) to alleviate the occurrence of the over-fitting problem, and to some extent to achieve the regularization effect, drop is used after each convolution layer of the network, and the activation function used in the present invention is a Rectified linear unit (ReLU), which is defined as follows:
（2）
4) randomly selecting shadow images in the data set to form a training set required by network training;
5) training an end-to-end convolutional neural network in a diversified manner by using a training set:
5-1) training a shallow neural network and a deep neural network independently, and training the two networks in a cascade mode when the two networks reach certain precision, so that the combined optimization of the two networks is finally realized, and the effect of staged training is realized;
5-2) setting original images of shadow conditions of different levels to respectively train according to the size of the shadow scale factor, for example, firstly training an image data set of hard shadow, then training an image data set of soft shadow, and finally combining the image data set and the soft shadow to form a data set to train so as to realize a multi-level training effect;
5-3) considering the difference of pixel sizes of images input by users, dividing the images with different pixel sizes into a plurality of layers for training, realizing the training effect of the plurality of layers, finally realizing rapid convergence and preventing overfitting, and ensuring the diversification of training modes;
5-4) shadow imageAnd its shadow maskThe relation between the shadow masks is given by formula (1), and during training, a real shadow mask is calculated according to a given shadow-free image pairThen, an attempt is made to learn a mapping function to create shadow imagesAnd a shadow maskThe relationship between;
5-5) during training, the network is constrained by the following overall loss, and the overall loss function comprises two loss functions and the fusion mode thereof as follows:
the first loss is calledPredicting the loss, approximating it as a loss function using the following formula:
（3）
wherein,representing true values at pixel q, if it is a shaded areaOtherwise, 0 is adopted;representing network structure based on parametersThe prediction obtained at pixel q.
The second loss, called composition loss, is the real shadow image RGB color and the shadow masking factor output by the unshaded image and the prediction layerThe difference between the RGB colors of the synthesized predicted shadow image is approximated using the following formula:
（4）
wherein,Nrepresenting the total number of training samples in the batch,representing shadow masking factors by predictionThe resulting predicted shadow image RGB channel,representing the actual shadow image RGB channel.
The overall loss experienced by the network is a linear fusion of the first two losses, following the equation:
（5）
here will beSet to 0.5.
6) Forming a test set by using the real image and the shadow image in the randomly selected data set;
7) and (4) carrying out shadow removal by using the trained end-to-end convolutional neural network by utilizing the test set to obtain a high-quality shadow-free image.
The contents of the present invention can be further explained by the following simulation results.
1. Simulation content: the method of the invention is applied to remove the image shadows in different scenes.
2. Simulation result
Fig. 3 is a shadow image applied to a simple scene by the method of the present invention. Fig. 3 (a), (d), and (g) show shadow images in a simple scene, respectively; (c) (f) and (i) respectively represent real shadow-free images corresponding to the scenes (a), (d) and (g); (b) the (e) and (h) respectively represent the image shadow removal result under the simple scene obtained by the invention. Therefore, the shadow removing method has a good shadow removing effect on a simple scene, and a high-quality shadow-free image is obtained.
Fig. 4 is a shadow image of a complex scene in which the method of the present invention is applied. Fig. 4 (a), (d), and (g) show shadow images in a complex scene, respectively; (c) (f) and (i) respectively represent real shadow-free images corresponding to the scenes (a), (d) and (g); (b) the (e) and (h) respectively represent the image shadow removal result under the complex scene obtained by the invention. By utilizing the method provided by the invention, the shadow removing effect under the complex scene is good, the shadow image is effectively restored to the shadow-free image which is consistent with the texture, color and the like of the real shadow-free image, and the method is particularly good in the aspect of detail processing. By comprehensively analyzing the images in fig. 3 and fig. 4, it can be found that the invention shows a more ideal removal effect in both a simple scene and a complex scene, and reduces the quality degradation effect caused by the existence of shadows.
In summary, the invention provides a fully automatic image shadow removal model based on a convolutional neural network. The method can realize full-automatic image shadow removal through the deep convolutional neural network, reduces interactive behaviors, obtains an ideal image shadow removal effect, improves the efficiency of the method, well eliminates the quality reduction phenomenon caused by the shadow, and has great application value for target identification and target tracking in the later period.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.
Claims (7)
1. An image shadow removing method based on a convolutional neural network is characterized in that an image shadow removing data set is collected and preprocessed, then a two-layer network structure of a shallow neural network and a deep neural network is used for training and learning, then an original image (shadow image) is input through the trained network structure, so that full-automatic shadow removing is achieved, and finally a high-quality shadow-free image is obtained.
2. The method for removing the image shadow based on the convolutional neural network as claimed in claim 1, which mainly comprises the following steps:
1) acquiring shadow images and shadow-free images in a real scene to form an image shadow removal data set;
2) preprocessing the image shadow removal data set;
3) constructing an end-to-end convolutional neural network structure;
4) randomly selecting shadow images in the data set to form a training set required by network training;
5) training an end-to-end convolutional neural network by using a training set in a diversified manner;
6) forming a test set by using the real image and the shadow image in the randomly selected data set;
7) and (4) carrying out shadow removal by using the trained end-to-end convolutional neural network by utilizing the test set to obtain a high-quality shadow-free image.
3. The method for removing image shadows based on the convolutional neural network as claimed in claim 2, wherein in the step 1), shadow images and shadow-free images in real scenes are collected to obtain an image shadow-removed data set: under different illumination intensity, the camera is fixed by using the tripod, shadow images and shadow-free image pairs generated by different projection objects in different scenes are shot by using the Bluetooth remote controller, and the image shadow removal data set is acquired.
4. The convolutional neural network-based image shadow removal method as claimed in claim 2, wherein said step 2) preprocesses the image shadow removal data set: the data set is expanded by cutting, rotating and the like, the collected image shadow data set is classified and sorted according to soft and hard shadows and scene characteristics to form an image pair, and meanwhile, images with different sizes and different pixels are sorted into images with specified pixels.
5. The method as claimed in claim 2, wherein the step 3) is configured to construct an end-to-end convolutional neural network structure, which includes two layers of network structures, namely a shallow neural network and a deep neural network: the shallow neural network comprises 16 convolutional layers, 5 maximum pooling layers and 1 prediction layer; the deep neural network two layers comprise 5 convolutional layers, 2 pooling layers and 1 prediction layer; dropout is applied after each convolutional layer of the network, using the activation function ReLU.
6. The method as claimed in claim 2, wherein said step 5) trains the end-to-end convolutional neural network in a diversified manner by using the training set: the invention adopts a training mode of diversification such as grading, layering and the like to train the network, thereby realizing rapid convergence and preventing overfitting; while linearly fusing the predicted loss with the patterning loss as a function of the total loss of the network.
7. The method for removing the image shadow based on the convolutional neural network as claimed in claim 2, wherein the characteristics of the convolutional neural network are utilized, and the shadow area existing in the shadow image is removed aiming at the defect of the full-automatic end-to-end shadow removing method for removing the image shadow, so that the influence of the shadow on the image is effectively eliminated; meanwhile, multi-context scenes are considered, local information and edge information are processed, and the network comprises a shallow layer neural network structure and a deep layer neural network structure, so that a high-quality shadow-free image is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910705551.0A CN110443763B (en) | 2019-08-01 | 2019-08-01 | Convolutional neural network-based image shadow removing method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910705551.0A CN110443763B (en) | 2019-08-01 | 2019-08-01 | Convolutional neural network-based image shadow removing method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110443763A true CN110443763A (en) | 2019-11-12 |
| CN110443763B CN110443763B (en) | 2023-10-13 |
Family
ID=68432691
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910705551.0A Active CN110443763B (en) | 2019-08-01 | 2019-08-01 | Convolutional neural network-based image shadow removing method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110443763B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112115934A (en) * | 2020-09-16 | 2020-12-22 | 四川长虹电器股份有限公司 | Bill image text detection method based on deep learning example segmentation |
| CN112862714A (en) * | 2021-02-03 | 2021-05-28 | 维沃移动通信有限公司 | Image processing method and device |
| CN113139917A (en) * | 2021-04-23 | 2021-07-20 | Oppo广东移动通信有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN113178010A (en) * | 2021-04-07 | 2021-07-27 | 湖北地信科技集团股份有限公司 | High-resolution image shadow region restoration and reconstruction method based on deep learning |
| CN113222826A (en) * | 2020-01-21 | 2021-08-06 | 深圳富泰宏精密工业有限公司 | Document shadow removing method and device |
| CN113222845A (en) * | 2021-05-17 | 2021-08-06 | 东南大学 | Portrait external shadow removing method based on convolution neural network |
| CN113628129A (en) * | 2021-07-19 | 2021-11-09 | 武汉大学 | Method for removing shadow of single image by edge attention based on semi-supervised learning |
| US20220188991A1 (en) * | 2020-12-12 | 2022-06-16 | Samsung Electronics Co., Ltd. | Method and electronic device for managing artifacts of image |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6701026B1 (en) * | 2000-01-26 | 2004-03-02 | Kent Ridge Digital Labs | Method and apparatus for cancelling lighting variations in object recognition |
| US7366323B1 (en) * | 2004-02-19 | 2008-04-29 | Research Foundation Of State University Of New York | Hierarchical static shadow detection method |
| CN101477628A (en) * | 2009-01-06 | 2009-07-08 | 青岛海信电子产业控股股份有限公司 | Method and apparatus for vehicle shape removing |
| CN104079802A (en) * | 2013-03-29 | 2014-10-01 | 现代Mnsoft公司 | Method and apparatus for removing shadow from aerial or satellite photograph |
| CN105574821A (en) * | 2015-12-10 | 2016-05-11 | 浙江传媒学院 | Data-based soft shadow removal method |
| US9430715B1 (en) * | 2015-05-01 | 2016-08-30 | Adobe Systems Incorporated | Identifying and modifying cast shadows in an image |
| CN106447721A (en) * | 2016-09-12 | 2017-02-22 | 北京旷视科技有限公司 | Image shadow detection method and device |
| US20170169590A1 (en) * | 2015-12-09 | 2017-06-15 | Oregon Health & Science University | Systems and methods to remove shadowgraphic flow projections in oct angiography |
| KR20190071452A (en) * | 2017-12-14 | 2019-06-24 | 동국대학교 산학협력단 | Apparatus and method for object detection with shadow removed |
| CN109978807A (en) * | 2019-04-01 | 2019-07-05 | 西北工业大学 | A kind of shadow removal method based on production confrontation network |
-
2019
- 2019-08-01 CN CN201910705551.0A patent/CN110443763B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6701026B1 (en) * | 2000-01-26 | 2004-03-02 | Kent Ridge Digital Labs | Method and apparatus for cancelling lighting variations in object recognition |
| SG103253A1 (en) * | 2000-01-26 | 2004-04-29 | Kent Ridge Digital Labs | Method and apparatus for cancelling lighting variations in object recognition |
| US7366323B1 (en) * | 2004-02-19 | 2008-04-29 | Research Foundation Of State University Of New York | Hierarchical static shadow detection method |
| CN101477628A (en) * | 2009-01-06 | 2009-07-08 | 青岛海信电子产业控股股份有限公司 | Method and apparatus for vehicle shape removing |
| CN104079802A (en) * | 2013-03-29 | 2014-10-01 | 现代Mnsoft公司 | Method and apparatus for removing shadow from aerial or satellite photograph |
| US9430715B1 (en) * | 2015-05-01 | 2016-08-30 | Adobe Systems Incorporated | Identifying and modifying cast shadows in an image |
| US20170169590A1 (en) * | 2015-12-09 | 2017-06-15 | Oregon Health & Science University | Systems and methods to remove shadowgraphic flow projections in oct angiography |
| CN105574821A (en) * | 2015-12-10 | 2016-05-11 | 浙江传媒学院 | Data-based soft shadow removal method |
| CN106447721A (en) * | 2016-09-12 | 2017-02-22 | 北京旷视科技有限公司 | Image shadow detection method and device |
| KR20190071452A (en) * | 2017-12-14 | 2019-06-24 | 동국대학교 산학협력단 | Apparatus and method for object detection with shadow removed |
| CN109978807A (en) * | 2019-04-01 | 2019-07-05 | 西北工业大学 | A kind of shadow removal method based on production confrontation network |
Non-Patent Citations (11)
| Title |
|---|
| HIEU LE 等: "Shadow Removal via Shadow Image Decomposition", pages 8578 - 8587, Retrieved from the Internet <URL:https://arxiv.org/abs/1908.08628> * |
| O天涯海阁O: "CNN阴影去除--DeshadowNet: A Multi-context Embedding Deep Network for Shadow Removal", pages 1, Retrieved from the Internet <URL:https://blog.csdn.net/zhangjunhit/article/details/77675481> * |
| XIAODONG GU 等: "Image shadow removal using pulse coupled neural network", IEEE TRANSACTIONS ON NEURAL NETWORKS, pages 692 * |
| 宋全恒: ""基于计算机视觉的目标检测和阴影检测算法的研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 宋全恒: ""基于计算机视觉的目标检测和阴影检测算法的研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》, 15 September 2017 (2017-09-15), pages 138 - 295 * |
| 徐晓燕 等: "室外光源光谱辐照度与K-means结合的单幅图像阴影检测", 科学技术与工程, no. 04, pages 286 - 291 * |
| 月如辰: "ST-CGAN 用GAN实现阴影检测和阴影去除", pages 1, Retrieved from the Internet <URL:https://www.jianshu.com/p/cb7545ddd944> * |
| 熊俊涛 等: "自然光照条件下采摘机器人果实识别的表面阴影去除方法", 农业工程学报, no. 22, pages 147 - 154 * |
| 王荣本等: "识别阴影中智能车辆导航路径的神经网络方法研究", 《公路交通科技》 * |
| 王荣本等: "识别阴影中智能车辆导航路径的神经网络方法研究", 《公路交通科技》, no. 05, 20 October 2002 (2002-10-20), pages 99 - 102 * |
| 闫凤 等: "纹理损失最小约束下的跟踪图像阴影去除算法的改进", 现代电子技术, no. 24, pages 104 - 108 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113222826A (en) * | 2020-01-21 | 2021-08-06 | 深圳富泰宏精密工业有限公司 | Document shadow removing method and device |
| CN112115934A (en) * | 2020-09-16 | 2020-12-22 | 四川长虹电器股份有限公司 | Bill image text detection method based on deep learning example segmentation |
| US20220188991A1 (en) * | 2020-12-12 | 2022-06-16 | Samsung Electronics Co., Ltd. | Method and electronic device for managing artifacts of image |
| CN112862714A (en) * | 2021-02-03 | 2021-05-28 | 维沃移动通信有限公司 | Image processing method and device |
| CN113178010A (en) * | 2021-04-07 | 2021-07-27 | 湖北地信科技集团股份有限公司 | High-resolution image shadow region restoration and reconstruction method based on deep learning |
| CN113139917A (en) * | 2021-04-23 | 2021-07-20 | Oppo广东移动通信有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN113222845A (en) * | 2021-05-17 | 2021-08-06 | 东南大学 | Portrait external shadow removing method based on convolution neural network |
| CN113628129A (en) * | 2021-07-19 | 2021-11-09 | 武汉大学 | Method for removing shadow of single image by edge attention based on semi-supervised learning |
| CN113628129B (en) * | 2021-07-19 | 2024-03-12 | 武汉大学 | Edge attention single image shadow removing method based on semi-supervised learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110443763B (en) | 2023-10-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110443763B (en) | Convolutional neural network-based image shadow removing method | |
| Fan et al. | Integrating semantic segmentation and retinex model for low-light image enhancement | |
| Ram Prabhakar et al. | Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs | |
| Liu et al. | Benchmarking low-light image enhancement and beyond | |
| Engin et al. | Cycle-dehaze: Enhanced cyclegan for single image dehazing | |
| Ren et al. | Single image dehazing via multi-scale convolutional neural networks with holistic edges | |
| Fu et al. | LE-GAN: Unsupervised low-light image enhancement network using attention module and identity invariant loss | |
| Liu et al. | HoLoCo: Holistic and local contrastive learning network for multi-exposure image fusion | |
| CN108764372B (en) | Construction method and device, mobile terminal, the readable storage medium storing program for executing of data set | |
| CN106886977B (en) | Multi-image automatic registration and fusion splicing method | |
| US11636639B2 (en) | Mobile application for object recognition, style transfer and image synthesis, and related systems, methods, and apparatuses | |
| Le et al. | Deeply Supervised 3D Recurrent FCN for Salient Object Detection in Videos. | |
| CN103262119B (en) | For the method and system that image is split | |
| CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
| CN113052210A (en) | Fast low-illumination target detection method based on convolutional neural network | |
| CN110956681B (en) | Portrait background automatic replacement method combining convolution network and neighborhood similarity | |
| CN110070517A (en) | Blurred picture synthetic method based on degeneration imaging mechanism and generation confrontation mechanism | |
| Vidal et al. | Ug^ 2: A video benchmark for assessing the impact of image restoration and enhancement on automatic visual recognition | |
| CN111462162B (en) | Foreground segmentation algorithm for specific class pictures | |
| Hou et al. | Text-aware single image specular highlight removal | |
| CN114627269A (en) | Virtual reality security protection monitoring platform based on degree of depth learning target detection | |
| CN109002771A (en) | A kind of Classifying Method in Remote Sensing Image based on recurrent neural network | |
| Chen et al. | Improving dynamic hdr imaging with fusion transformer | |
| CN109840498A (en) | A kind of real-time pedestrian detection method and neural network, target detection layer | |
| Chen et al. | LENFusion: A Joint Low-Light Enhancement and Fusion Network for Nighttime Infrared and Visible Image Fusion |
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 |