CN112614073A - Image rain removing method based on visual quality evaluation feedback and electronic device - Google Patents
Image rain removing method based on visual quality evaluation feedback and electronic device Download PDFInfo
- Publication number
- CN112614073A CN112614073A CN202011593324.2A CN202011593324A CN112614073A CN 112614073 A CN112614073 A CN 112614073A CN 202011593324 A CN202011593324 A CN 202011593324A CN 112614073 A CN112614073 A CN 112614073A
- Authority
- CN
- China
- Prior art keywords
- image
- rain
- rainy
- day
- quality evaluation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 36
- 238000013441 quality evaluation Methods 0.000 title claims abstract description 20
- 230000000007 visual effect Effects 0.000 title claims abstract description 13
- 238000012549 training Methods 0.000 claims abstract description 14
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 12
- 238000013528 artificial neural network Methods 0.000 claims abstract description 3
- 238000004590 computer program Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 6
- 238000001303 quality assessment method Methods 0.000 claims description 5
- 238000002834 transmittance Methods 0.000 claims description 5
- 230000015556 catabolic process Effects 0.000 abstract description 3
- 238000006731 degradation reaction Methods 0.000 abstract description 3
- 230000008569 process Effects 0.000 abstract description 3
- 230000000694 effects Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000004451 qualitative analysis Methods 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 238000007670 refining Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- 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
-
- 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
-
- 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
- G06T2207/10004—Still image; Photographic 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]
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
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)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image rain removing method and an electronic device based on visual quality evaluation feedback, which comprises the steps of constructing a paired image data set based on a plurality of sample rain-free images and generated rain marks and rain fog; carrying out rain removing treatment on collected sample rainy-day images, carrying out artificial visual quality evaluation on each treated sample rainy-day image, and constructing a non-paired rainy-day image quality data set; training a first convolution neural network by using an unpaired rainy day image quality data set to obtain a quality evaluation network; training a second convolutional neural network by using the paired image data set and the quality evaluation network constraint to obtain an image rain removal model; and inputting the image to be processed into an image rain removal model to obtain an image after rain removal. The invention uses real rainy images to participate in model training, so that the model can learn to process richer and real degradation types, and visual quality evaluation feedback is introduced, so that the generated rain-removing image has better quality in the subjective sense of human eyes.
Description
Technical Field
The invention belongs to the field of image processing and enhancement, and relates to an image rain removing method based on visual quality evaluation feedback and an electronic device.
Background
The deep learning rain-removing era started in 2017. Yang et al construct a network that combines rain mark detection and removal to handle heavy rain, overlapping rain marks and rain fog. The network can detect the position of rain by predicting a binary mask, and remove rain marks by adopting a recursive framework to gradually remove rain fog. The method achieves good effect under the condition of heavy rain. However, the method may erroneously remove the vertical texture and cause underexposure.
In the same year, Fu et al tried to remove rain marks by constructing a deep detail network. The network takes only high frequency details as input and predicts rain marks and clean, rain-free images. This work shows that removing background information from the network input facilitates network training.
Following the work of Yang and Fu et al, in subsequent work, a number of convolutional neural network-based approaches have been proposed. These methods employ more advanced network structures and embed new priors associated with rain, yielding better results in both quantitative and qualitative analysis. However, since these methods are limited by the fully supervised learning paradigm (using a composite rain map and using only constraints based on signal fidelity metrics), the model training targets are not completely consistent with human perception when dealing with real rain scenes never seen in the training process, and in practical application scenarios, the enhanced model may fail.
Disclosure of Invention
Aiming at the problems and the defects of related methods, the invention provides an image rain removing method based on visual quality evaluation feedback and an electronic device, wherein a rain removing neural network model is trained based on a signal fidelity reconstruction loss function and a quality evaluation loss function based on a subjective quality evaluation method, the former ensures that the trained model can realize a rain mark removing effect on synthetic data, and the latter ensures the generalization performance of the method, including the processing effect of non-visible degradation and the improvement of subjective intuition quality.
The technical scheme adopted by the invention comprises the following steps:
an image rain removing method based on visual quality evaluation feedback comprises the following steps:
1) generating a plurality of rainy-day images based on the rain-free images of the samples and the generated rain marks and rain fog, and constructing a paired image data set;
2) carrying out rain removing treatment on collected sample rainy-day images, carrying out artificial visual quality evaluation on each treated sample rainy-day image to obtain an image quality label of each treated sample rainy-day image, and constructing a non-paired rainy-day image quality data set;
3) training a first convolution neural network by using an unpaired rainy day image quality data set to obtain a quality evaluation network;
4) training a second convolutional neural network by using the paired image data set and the quality evaluation network constraint to obtain an image rain removal model;
5) and inputting the image to be processed into an image rain removal model to obtain an image after rain removal.
Further, the method for generating rain marks and rain fog comprises the following steps: a rain mark appearance model was used.
Further, the parameters of the rain fog include: light transmittance and background light.
Further, the rainy day image y is x (1-t) + t α + s, where x is a sample no-rain image, s is rain mark, t is light transmittance, and α is background light.
Further, the method for carrying out rain removing treatment on the sample rainy day image comprises the following steps: DID-MDN, SSIR, SPANet, or JORDER-E.
Further, the first convolutional neural network has a structure of replacing the last layer with a VGG16 network having n cell full-connected layers and a softmax layer, where n is the number of kinds of image quality labels.
Further, the first convolutional neural network is pre-trained on ImageNet prior to training the first convolutional neural network.
Further, the structure of the second convolutional network includes: and a U-Net structure.
Further, by a signal fidelity measure LFidAnd a mass loss function LQualityConstraining learning of a second convolutional neural network, wherein the signal fidelity measureFunction of mass lossIs an index of the similarity, and is,for the image after rain removal, y is the image in rainy days in the paired image dataset, D is the quality assessment network, and l is a random number.
A storage medium having a computer program stored therein, wherein the computer program is arranged to perform the above-mentioned method when executed.
An electronic device comprising a memory having a computer program stored therein and a processor arranged to run the computer to perform the method as described above.
Compared with the prior art, the invention has the following advantages:
1) besides the synthesized training data, the real rainy-day images are used to participate in model training, so that the model can learn to process richer and real degradation types;
2) visual quality evaluation feedback is introduced, so that the generated rain-removing image has better quality in the subjective sense of human eyes.
Drawings
FIG. 1 is a frame diagram of subjective quality assessment of a rain removal method.
FIG. 2 is a frame diagram of an image rain removal model of the present invention.
Detailed Description
In order to further explain the technical method of the invention, the invention is further described in detail by combining the drawings and the specific examples in the specification.
The image rain removing method comprises the following steps:
step 1: a rain/no rain paired image dataset was constructed for a total of 1800 rain/no rain image pairs. According to a rainless image x, corresponding raindrop s and rainfog parameters (light transmittance t and background light alpha) are generated based on a raindrop appearance model (random sampling generation illumination direction parameters, viewing angle parameters and raindrop vibration parameters) [ Garg and Nayar,2016], relevant variables are superposed, and a rainy day image y is generated as x (1-t) + t alpha + s, so that a paired image data set is constructed.
Step 2: constructing an unpaired rainy day image quality data set, collecting rainy day images through a public channel, running a plurality of forefront rain removing methods (such as DID-MDN [ Zhang and Patel,2018], SSIR [ Wei et al, 2019], SPANet [ Hu et al, 2019] and JORDER-E [ Yang et al, 2020] and the like) to remove rain from the rainy day images, and evaluating the visual quality of a processing result by an applicant, wherein corresponding image quality labels (1-10 grades, 10 represents the highest quality, and 1 represents the lowest quality).
And step 3: the quality assessment network D is trained using unpaired rain image quality data sets. As shown in fig. 1, D uses a network structure of VGG16 and replaces the last layer with a fully connected layer of 10 cells. Then softmax is followed. The network was pre-trained on ImageNet and then refined using a rain image quality data set. The score range [1,10], 10 is the highest score.
And 4, step 4: as shown in fig. 2, the rain-removed network is trained using paired image datasets and quality assessment network D constraints. The rain removal network uses a U-Net structure, uses the convolution layer with the step length of 1 to carry out feature extraction and refining, and uses the convolution layer with the step length of 2 to gradually reduce the space size of the feature. Then, feature extraction and refinement are continued using the convolution layer with step size 1, and the spatial size of the feature is gradually enlarged to the original image size using the deconvolution layer with step size 2. Features at the same spatial scale but belonging to different ends of the size scaling are connected using a jump connection, thereby ensuring that information flows smoothly at each scale and maintaining the underlying fine signal structure. Using a signal fidelity metric LFidAnd a mass loss function LQualityAnd (5) restricting the learning of the network. The total loss function L can be expressed as:
L=LFid+λLQuality,
where λ is a weight parameter. l is a random number between 7 and 12,the image after rain removal, where 10 represents the highest quality in the database. Phi (-) is a structural similarity index.
And 5: in practical application, the image x in rainy days is directly input into the image rain removing model, and the image after rain removal is output
The above examples are provided only for the purpose of describing the present invention, and are not intended to limit the scope of the present invention. The scope of the invention is defined by the appended claims. Various equivalent substitutions and modifications can be made without departing from the spirit and principles of the invention, and are intended to be within the scope of the invention.
Claims (10)
1. An image rain removing method based on visual quality evaluation feedback comprises the following steps:
1) generating a plurality of rainy-day images based on the rain-free images of the samples and the generated rain marks and rain fog, and constructing a paired image data set;
2) carrying out rain removing treatment on collected sample rainy-day images, carrying out artificial visual quality evaluation on each treated sample rainy-day image to obtain an image quality label of each treated sample rainy-day image, and constructing a non-paired rainy-day image quality data set;
3) training a first convolution neural network by using an unpaired rainy day image quality data set to obtain a quality evaluation network;
4) training a second convolutional neural network by using the paired image data set and the quality evaluation network constraint to obtain an image rain removal model;
5) and inputting the image to be processed into an image rain removal model to obtain an image after rain removal.
2. The method of claim 1, wherein the method of generating rain marks and fog comprises: using a rain mark appearance model; the method for carrying out rain removing treatment on the sample rainy day image comprises the following steps: DID-MDN, SSIR, SPANet, or JORDER-E.
3. The method of claim 1, wherein the parameters of rain fog comprise: light transmittance and background light.
4. The method of claim 3, wherein the rain image y is x (1-t) + t α + s, where x is a sample no rain image, s is rain mark, t is light transmittance, and α is background light.
5. The method of claim 1, wherein the first convolutional neural network is structured as a VGG16 network with n cell fully-connected layers replacing the last layer and a softmax layer, where n is the number of kinds of image quality labels.
6. The method of claim 1, wherein the first convolutional neural network is pre-trained on ImageNet prior to training the first convolutional neural network.
7. The method of claim 1, wherein the structure of the second convolutional network comprises: and a U-Net structure.
8. The method of claim 1, wherein the signal fidelity measure L is passedFidAnd a mass loss function LQualityConstraining learning of a second convolutional neural network, wherein the signal fidelity measureFunction of mass loss Phi (-) is the similarity index,for the image after rain removal, y is the image in rainy days in the paired image dataset, D is the quality assessment network, and l is a random number.
9. A storage medium having a computer program stored thereon, wherein the computer program is arranged to, when run, perform the method of any of claims 1-8.
10. An electronic device comprising a memory having a computer program stored therein and a processor arranged to run the computer program to perform the method according to any of claims 1-8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011593324.2A CN112614073A (en) | 2020-12-29 | 2020-12-29 | Image rain removing method based on visual quality evaluation feedback and electronic device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011593324.2A CN112614073A (en) | 2020-12-29 | 2020-12-29 | Image rain removing method based on visual quality evaluation feedback and electronic device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112614073A true CN112614073A (en) | 2021-04-06 |
Family
ID=75248816
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011593324.2A Pending CN112614073A (en) | 2020-12-29 | 2020-12-29 | Image rain removing method based on visual quality evaluation feedback and electronic device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112614073A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2023160645A1 (en) * | 2022-02-25 | 2023-08-31 | 索尼集团公司 | Image enhancement method and device |
CN117152000A (en) * | 2023-08-08 | 2023-12-01 | 华中科技大学 | Rainy day image-clear background paired data set manufacturing method and device and application thereof |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110111288A (en) * | 2019-04-15 | 2019-08-09 | 电子科技大学 | Image enhancement and blind plot quality based on depth assisted learning evaluate network |
CN110942436A (en) * | 2019-11-29 | 2020-03-31 | 复旦大学 | Image deblurring method based on image quality evaluation |
-
2020
- 2020-12-29 CN CN202011593324.2A patent/CN112614073A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110111288A (en) * | 2019-04-15 | 2019-08-09 | 电子科技大学 | Image enhancement and blind plot quality based on depth assisted learning evaluate network |
CN110942436A (en) * | 2019-11-29 | 2020-03-31 | 复旦大学 | Image deblurring method based on image quality evaluation |
Non-Patent Citations (1)
Title |
---|
WENHAN YANG等: "Frame-Consistent Recurrent Video Deraining with Dual-Level Flow", 《IEEE》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2023160645A1 (en) * | 2022-02-25 | 2023-08-31 | 索尼集团公司 | Image enhancement method and device |
CN117152000A (en) * | 2023-08-08 | 2023-12-01 | 华中科技大学 | Rainy day image-clear background paired data set manufacturing method and device and application thereof |
CN117152000B (en) * | 2023-08-08 | 2024-05-14 | 华中科技大学 | Rainy day image-clear background paired data set manufacturing method and device and application thereof |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Din et al. | A novel GAN-based network for unmasking of masked face | |
Zhu et al. | Hard sample aware noise robust learning for histopathology image classification | |
Huang et al. | Semi-supervised neuron segmentation via reinforced consistency learning | |
Tursun et al. | MTRNet++: One-stage mask-based scene text eraser | |
CN112614073A (en) | Image rain removing method based on visual quality evaluation feedback and electronic device | |
CN113762138A (en) | Method and device for identifying forged face picture, computer equipment and storage medium | |
CN111127354A (en) | Single-image rain removing method based on multi-scale dictionary learning | |
Zhang et al. | Diffusionad: Denoising diffusion for anomaly detection | |
CN110852199A (en) | Foreground extraction method based on double-frame coding and decoding model | |
CN115358337A (en) | Small sample fault diagnosis method and device and storage medium | |
CN114266894A (en) | Image segmentation method and device, electronic equipment and storage medium | |
CN114494168A (en) | Model determination, image recognition and industrial quality inspection method, equipment and storage medium | |
CN114399638A (en) | Semantic segmentation network training method, equipment and medium based on patch learning | |
CN114445356A (en) | Multi-resolution-based full-field pathological section image tumor rapid positioning method | |
Xiao et al. | Self-explanatory deep salient object detection | |
CN116701647A (en) | Knowledge graph completion method and device based on fusion of embedded vector and transfer learning | |
Liu et al. | Explaining deep neural networks using unsupervised clustering | |
CN115810128A (en) | Image classification model compression method based on greedy strategy | |
Mandal et al. | Neural architecture search for image dehazing | |
CN114496099A (en) | Cell function annotation method, device, equipment and medium | |
CN113095328A (en) | Self-training-based semantic segmentation method guided by Gini index | |
Punia et al. | Automatic detection of liver in CT images using optimal feature based neural network | |
Shi et al. | Refactoring ISP for High-Level Vision Tasks | |
Agnihotri | DeepFake Detection using Deep Neural Networks | |
CN112257769B (en) | Multilayer nuclear magnetic image classification method and system based on reinforcement learning type brain reading |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210406 |