CN114170106A - High-reflection noise removing method based on improved U-Net model - Google Patents
High-reflection noise removing method based on improved U-Net model Download PDFInfo
- Publication number
- CN114170106A CN114170106A CN202111510031.8A CN202111510031A CN114170106A CN 114170106 A CN114170106 A CN 114170106A CN 202111510031 A CN202111510031 A CN 202111510031A CN 114170106 A CN114170106 A CN 114170106A
- Authority
- CN
- China
- Prior art keywords
- improved
- net model
- image
- point cloud
- cloud data
- 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 50
- 238000010606 normalization Methods 0.000 claims abstract description 19
- 210000002569 neuron Anatomy 0.000 claims abstract description 14
- 238000011176 pooling Methods 0.000 claims abstract description 12
- 238000011423 initialization method Methods 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims abstract description 10
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 230000011218 segmentation Effects 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 17
- 230000004913 activation Effects 0.000 claims description 11
- 238000009499 grossing Methods 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 abstract description 14
- 238000010586 diagram Methods 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 7
- 239000000463 material Substances 0.000 description 4
- 230000006835 compression Effects 0.000 description 3
- 238000007906 compression Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 239000002699 waste material Substances 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 229910052751 metal Inorganic materials 0.000 description 2
- 239000002184 metal Substances 0.000 description 2
- 150000002739 metals Chemical class 0.000 description 2
- 238000002310 reflectometry Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000002146 bilateral effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000007619 statistical method 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
- G06T5/70—Denoising; Smoothing
-
- 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a high-reflection noise removing method based on an improved U-Net model, which can improve the running speed and the detection accuracy of a high-reflection region in an image, wherein the improved U-Net model comprises three convolution modules of an up-sampling path and a down-sampling path, and each convolution module comprises two convolution layers connected with a linear rectification function and is linked by a maximum pooling layer; the method comprises the following steps: acquiring an original image of an object to be detected, preprocessing the original image, training an improved U-Net model to construct the improved U-Net model: adding batch normalization layers, adding one normalization layer after each convolution layer in the convolution module, abandoning neuron operation, initializing the weight by using an Xavier initialization method, outputting a segmentation image, taking original point cloud data, registering the original point cloud data to obtain registered point cloud data, and registering the point cloud data to reduce noise, smooth, simplify and reconstruct three-dimensionally.
Description
Technical Field
The invention relates to the technical field of computer vision, in particular to a high-precision visual detection and removal method for high-reflection noise based on a U-Net improved model.
Background
When the surface quality of objects such as industrial parts, metals and the like is detected, the computer vision technology is one of the most commonly used technical means at present, but the surfaces of the objects such as industrial parts, metals, mirrors and the like are easy to have high light reflection areas, the surface images of the objects such as the industrial parts and the like with high light reflection characteristics are collected through a camera in the computer vision technology, and when the collected images are processed by adopting the image processing technology, the high light reflection areas in the images can seriously affect the processing effect of the images, so that the image processing difficulty is increased, and the detection result is easy to cause inaccuracy.
Detecting and separating high reflectance regions in an image is not simple, and conventional separation methods are mainly performed on single images and multiple images. The common method and strategy for removing the highlight area of the multi-image is as follows: finding matching specular and diffuse reflectance pixel features from multiple images typically takes a significant amount of time. While the method for removing highlights of a single image is implemented based on a model, for example, the name: the article of color identification of color structured light based on a two-color reflection model (published in computer science and newspapers, 9.2013, vol.36, 9) provides a projection light color identification algorithm of a two-color reflection model, which can realize the detection of the color structured light. However, the performance of the model is very susceptible to the geometric shape, lighting condition, reflectivity, material characteristics, and other factors of the object, and when the difference between these factors is large, the detection results of the model may be completely different, so that the purpose of effectively detecting and removing the high reflection noise cannot be achieved.
There have also been some recent studies to achieve high optical noise removal in images by models, such as the application of a model generating a countermeasure network (GAN) in a single endoscopic image; for another example, the effect of removing specular highlight noise from a single image is achieved by generating a diffuse reflection component by a Convolutional Neural Network (CNN) model, and when image classification is performed by using a convolutional neural network, the output result is a class label of the entire image. A typical representation of a Convolutional Neural Network (CNN) model is a U-Net network, the U-Net network has a U-shaped structure, and by using the structure, not only can used training pictures be reduced, the training speed is improved, but also accurate segmentation of images can be realized. However, the U-Net network is pixel-level, which outputs a category of each pixel point, and different categories of pixels display different colors, although the U-Net network is proposed to solve the related problems of biomedical images, it is also favored by other computer vision tasks due to its powerful performance and good generalization capability. However, U-Net networks also have two significant drawbacks: (1) the operation convergence speed is slow, the over-fitting problem exists, and the original network characteristics can cause the same characteristics to be trained for multiple times, so that the resource waste is caused. (2) Positioning accuracy and obtaining context information are not compatible.
Compared with the traditional method, the method based on deep learning also depends on training data, when the training data is insufficient, the output result is poor, the processed image has the problems of color distortion, gray value residual error and the like, and the accuracy standard required by the industry is difficult to achieve. Therefore, the invention of a high-precision visual detection method which is less affected by the geometric shape, illumination condition, reflectivity, material characteristics, training data and other factors of an object and can improve the detection accuracy and the removal effect of high-reflection noise of an image becomes a problem to be solved by the technical field.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a high-precision visual detection and removal method for high-reflection noise, which can improve the running speed and the detection accuracy of a high-reflection area in an image.
In order to achieve the purpose, the invention adopts the following technical scheme:
a high-reflection noise removing method based on an improved U-Net model is characterized in that the method is realized based on the improved U-Net model, the improved U-Net model comprises three convolution modules of an up-sampling path and a down-sampling path, each convolution module comprises two convolution layers connected with a linear rectification function (ReLU activation function) and is linked by a maximum pooling layer; the method comprises the following steps: s1, acquiring an original image of the object to be detected;
s2, preprocessing the original image to obtain a preprocessed image;
s3, taking the preprocessed image as an input image, training the improved U-Net model, and constructing the improved U-Net model; the training step comprises: s31, adding batch normalization layers in the improved U-Net model, and adding one normalization layer after each convolution layer in the convolution module;
s32, discarding neuron operation;
s33, initializing the weight by using an Xavier initialization method, and outputting an improved image;
s4, acquiring original point cloud data;
s5, registering the original point cloud data to obtain registered point cloud data;
s6, denoising, smoothing and simplifying the registration point cloud data to obtain processed point cloud data;
and S7, performing three-dimensional reconstruction based on the processed point cloud data.
It is further characterized in that the method further comprises the steps of,
the convolution layer is a 3 x 3 convolution layer;
the maximum pooling layer is a 2 x 2 maximum pooling layer;
in step S1, a binocular camera is used to collect an original image, which is a two-dimensional image;
in step S2, the preprocessing step includes: adjusting the size of the original image to the input size required by the improved U-Net model; processing the original image after size adjustment by applying a mean subtraction, wherein the mean subtraction refers to: subtracting the average value of each corresponding feature from the feature data of the original image after the size adjustment;
in step S31, inserting one batch normalization layer after each 3 × 3 convolutional layer specifically includes: in the training process, S311, carrying out batch standardization on the activation functions in the model;
s312, calculating the mean and variance of each individual feature of each input image;
s313, obtaining the mean and variance values of each individual feature and dividing the difference values by the standard deviation;
s314, normalizing the activation function after batch normalization and adding a constant;
in step S32, in the training process of the improved U-Net model, a discarding neuron operation is added, which specifically includes: discarding a single node with probability p in each training stage, and reducing the scale of the model;
the probability p is 0.2;
in step S33, in the training process of the improved U-Net model, the Xavier initialization method refers to extracting samples from the uniform distribution by comprehensively using an Adadelta optimizer and default parameters, and using the classification cross entropy as a loss function.
By adopting the structure of the invention, the following beneficial effects can be achieved: according to the method, high-reflection noise in the image is removed based on the improved U-Net model, batch normalization layers are added in the training process of the improved U-Net model, namely, a normalization layer is added behind each convolution layer in the convolution module, and the normalization layers are used for normalizing the activation function in the model, so that the convergence rate in the training process can be increased, and the problem of overfitting of the model is avoided by applying indirect regularization terms.
The discarding neuron operation is added in the improved U-Net model training process, neurons depend on each other in the training process, the individual ability of each neuron is limited, the training data overfitting is easily caused, the adding of the discarding neuron operation can provide implicit data to relieve the overfitting, in the operation, a single node is discarded in each training stage according to the probability p, the probability p is the implicit data, the model scale is indirectly reduced, the operation convergence speed is accelerated, the output volume size is not changed, the interdependency of the neurons in the training process is reduced, and the individual ability of each neuron is prevented from being limited and the overfitting of the training data is avoided. And abandoning neuron operation avoids the characteristic of the original model from being trained for many times, thereby reducing resource waste.
The improved U-Net model training process also uses an Xavier initialization method, the Xavier initialization method is beneficial to accelerating convergence and reducing overfitting, and inaccurate image stripe identification caused by high light reflection is reduced, so that the point cloud reconstruction accuracy of a high light reflection region is improved.
Drawings
FIG. 1 is a flow chart of a method for removing high reflection noise according to the present invention;
FIG. 2 is a model structure diagram of an original U-Net model;
FIG. 3 is a schematic diagram of a three-dimensional image generated by object reconstruction acquired by a binocular camera in a natural environment;
FIG. 4 is a model structure diagram of the improved U-Net model of the present invention;
FIG. 5 is a schematic structural diagram of a neural network layer added with a batch normalization layer in the improved U-Net model according to the present invention;
FIG. 6 is a schematic representation of a three-dimensional image generated by object reconstruction after the improved U-Net model of the present invention is used, wherein mirror portions are highlighted in a cross-hatch manner;
FIG. 7 is a schematic diagram of the three-dimensional image reduction generated by object reconstruction after the U-Net model is improved by the method.
Detailed Description
The following disclosure provides many different embodiments or examples for implementing different features of the invention. To simplify the disclosure of the present invention, the components and arrangements of specific examples are described below. Of course, they are merely examples and are not intended to limit the present invention. Furthermore, the present invention may repeat reference numerals and/or letters in the various examples, such repetition is for the purpose of simplicity and clarity and does not in itself dictate a relationship between the various embodiments and/or configurations discussed. In addition, the present invention provides examples of various specific processes and materials, but one of ordinary skill in the art may recognize applications of other processes and/or uses of other materials.
The method takes a mirror as a detection target, the mirror surface of the mirror is a high-reflection area, a binocular camera is adopted to collect mirror images, the mirror images are detected and restored based on a model, the conventional model for mirror detection is a U-Net model in a Convolutional Neural Network (CNN) model, the U-Net model is a common algorithm, a full convolutional network is used for semantic segmentation, the algorithm is shown in figure 2, and a symmetrical U-shaped structure containing a compression Path (compressing Path) and an expansion Path is used in the original U-Net model. The input to the U-Net model is a frame of image 572pixel by 572pixel edge subjected to a mirroring operation (Overlay-tile template) that facilitates better processing of the image's boundary pixels. On the left side of the original U-Net model is a series of down-sampling operations consisting of convolutional layers and max-pooling layers, this part being called the compression path. The compression path is composed of four convolution modules (blocks), each module uses three convolution layer operations and one maximum pooling layer for down-sampling, and the Feature Map with the size of 32 pixels × 32 pixels is obtained by multiplying the number of Feature maps (Feature maps) after each down-sampling by 2, which explains the principle of Feature Map size change in the Map. However, the operation convergence speed of the original U-Net model is slow, the over-fitting problem exists, the same characteristics can be trained for multiple times due to the original network characteristics, resource waste is caused, and in addition, the original U-Net model also has the problems that the positioning accuracy and the context information acquisition cannot be achieved at the same time.
In order to solve the problems, the application provides a high-reflection region detection method which is based on an improved model of a U-Net model and used for segmenting and removing a high-reflection noise part in an image, so that high-precision detection of high-reflection noise of a workpiece is achieved. The improved U-Net model comprises the following steps: (1) three convolution modules of the up-sampling path (up-conv); (2) each convolution module contains two 3 x 3 convolution layers connected to the ReLU activation function and is linked by a 2 x 2 Max Pooling layer (Max Pooling); (3) the modified U-Net model inserts a Batch Normalization Layer (BN) after each convolutional Layer; (4) adding a discarding (Dropout) step in the training stage of the improved model, and discarding a single node with a probability p; (5) unlike the original parameter initialization method according to gaussian distribution, the improved model uses Xavier unified initializer and uses class cross entropy as loss function.
Fig. 4 shows a model structure diagram of an improved U-Net model that repeatedly applies two 3 x 3 unfilled convolutional layers, each followed by an activation function ReLU and a 2 x 2 max pooling (max-pool 2 x 2) operation, setting the downsampling step size to 2. The signature is upsampled in each step of the decoder path and a 2 x 2 convolutional layer is applied that halves the number of channels. After this, the systolic path and the corresponding profile will be connected. Two further successive 3 x 3 convolutional layers are then applied, each followed by a ReLU activation function. At the last level, each feature vector is mapped to the required number of classes using the 1 × 1 convolutional layer. In macroscopic view, the model can be interpreted as having four levels in each path, with a bridge connection in between, which is a 256pixel by 256pixel image obtained after 3 successive downsamplings. In fig. 4, for example, the side length N of a convolution layer (COVN) image is 32, 64, 128, 256 or the image pixels are: 32pixel × 32pixel, 64pixel × 64pixel, 128pixel × 128pixel, 256pixel × 256pixel, "copy and concat" indicates that an image is copied and cropped.
Referring to fig. 1, the following is a high-light-reflection region detection method based on an improved U-Net model, and the method specifically includes the steps: s1, acquiring an image by using a binocular camera, and acquiring a two-dimensional original image of an object to be detected, wherein the object to be detected is a mirror in the embodiment;
s2, preprocessing the original image to obtain a preprocessed image, wherein the preprocessing step comprises: s21, adjusting the size of the original image to the input size required by the improved U-Net model; s22, processing the original image after size adjustment by applying a mean subtraction, wherein the mean subtraction refers to: subtracting the average value of each corresponding feature from the feature data of the original image after size adjustment;
s3, taking the preprocessed image as an input image, training the improved U-Net model, and constructing the improved U-Net model; the training step comprises: s31, adding a normalization layer after adding each convolution layer in a batch normalization layer (BN) module in the improved U-Net model; the method comprises the following specific operation steps: after each 3 × 3 convolutional layer, a batch normalization layer is inserted, see fig. 5, which specifically includes: in the training process, S311, carrying out batch standardization on the activation functions in the model;
s312, calculating the mean and variance of each individual feature of each input image;
s313, obtaining the mean and variance values of each individual feature and dividing the difference values by the standard deviation;
and S314, normalizing the activation function after batch normalization and adding a constant.
And S32, discarding neuron operation, wherein in the training process of the improved U-Net model, the discarding neuron operation is added, namely, a single node is discarded by a probability p in each training stage, and the model size is reduced, wherein the probability p is 0.2.
S33, in the training process of the improved U-Net model, initializing the weight by using an Xavier initialization method, and outputting an improved image; the Xavier initialization method is realized based on a Xavier unified initializer, and the Xavier initialization method is realized by comprehensively using an Adadelta optimizer and default parameters, extracting samples from uniform distribution and using classification cross entropy as a loss function.
S4, acquiring the original point cloud data of the output improved image;
s5, adopting an Iterative Closest Point algorithm (ICP) to register the original Point cloud data to obtain registered Point cloud data;
s6, performing noise reduction, smoothing and simplification processing on the registration point cloud data by adopting a bilateral filtering method to obtain processed point cloud data, wherein noise reduction refers to denoising outliers which are obviously deviated in the registration point cloud data, and in the denoising process, the neighborhood of each point is subjected to statistical analysis, and neighborhood points which do not meet the standard are removed, so that the processed point cloud data is obtained;
and S7, performing three-dimensional reconstruction based on the processed point cloud data to obtain a three-dimensional reconstructed image.
Fig. 6 shows a schematic diagram of a three-dimensional image generated by object reconstruction after the improved U-Net model is adopted, wherein mirror portions are highlighted in a crossed manner, and it can be seen from the diagram that high reflection noise in the image is removed and accurately identified. As can be seen from FIG. 7, after the three-dimensional gray scale image obtained by improving the U-Net model is restored to be an RGB image, the image is basically consistent with a real object, and the method for detecting the high-reflection area based on the improved U-Net model can further realize the segmentation and accurate identification of the high-reflection area of the mirror.
The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is to be understood that other modifications and variations directly derivable or suggested by those skilled in the art without departing from the spirit and concept of the present invention are to be considered as included within the scope of the present invention.
Claims (8)
1. A high-reflection noise removing method based on an improved U-Net model is characterized in that the method is realized based on the improved U-Net model, the improved U-Net model comprises three convolution modules of an up-sampling path and a down-sampling path, each convolution module comprises two convolution layers connected with a linear rectification function, and the convolution layers are linked by a maximum pooling layer; the method comprises the following steps: s1, acquiring an original image of the object to be detected;
s2, preprocessing the original image to obtain a preprocessed image;
s3, taking the preprocessed image as an input image, training the improved U-Net model, and constructing the improved U-Net model; the training step comprises: s31, adding batch normalization layers in the improved U-Net model, and adding one normalization layer after each convolution layer in the convolution module;
s32, discarding neuron operation;
s33, initializing the weight by using an Xavier initialization method, and outputting a segmentation image;
s4, acquiring original point cloud data;
s5, registering the original point cloud data to obtain registered point cloud data;
s6, denoising, smoothing and simplifying the registration point cloud data to obtain processed point cloud data;
and S7, performing three-dimensional reconstruction based on the processed point cloud data.
2. The improved U-Net model-based highly reflective noise removal method according to claim 1, wherein the convolutional layer is a 3 x 3 convolutional layer; the maximum pooling layer is a 2 × 2 maximum pooling layer.
3. The improved U-Net model-based highly reflective noise removal method according to claim 1 or 2, wherein in step S1, a binocular camera is used to capture an original image, and the original image is a two-dimensional image.
4. The method for removing the high reflection noise based on the improved U-Net model according to claim 3, wherein in the step S2, the preprocessing step comprises: adjusting the size of the original image to the input size required by the improved U-Net model; processing the original image after size adjustment by applying a mean subtraction, wherein the mean subtraction refers to: the average value of each corresponding feature is subtracted from the feature data of the resized original image.
5. The improved U-Net model based highly reflective noise removing method according to claim 4, wherein in step S31, inserting one batch normalization layer after each 3 x 3 convolutional layer specifically comprises: in the training process, S311, carrying out batch standardization on the activation functions in the model;
s312, calculating the mean and variance of each individual feature of each input image;
s313, obtaining the mean and variance values of each individual feature and dividing the difference values by the standard deviation;
s314, normalizing the activation function after batch normalization and adding a constant.
6. The method for removing the high reflection noise based on the improved U-Net model of claim 5, wherein in step S32, in the training process of the improved U-Net model, a discarding neuron operation is added, specifically comprising: and (4) discarding a single node with the probability p in each training stage, and reducing the model size.
7. The improved U-Net model-based hyperreflective noise removal method according to claim 6, wherein the probability p is 0.2.
8. The improved U-Net model based highly reflective noise removing method according to claim 7, wherein in step S33, in the training process of the improved U-Net model, the Xavier initialization method refers to using Adadelta optimizer in combination with default parameters, extracting samples from uniform distribution, and using classification cross entropy as a loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111510031.8A CN114170106A (en) | 2021-12-10 | 2021-12-10 | High-reflection noise removing method based on improved U-Net model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111510031.8A CN114170106A (en) | 2021-12-10 | 2021-12-10 | High-reflection noise removing method based on improved U-Net model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114170106A true CN114170106A (en) | 2022-03-11 |
Family
ID=80485508
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111510031.8A Pending CN114170106A (en) | 2021-12-10 | 2021-12-10 | High-reflection noise removing method based on improved U-Net model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114170106A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114719780A (en) * | 2022-04-07 | 2022-07-08 | 昆明理工大学 | Three-dimensional measurement method for structured light projection of surface of reflective object based on deep learning |
| CN114782387A (en) * | 2022-04-29 | 2022-07-22 | 苏州威达智电子科技有限公司 | Surface defect detection system |
-
2021
- 2021-12-10 CN CN202111510031.8A patent/CN114170106A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114719780A (en) * | 2022-04-07 | 2022-07-08 | 昆明理工大学 | Three-dimensional measurement method for structured light projection of surface of reflective object based on deep learning |
| CN114719780B (en) * | 2022-04-07 | 2022-12-16 | 昆明理工大学 | Three-dimensional measurement method for surface structured light projection of reflective object based on deep learning |
| CN114782387A (en) * | 2022-04-29 | 2022-07-22 | 苏州威达智电子科技有限公司 | Surface defect detection system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109580630B (en) | Visual inspection method for defects of mechanical parts | |
| CN110675368B (en) | Cell image semantic segmentation method integrating image segmentation and classification | |
| CN110458831B (en) | Scoliosis image processing method based on deep learning | |
| CN111814867A (en) | Defect detection model training method, defect detection method and related device | |
| CN112819772B (en) | High-precision rapid pattern detection and recognition method | |
| Li et al. | Robust retinal image enhancement via dual-tree complex wavelet transform and morphology-based method | |
| CN113240626B (en) | Glass cover plate concave-convex type flaw detection and classification method based on neural network | |
| CN109035172B (en) | Non-local mean ultrasonic image denoising method based on deep learning | |
| CN114170106A (en) | High-reflection noise removing method based on improved U-Net model | |
| CN111462120A (en) | Defect detection method, device, medium and equipment based on semantic segmentation model | |
| CN114550021B (en) | Surface defect detection method and device based on feature fusion | |
| CN112465759A (en) | Convolutional neural network-based aeroengine blade defect detection method | |
| CN113160226B (en) | Classification segmentation method and system for AMD lesion OCT image based on bidirectional guidance network | |
| GB2466818A (en) | Cell image segmentation using binary threshold and greyscale image processing | |
| CN109166092A (en) | A kind of image defect detection method and system | |
| CN116012291A (en) | Industrial part image defect detection method and system, electronic equipment and storage medium | |
| CN114841992A (en) | Defect detection method based on cyclic generation countermeasure network and structural similarity | |
| CN116934762B (en) | System and method for detecting surface defects of lithium battery pole piece | |
| Cho et al. | Modified perceptual cycle generative adversarial network-based image enhancement for improving accuracy of low light image segmentation | |
| CN117217997A (en) | Remote sensing image super-resolution method based on context perception edge enhancement | |
| CN115330874B (en) | Monocular depth estimation method based on superpixel processing shielding | |
| CN114627136A (en) | Tongue picture segmentation and alignment method based on feature pyramid network | |
| CN113822869A (en) | Transparent soft packaging bag static detection method and system based on machine vision | |
| Jimenez et al. | Deep convolutional autoencoders for deblurring and denoising low-resolution images | |
| Khong et al. | A novel approach for efficient implementation of nucleus detection and segmentation using correlated dual color space |
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 |