CN112435192A - Lightweight image definition enhancing method - Google Patents
Lightweight image definition enhancing method Download PDFInfo
- Publication number
- CN112435192A CN112435192A CN202011373558.6A CN202011373558A CN112435192A CN 112435192 A CN112435192 A CN 112435192A CN 202011373558 A CN202011373558 A CN 202011373558A CN 112435192 A CN112435192 A CN 112435192A
- Authority
- CN
- China
- Prior art keywords
- loss
- image
- vgg
- model
- true
- 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 27
- 230000002708 enhancing effect Effects 0.000 title claims abstract description 6
- 238000012545 processing Methods 0.000 claims abstract description 19
- 238000012549 training Methods 0.000 claims abstract description 19
- 230000000694 effects Effects 0.000 claims abstract description 10
- 238000010276 construction Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 5
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 11
- 238000013135 deep learning Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000007547 defect Effects 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
-
- G06T5/80—
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20132—Image cropping
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The invention provides a lightweight image definition enhancing method which is characterized by comprising the following steps: step 1: constructing a network, and performing training iteration through a set weighted LOSS to obtain a model which is trained; LOSS includes VGG Loss, Huber Loss and mean absolute error LOSS MAE Loss; step 2: the trained model receives the image to be processed and outputs an effect image subjected to definition enhancement processing; the method corrects the model by setting a plurality of Loss values, and can ensure that the corrected model can better process the edge characteristics of the image in the picture compared with the traditional method of correcting the model only by the MAE Loss value.
Description
Technical Field
The invention relates to the field of deep learning, in particular to a lightweight image definition enhancing method.
Background
With the evolution of image algorithms, sharpness correction for blurred images is now increasingly optimized. At present, various definition enhancement functions exist in mobile terminal APP or PC image processing software, and the purpose of definition enhancement is mainly achieved by sharpening images through algorithms, such as sharpening algorithms, algorithms based on deep learning technology and the like. Compared with a sharpening algorithm, the definition algorithm based on the deep learning technology has the advantages that the definition of the image is improved more remarkably, but the model established by the deep learning is large, the algorithm is complex, and the algorithm is difficult to apply to a mobile terminal. The image at the mobile terminal is often processed clearly through a traditional sharpening algorithm. Although the traditional sharpening algorithm can also improve the definition of an image, an ideal effect cannot be obtained.
On the other hand, as the number of pixels to be photographed increases, the composition of an image becomes more complicated, and a higher computational demand is required for processing the image clearly. There is therefore a need for a lightweight image sharpness enhancement method suitable for mobile terminals.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a lightweight image definition enhancing method which is simple in structure and convenient to use, can obtain a definition enhanced image and is suitable for a mobile terminal.
A lightweight image definition enhancement method is characterized by comprising the following steps:
step 1: constructing a network, and performing training iteration through a set weighted LOSS to obtain a model which is trained; LOSS includes VGG Loss, Huber Loss and mean absolute error LOSS MAE Loss;
step 2: the trained model receives the image to be processed and outputs an effect image subjected to definition enhancement processing;
the training of the model in the step 1 comprises the following steps:
step 11: constructing a definition enhancement network;
step 12: collecting data samples;
step 13: completing model iterative training according to the collected data samples;
step 14: and finishing lightweight setting of the model.
Further, when the sharpness enhancement network is constructed in the step 11, the Width of the input image is set to be Width, and the Height is set to be Height; the construction of the sharpness enhancement network comprises the following steps:
step 111: inputting image data of RGB24, selecting the number of channels as N, namely W × H × N of network output;
step 112: performing convolution downsampling for at least two times on Input image data; each convolution reduces W and H to 1/2, and the number of convoluted channels increases in sequence;
step 113: after convolution downsampling is completed, performing ResBlock processing on an output result m times; ResBlock denotes residual block;
step 114: performing Add operation on the output result after completing the ResBlock processing and the corresponding down-sampling result in sequence; performing UpSampling and convolution operation at least twice after the Add operation is completed, and selecting a tanh activation function when the last UpSampling and convolution operation is performed to obtain a first output of the definition enhancement network;
step 115: performing Add operation on the first output and the Input, and then performing convolution processing to complete construction of the definition enhancement network; and outputting a corresponding definition enhancement effect graph after the construction of the network is completed.
Further, Width of the input image in the step 111 is 1024, Height is 1024, the number N of channels of the image data of the input RGB24 is 3, and the obtained network output is 1024 × 1024 × 3; in step 112, 4 times of convolution downsampling are carried out; performing ResBlock processing for 4 times in step 113; the upsampling and convolution operations are performed 4 times in step 114.
Further, the step 12 of acquiring data samples includes the following steps:
step 121: preparing Z scene pictures, zooming each picture, and zooming the pictures to W multiplied by H to obtain a zoom image; the zoom factor is adjusted according to the original size of the picture;
step 122: performing Gaussian blur with Radius on the zoomed picture to obtain a blur picture; wherein Radius ═ 1 or 2;
the image obtained after scaling in step 121 is called a scaling map, and the image obtained after gaussian blurring processing in step 122 is called a blur map.
Further, the model training of step 13 firstly requires inputting the fuzzy graph into the sharpness enhancement network constructed in step 11 to obtain a real result y _ true; secondly, performing iterative training on the model according to a LOSS function LOSS of the network to obtain a corrected model; the LOSS function LOSS includes three parts, which are VGG Loss, Huber Loss, and the mean absolute error LOSS MAE Loss.
Further, the VGG Loss calculation includes the following steps:
step 131: respectively inputting images of a prediction result y _ pred and a real result y _ true of the model into a VGG-19 classification network; the image of the prediction result y _ pred is a zoom image, and the image of the real result y _ true is an image output by the sharpness enhancement network according to the blur image;
step 132: respectively calculating the minimum Mean Square Error (MSE) of the feature maps corresponding to the predicted result image and the real result image for the last three convolutional layers in the VGG-19 to obtain loss _ VGG _1, loss _ VGG _2 and loss _ VGG _ 3;
step 1333: VGG Loss is the mean value of Loss _ VGG _1, Loss _ VGG _2 and Loss _ VGG _ 3.
Further, the Huber Loss is obtained by a predicted result y _ pred and a real result y _ true of the model, the value of the Huber Loss is expressed by a Loss _ Huber, and the specific calculation of the Loss _ Huber is as follows:
wherein L (y _ true, y _ pred) represents loss _ huber; δ is a set constant.
Further, the MAE Loss is also obtained through a prediction result y _ pred and a real result y _ true of the model, a value of the MAE Loss is represented by a Loss _ MAE, and the Loss _ MAE is specifically calculated as follows:
L′(y_true,y_pred)=|y_true-y_pred|
where L' (y _ true, y _ pred) represents loss _ mae.
Further, the LOSS function LOSS of the sharpness enhancement network has a value of a weighted sum of VGG LOSS, Huber LOSS, and mean absolute error LOSS MAE LOSS, expressed as:
LOSS=loss_huber+loss_mae+(loss_vgg_1+loss_vgg_2+loss_vgg_3)*0.33。
further, the image to be processed input in the step 2 is represented by S; before an image to be processed is input, preprocessing any S, wherein the preprocessing comprises cutting the width and the height of the S, the width and the height of the S after cutting are all multiples of 4, and redundant pixels are discarded; in step 14, kernel _ size of the first layer of convolution is set to 7, and kernel _ size of the remaining layers of convolution is set to 3; the maximum number of channels was set to 48; the volume of the final model was set to 0.8M.
The invention has the beneficial effects that:
the model is corrected by setting a plurality of Loss values, compared with the traditional method that the model is corrected only by the MAE Loss value, the edge characteristics of the image in the picture can be better processed by the corrected model;
by completing lightweight setting of the model, the calculated amount is limited within 100M FLOPs, and the model can be ensured to run at the mobile end.
Drawings
FIG. 1 is a block diagram of the overall structure of a first embodiment of the present invention;
fig. 2 is a schematic diagram of a sharpness enhancement network structure according to a first embodiment of the present invention.
Detailed Description
The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It is to be noted that the features in the following embodiments and examples may be combined with each other without conflict.
It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention, and the components related to the present invention are only shown in the drawings rather than drawn according to the number, shape and size of the components in actual implementation, and the type, quantity and proportion of the components in actual implementation may be changed freely, and the layout of the components may be more complicated.
The first embodiment is as follows:
as shown in fig. 1, a lightweight image sharpness enhancement method includes the following steps:
step 1: constructing a network, and performing training iteration through a set weighted LOSS to obtain a model which is trained; LOSS includes VGG Loss, Huber Loss and mean absolute error LOSS MAE Loss;
step 2: and the trained model receives the image to be processed and outputs the image subjected to the definition enhancement processing.
The LOSS in the step 1 is a LOSS function, and the training of the model in the step 1 comprises the following steps:
step 11: constructing a definition enhancement network;
step 12: collecting data samples;
step 13: completing model iterative training according to the collected data samples;
step 14: and finishing lightweight setting of the model.
As shown in fig. 2, the idea of designing the sharpness enhancement network in step 11 is to learn an edge deviation rule between a sharp target picture and a blurred picture through a large number of data samples, and implement sharpness enhancement operation on an arbitrary picture according to the deviation rule. The Width of the input image is set to Width, and the Height is set to Height, and in this example, Width is 1024 pixels. The construction of the sharpness enhancement network comprises the following steps:
step 111: inputting image data of RGB24, selecting channel number N as 3, namely W multiplied by H multiplied by N of network output is 1024 multiplied by 3;
step 112: performing convolution downsampling for at least two times on Input image data; in the example, the samples are sampled by 4 times of convolution, namely C1, C2, C3 and C4, each time of convolution, W and H are reduced to 1/2 each time, and the number of channels of convolution is increased in sequence;
step 113: after convolution downsampling is completed, performing ResBlock processing on an output result m times; m is 4 in this example; ResBlock denotes residual block;
step 114: performing Add operation on the output result after completing the ResBlock processing and the corresponding down-sampling result in sequence; performing UpSampling and convolution operation at least twice after the Add operation is completed, and selecting a tanh activation function when the last UpSampling and convolution operation is performed to obtain a first output of the definition enhancement network; in this example, a total of 4 upsampling and convolution operations are performed, each of which is denoted by U1, U2, U3 and U4, respectively;
step 115: performing Add operation on the first output and the Input, and then performing convolution processing to complete construction of the definition enhancement network; and outputting a corresponding definition enhancement effect graph after the construction of the network is completed.
The step 12 of collecting data samples comprises the following steps:
step 121: preparing Z scene pictures, 10 ten thousand in this example; zooming each picture to obtain a zoomed picture, wherein the zoomed picture is zoomed to W multiplied by H (1024 multiplied by 1024); the actual zoom factor is adjusted according to the original size of the picture;
step 122: performing Gaussian blur with Radius on the zoomed picture to obtain a blur picture; in this example Radius is 1 or 2.
The equation for gaussian blur in step 122 is as follows:
wherein x represents the current pixel value; mu represents the mean value of all pixels in a square area with the Radius of Radius and the x pixel position as the center position; sigma represents the variance of all pixels in a square area with the x pixel position as the center position and Radius as Radius; GaussFilter (x) indicates that the pixel is the blurred result.
Wherein, the 10 ten thousand zoom images after being zoomed in the step 121 and the 10 ten thousand blur images after being processed by the gaussian blur processing in the step 122 are used as a data sample pair for model training, wherein the zoom images are used as output, and the blur images are used as input.
The model training in step 13 firstly requires inputting the fuzzy graph in the data sample pair into the sharpness enhancement network constructed in step 11, and obtains a true result y _ true. And then, carrying out iterative training on the model according to the LOSS function LOSS of the network to obtain the corrected model. The LOSS function LOSS includes three parts, which are VGG Loss, Huber Loss, and mean absolute error LOSS (MAE Loss), respectively.
The VGG Loss calculation comprises the following steps:
step 131: respectively inputting images of a prediction result y _ pred and a real result y _ true of the model into a VGG-19 classification network; the image of the prediction result y _ pred is a zoom image, and the image of the real result y _ true is an image output by the sharpness enhancement network according to the blur image;
step 132: respectively calculating the minimum Mean Square Error (MSE) of the feature maps corresponding to the predicted result image and the real result image for the last three convolutional layers in the VGG-19 to obtain loss _ VGG _1, loss _ VGG _2 and loss _ VGG _ 3;
step 1333: VGG Loss is the mean value of Loss _ VGG _1, Loss _ VGG _2 and Loss _ VGG _ 3.
The Huber Loss is obtained through a prediction result y _ pred and a real result y _ true of the model, the value of the Huber Loss is represented by a Loss _ hub, and the specific calculation of the Loss _ hub is as follows:
wherein L (y _ true, y _ pred) represents loss _ huber; δ is a set constant, in this example 1.0.
The MAE Loss is also obtained through a prediction result y _ pred and a real result y _ true of the model, the value of the MAE Loss is represented by a Loss _ MAE, and the specific calculation of the Loss _ MAE is as follows:
L′(y_true,y_pred)=|y_true-y_pred|
where L' (y _ true, y _ pred) represents loss _ mae.
The LOSS function LOSS of the final sharpness enhancement network has a value of the weighted sum of VGG LOSS, Huber LOSS, and mean absolute error LOSS (MAE LOSS), expressed as:
LOSS=loss_huber+loss_mae+(loss_vgg_1+loss_vgg_2+loss_vgg_3)*0.33
and training the model according to the obtained LOSS function to obtain a training model.
Step 14, in the setting of model lightweight, setting kernel _ size of the convolution of the first layer to be 7, and setting kernel _ size of the convolution of the other layers to be 3; the maximum number of channels was set to 48; the volume of the final model was set to 0.8M. The calculation amount is limited within 100M FLOPs through the lightweight setting of the model, so that the model can be adapted to the mobile terminal. When the GPU processes the 720P resolution image, the processing speed is within 200 ms; by applying various weighted Loss values including values of VGG Loss, Huber Loss and MAE Loss and replacing the value of MSE Loss, the detailed characteristic effect of the image edge is richer and more real.
The image to be processed input in the step 2 is represented by S, and because the network is designed as a convolutional network and can be expanded to any size input, any S is preprocessed, where the preprocessing includes clipping S, so that the width and height of S are all multiples of 4, and the redundant pixels are discarded, it needs to be noted that only the minimum pixel value meeting the condition is clipped during clipping, for example, the width of the original image is 9 pixels, and the clipped part is 1 pixel instead of 5 pixels; and secondly, inputting the S into the model which completes lightweight setting to obtain a corresponding output effect image with enhanced definition.
In the implementation process, the model is corrected by setting a plurality of Loss values, compared with the traditional method that the model is corrected only by the MAE Loss value, the method can ensure that the corrected model can better process the edge characteristics of the image in the picture; by completing lightweight setting of the model, the calculated amount is limited within 100M FLOPs, and the model can be ensured to run at the mobile end.
The above description is only one specific example of the present invention and should not be construed as limiting the invention in any way. It will be apparent to persons skilled in the relevant art(s) that, having the benefit of this disclosure and its principles, various modifications and changes in form and detail can be made without departing from the principles and structures of the invention, which are, however, encompassed by the appended claims.
Claims (10)
1. A lightweight image definition enhancement method is characterized by comprising the following steps:
step 1: constructing a network, and performing training iteration through a set weighted LOSS to obtain a model which is trained; LOSS includes VGG Loss, Huber Loss and mean absolute error LOSS MAE Loss;
step 2: the trained model receives the image to be processed and outputs an effect image subjected to definition enhancement processing;
the training of the model in the step 1 comprises the following steps:
step 11: constructing a definition enhancement network;
step 12: collecting data samples;
step 13: completing model iterative training according to the collected data samples;
step 14: and finishing lightweight setting of the model.
2. A lightweight image sharpness enhancement method according to claim 1, wherein when constructing the sharpness enhancement network in step 11, the Width of the input image is set to Width and the Height is set to Height; the construction of the sharpness enhancement network comprises the following steps:
step 111: inputting image data of RGB24, selecting the number of channels as N, namely W × H × N of network output;
step 112: performing convolution downsampling for at least two times on Input image data; each convolution reduces W and H to 1/2, and the number of convoluted channels increases in sequence;
step 113: after convolution downsampling is completed, performing ResBlock processing on an output result m times; ResBlock denotes residual block;
step 114: performing Add operation on the output result after completing the ResBlock processing and the corresponding down-sampling result in sequence; performing UpSampling and convolution operation at least twice after the Add operation is completed, and selecting a tanh activation function when the last UpSampling and convolution operation is performed to obtain a first output of the definition enhancement network;
step 115: performing Add operation on the first output and the Input, and then performing convolution processing to complete construction of the definition enhancement network; and outputting a corresponding definition enhancement effect graph after the construction of the network is completed.
3. A lightweight image sharpness enhancement method according to claim 2, wherein in said step 111, Width of the input image is 1024, Height is 1024, the number N of channels of the image data of the input RGB24 is 3, and the net output is 1024 × 1024 × 3; in step 112, 4 times of convolution downsampling are carried out; performing ResBlock processing for 4 times in step 113; the upsampling and convolution operations are performed 4 times in step 114.
4. A lightweight method for enhancing image sharpness as claimed in claim 2, wherein the step 12 of acquiring data samples comprises the steps of:
step 121: preparing Z scene pictures, zooming each picture, and zooming the pictures to W multiplied by H to obtain a zoom image; the zoom factor is adjusted according to the original size of the picture;
step 122: performing Gaussian blur with Radius on the zoomed picture to obtain a blur picture; wherein Radius is 1 or 2.
5. A lightweight image sharpness enhancement method according to claim 1, wherein the model training of step 13 first requires inputting a blur image into the sharpness enhancement network constructed in step 11 to obtain a true result y _ true; secondly, performing iterative training on the model according to a LOSS function LOSS of the network to obtain a corrected model; the LOSS function LOSS includes three parts, which are VGG Loss, Huber Loss, and the mean absolute error LOSS MAE Loss.
6. A method as claimed in claim 5, wherein the VGG Loss calculation includes the following steps:
step 131: respectively inputting images of a prediction result y _ pred and a real result y _ true of the model into a VGG-19 classification network; the image of the prediction result y _ pred is a zoom image, and the image of the real result y _ true is an image output by the sharpness enhancement network according to the blur image;
step 132: respectively calculating the minimum Mean Square Error (MSE) of the feature maps corresponding to the predicted result image and the real result image for the last three convolutional layers in the VGG-19 to obtain loss _ VGG _1, loss _ VGG _2 and loss _ VGG _ 3; step 1333: VGG Loss is the mean value of Loss _ VGG _1, Loss _ VGG _2 and Loss _ VGG _ 3.
7. The method according to claim 6, wherein the Huber Loss is obtained from a predicted result y _ pred and a true result y _ true of the model, the value of the Huber Loss is expressed by a Loss _ Huber, and the Loss _ Huber is specifically calculated as follows:
wherein L (y _ true, y _ pred) represents loss _ huber; δ is a set constant.
8. The method according to claim 7, wherein the MAE Loss is also obtained from the predicted result y _ pred and the true result y _ true of the model, and the value of MAE Loss is expressed as Loss _ MAE, which is specifically calculated as follows:
L′(y_true,y_pred)=|y_true-y_pred|
where L' (y _ true, y _ pred) represents loss _ mae.
9. A lightweight image sharpness enhancement method according to claim 8, wherein the LOSS function LOSS of the sharpness enhancement network has a value of a weighted sum of VGG LOSS, Huber LOSS and mean absolute error LOSS, MAE LOSS, expressed as:
LOSS=loss_huber+loss_mae+(loss_vgg_1+loss_vgg_2+loss_vgg_3)*0.33。
10. a lightweight image sharpness enhancement method according to claim 2, wherein the image to be processed input in step 2 is denoted by S; before an image to be processed is input, preprocessing any S, wherein the preprocessing comprises cutting the width and the height of the S, the width and the height of the S after cutting are all multiples of 4, and redundant pixels are discarded; in step 14, kernel _ size of the first layer of convolution is set to 7, and kernel _ size of the remaining layers of convolution is set to 3; the maximum number of channels was set to 48; the volume of the final model was set to 0.8M.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011373558.6A CN112435192B (en) | 2020-11-30 | 2020-11-30 | Lightweight image definition enhancing method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011373558.6A CN112435192B (en) | 2020-11-30 | 2020-11-30 | Lightweight image definition enhancing method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112435192A true CN112435192A (en) | 2021-03-02 |
CN112435192B CN112435192B (en) | 2023-03-14 |
Family
ID=74697583
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011373558.6A Active CN112435192B (en) | 2020-11-30 | 2020-11-30 | Lightweight image definition enhancing method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112435192B (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108376387A (en) * | 2018-01-04 | 2018-08-07 | 复旦大学 | Image deblurring method based on polymerization expansion convolutional network |
CN108549892A (en) * | 2018-06-12 | 2018-09-18 | 东南大学 | A kind of license plate image clarification method based on convolutional neural networks |
CN108898579A (en) * | 2018-05-30 | 2018-11-27 | 腾讯科技(深圳)有限公司 | A kind of image definition recognition methods, device and storage medium |
US10646156B1 (en) * | 2019-06-14 | 2020-05-12 | Cycle Clarity, LLC | Adaptive image processing in assisted reproductive imaging modalities |
CN111192206A (en) * | 2019-12-03 | 2020-05-22 | 河海大学 | Method for improving image definition |
-
2020
- 2020-11-30 CN CN202011373558.6A patent/CN112435192B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108376387A (en) * | 2018-01-04 | 2018-08-07 | 复旦大学 | Image deblurring method based on polymerization expansion convolutional network |
CN108898579A (en) * | 2018-05-30 | 2018-11-27 | 腾讯科技(深圳)有限公司 | A kind of image definition recognition methods, device and storage medium |
CN108549892A (en) * | 2018-06-12 | 2018-09-18 | 东南大学 | A kind of license plate image clarification method based on convolutional neural networks |
US10646156B1 (en) * | 2019-06-14 | 2020-05-12 | Cycle Clarity, LLC | Adaptive image processing in assisted reproductive imaging modalities |
CN111192206A (en) * | 2019-12-03 | 2020-05-22 | 河海大学 | Method for improving image definition |
Non-Patent Citations (1)
Title |
---|
丁予康: "基于卷积神经网络的图像超分辨方法研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》, no. 1, 15 January 2019 (2019-01-15), pages 138 - 2990 * |
Also Published As
Publication number | Publication date |
---|---|
CN112435192B (en) | 2023-03-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | Real-esrgan: Training real-world blind super-resolution with pure synthetic data | |
CN112233038B (en) | True image denoising method based on multi-scale fusion and edge enhancement | |
CN109389556B (en) | Multi-scale cavity convolutional neural network super-resolution reconstruction method and device | |
CN110008817B (en) | Model training method, image processing method, device, electronic equipment and computer readable storage medium | |
Yu et al. | A unified learning framework for single image super-resolution | |
CN110136062B (en) | Super-resolution reconstruction method combining semantic segmentation | |
CN112446383B (en) | License plate recognition method and device, storage medium and terminal | |
CN110766632A (en) | Image denoising method based on channel attention mechanism and characteristic pyramid | |
CN112102177A (en) | Image deblurring method based on compression and excitation mechanism neural network | |
CN112419191B (en) | Image motion blur removing method based on convolution neural network | |
CN110942436B (en) | Image deblurring method based on image quality evaluation | |
EP4075373A1 (en) | Image processing method and apparatus | |
CN112529776A (en) | Training method of image processing model, image processing method and device | |
CN113658044A (en) | Method, system, device and storage medium for improving image resolution | |
CN111951164A (en) | Image super-resolution reconstruction network structure and image reconstruction effect analysis method | |
CN112598587A (en) | Image processing system and method combining face mask removal and super-resolution | |
CN116547694A (en) | Method and system for deblurring blurred images | |
CN113096032B (en) | Non-uniform blurring removal method based on image region division | |
CN113724134A (en) | Aerial image blind super-resolution reconstruction method based on residual distillation network | |
CN116029943B (en) | Infrared image super-resolution enhancement method based on deep learning | |
CN112435192B (en) | Lightweight image definition enhancing method | |
CN110717913B (en) | Image segmentation method and device | |
CN109996085B (en) | Model training method, image processing method and device and electronic equipment | |
CN114187174A (en) | Image super-resolution reconstruction method based on multi-scale residual error feature fusion | |
CN113240589A (en) | Image defogging method and system based on multi-scale feature 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 | ||
CP02 | Change in the address of a patent holder | ||
CP02 | Change in the address of a patent holder |
Address after: 22nd floor, block a, Huaxing Times Square, 478 Wensan Road, Xihu District, Hangzhou, Zhejiang 310000 Patentee after: Hangzhou Xiaoying Innovation Technology Co.,Ltd. Address before: 16 / F, HANGGANG Metallurgical Science and technology building, 294 Tianmushan Road, Xihu District, Hangzhou City, Zhejiang Province, 310012 Patentee before: Hangzhou Xiaoying Innovation Technology Co.,Ltd. |