CN109829903B - Chip surface defect detection method based on convolution denoising autoencoder - Google Patents
Chip surface defect detection method based on convolution denoising autoencoder Download PDFInfo
- Publication number
- CN109829903B CN109829903B CN201910078674.6A CN201910078674A CN109829903B CN 109829903 B CN109829903 B CN 109829903B CN 201910078674 A CN201910078674 A CN 201910078674A CN 109829903 B CN109829903 B CN 109829903B
- Authority
- CN
- China
- Prior art keywords
- image
- img
- chip surface
- residual
- defect
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
- Testing Or Measuring Of Semiconductors Or The Like (AREA)
Abstract
The invention discloses a chip surface defect detection method based on a convolution denoising autoencoder, which comprises the following steps: 1, constructing a defect-free image reconstruction based on a convolution denoising self-encoder; 2, constructing a residual error map based on the overlapped area; and 3, detecting the defects based on the residual error map. The method can effectively detect weak defects with the characteristics of low contrast between the defects and the background, small defects and the like, has strong robustness, and improves the detection precision of the chip surface.
Description
Technical Field
The invention belongs to the technical field of image surface detection, and relates to a chip surface defect detection method based on a convolution denoising autoencoder.
Background
Chips are a general term for semiconductor device products, have been used in almost all fields such as computing, communications, manufacturing, and transportation, and have become an indispensable part of modern society. During the production process, the chip surface may have scratches, melting, cracks, etc. which affect the appearance and performance of the chip and thus need to be detected and processed. The detection method based on machine vision has the advantages of low cost, high efficiency and the like, and is widely applied to a plurality of links of chip production, such as identification detection, pin detection, wafer detection, packaging detection and the like. The chip surface defect detection belongs to the field of surface detection research. Although there are general methods and theories for surface inspection, because the problem of surface inspection in different application fields is very different, a great deal of research and development of distinctive methods such as fabric surface inspection, wood surface inspection, steel surface inspection, etc. are aimed at specific application fields. With the development of the surface detection technology, chip surface detection is also concerned, but due to weak defects with the characteristics of low contrast between the defects and the background, small defects and the like, information such as defect targets, backgrounds, noise and the like is contained in a narrow gray scale range and is difficult to distinguish, and great challenges are brought to the chip surface defect automatic detection method based on the traditional image processing.
Disclosure of Invention
The invention provides a chip surface defect detection method based on a convolution denoising autoencoder, aiming at overcoming the problem of insufficient detection in weak defects in the traditional surface defect detection in the prior art, so that the weak defects can be effectively detected, the chip surface defects can be more volume-divided, and the chip surface defect detection precision can be improved.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
the invention relates to a chip surface detection method based on a convolution denoising autoencoder, which is characterized by comprising the following steps of:
step 1: defect-free image reconstruction based on a convolution denoising auto-encoder:
step 1.1: constructing a convolution denoising autoencoder and taking the autoencoder as a network model:
the network model is composed of an encoder, a full connection layer and a decoder; the encoder consists of n convolutional layers and n pooling layers; the decoder is composed of n deconvolution layers; and the encoder and decoder are connected by a full connection layer; the n deconvolution layers realize a deconvolution function by utilizing a nearest neighbor interpolation method and convolution;
step 1.2: constructing a training data set:
step 1.2.1: acquiring a chip surface defect-free image, carrying out image size normalization and Gaussian smooth denoising pretreatment, and obtaining a preprocessed chip surface image set which is recorded as I ═ I
1,I
2,…,I
i,...,I
n}，I
iRepresenting the ith chip surface picture in a chip surface image set I, wherein n represents the total number of images in a chip surface image data set I;
step 1.2.2: adopting a segmentation algorithm to carry out surface picture I on the ith chip
iDividing to obtain corresponding sub-image set
Picture I showing the surface of the ith chip
iA j-th sub-image after division, wherein m represents the number of image blocks obtained by division;
step 1.3: training a network model:
step 1.3.1: adding Gaussian noise to jth block sub-image
Carrying out damage processing to obtain corresponding data containing Gaussian noise
Thereby obtaining the ith sub-image set containing Gaussian noise
Further obtaining a chip surface data set containing Gaussian noise
Step 1.3.2: using a chip surface data set containing Gaussian noise
Training the network model to obtain a network model for reconstructing the chip surface image:
step 2: constructing a residual image based on the overlapping region:
step 2.1: carrying out size normalization and Gaussian smooth denoising pretreatment on a chip surface image to be detected to obtain a pretreated image Img; dividing the step length in the horizontal direction into s
wDivision step length in vertical direction is s
hThe preprocessed image Img is segmented to obtain a segmented image set to be detected, and the segmented image set is recorded as
Wherein, a
i,jRepresenting the image block in the ith row and the jth column, and the image block a in the ith row and the jth column
i,jCorresponds to the position of the ith row and jth column of the image Img after preprocessing; i is more than or equal to 1 and less than or equal to h; j is more than or equal to 1 and less than or equal to w; h and w represent the height and width of the image Img after preprocessing, respectively;
step 2.2: inputting the image block set A to be detected into a trained network model to obtain a reconstructed image block set A', wherein
Wherein, a'
i,jRepresenting the reconstructed block a of the ith row and the jth column
i,j；
Step 2.3: and (3) residual map synthesis:
step 2.3.1: obtaining an image block a of the ith row and the jth column by using the formula (1)
i,jAnd its corresponding reconstructed image block a'
i,jResidual theta between
i,jThus obtaining the residuals of all image blocks:
θ
i,j＝a
i′
,j-a
i,j(1)
step 2.3.2: synthesizing the residual errors of all the image blocks in the formula (2) to obtain a synthesized residual error image R
imgResidual matrix R (M, N):
in the formula (2), (m, n) represents the position coordinates of the pixel point in the m-th row and n-th column of any image block in the reconstructed image block set a', and θ
i,j(m, n) denotes the residual θ
i,jThe pixel points of the M-th row and the N-th column in the middle (M, N) represent a residual image R after synthesis
imgThe position coordinates of the pixel points in the M-th row and the N-th column, the initial value of R (M, N) is an all-zero matrix, N
pRepresents the number of image blocks, and n
pI × j; and comprises the following components:
M＝i×s
w+m (3)
N＝j×s
h+n (4)
and step 3: based on a residual map R
imgThe defect detection of (2):
step 3.1: obtaining residual error image R by maximum inter-class variance method
imgA segmentation threshold value T of (1);
step 3.2: according to the divided threshold value T, making a residual error map R
imgThe pixel points with the middle gray pixel value higher than T are defect areas, and the pixel points with the middle gray pixel value lower than or equal to T are non-defect areas, so that the residual image R is subjected to equation (5)
imgPerforming segmentation treatment to obtain a segmented residual map R'
img：
In the formula (5), R
img(i, j) denotes a residual map R
imgThe gray value at the ith row and jth column pixel point in the bit.
Compared with the prior art, the invention has the beneficial effects that:
1. the method provides a convolution denoising self-coding network model, obtains the mapping relation between a training sample and a non-defective template by learning a large number of Gaussian noise chip surface images, effectively extracts the characteristics of the chip surface images, and then realizes the defect detection of the sample by utilizing the contrast between a reconstructed image and a defect image, improves the detection efficiency of weak defects with the characteristics of low contrast between the defects and the background, smaller defects and the like, and has higher robustness compared with a chip surface defect automatic detection method based on the traditional image processing;
2. in the construction process of the residual error map, the strategy of the overlapping area is adopted, the contrast between the defect area and the non-defect area in the residual error map is effectively enhanced, the interference of non-defect noise on the defect area on the surface of a chip is inhibited, and the defect area in the residual error map can be easily segmented, so that the problem that information such as defect target, background, noise and the like is contained in a narrow gray scale range and is difficult to be distinguished is solved, and the defect area on the surface of the chip can be segmented by adopting a simple segmentation method;
3. the method is applied to the detection link of the surface defects in the chip production process. In the industrial production process of chips, the types of the defects on the surfaces of the chips cannot be predicted, all defect types are difficult to collect, and the workload of manually marking the defect images of the chips is overlarge, so that great difficulty is brought to the application of a method based on supervised learning. Compared with supervised learning, the unsupervised learning does not need to collect and manually mark a large number of defective samples in the training stage, so that the problem that defect types in chip surface images are difficult to collect in industrial production is solved, a large number of chip surface defect images are prevented from being manually marked, and the labor cost is saved;
4. the network model in the invention is simple to train, does not need users to have certain prior knowledge in the machine vision aspect, and is friendly to the users.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram of a network model architecture in the method of the present invention;
FIG. 3 is a comparison graph of the synthetic effect of the residual error map of the overlapping region of different segmentation step sizes in the method of the present invention;
FIG. 4 is a diagram showing the results of detecting the surface defects of the chip according to the present invention.
Detailed Description
In this embodiment, a method for detecting a chip surface defect based on a convolution denoising autoencoder, as shown in fig. 1, includes the following steps:
step 1: defect-free image reconstruction based on a convolution denoising auto-encoder:
step 1.1: constructing a convolution denoising autoencoder and taking the autoencoder as a network model:
the network model is composed of an encoder, a full connection layer and a decoder; the encoder consists of n-4 convolutional layers and 4 pooling layers; the decoder is composed of 4 deconvolution layers; and the encoder and decoder are connected through a full connection layer; the 4 deconvolution layers realize the deconvolution function by utilizing a nearest neighbor interpolation method and convolution. As shown in fig. 2, the specific parameters are as follows:
inputting: 28 × 28 × 1 single channel png format picture.
An encoder: consisting of 4 convolutional (C) layers and 4 pooling (P) layers, one after each convolutional layer. The first convolutional layer C1 contains 32 feature maps, the convolutional kernel size is 3 × 3, and the feature matrix size is 28 × 28; pooling layer P1 contained 32 feature maps with a feature matrix size of 28 × 28. The second convolutional layer C2 contains 32 feature maps, the convolutional kernel size is 3 × 3, and the feature matrix size is 14 × 14; pooling layer P2 contained 32 feature maps with a feature matrix size of 14 × 14. The third convolutional layer C3 contains 64 feature maps, the convolutional kernel size is 3 × 3, and the feature matrix size is 7 × 7; pooling layer P3 contained 64 feature maps with a feature matrix size of 7 × 7. The fourth convolution layer C4 contains 128 feature maps, the convolution kernel size is 3 x 3, and the feature matrix size is 4 x 4; pooling layer P4 contained 128 feature maps with a feature matrix size of 4 × 4. The pooling operation at this stage adopts mean pooling, so that the background information of the chip surface image can be better retained. The activation functions all adopt Relu activation functions.
Full connection layer: full connectivity layer (F) size 512, using the Relu activation function.
A decoder: the method is composed of 4 deconvolution layers, wherein a deconvolution function is realized by utilizing a nearest neighbor interpolation algorithm (N) and convolution (C) of an image. The first layer of deconvolution (N1, C5) contains 128 feature maps, with a convolution kernel size of 3 × 3 and a feature matrix size of 4 × 4; the second deconvolution layer (N2, C6) contains 64 feature maps, the convolution kernel size is 3 x 3, and the feature matrix size is 7 x 7; the third layer of deconvolution (N3, C7) contains 32 feature maps, the convolution kernel size is 3 x 3, and the feature matrix size is 14 x 14; the fourth deconvolution layer (N4, C8) contains 32 feature maps with convolution kernel size of 3 × 3 and feature matrix size of 28 × 28, where the activation functions of the first three deconvolution layers use Relu activation functions and the last deconvolution layer uses Sigmoid activation functions.
And (3) outputting: 28 × 28 × 1 single channel png format picture.
The above network model parameters are only preferred implementation parameters, and can effectively extract image features. Deepening the network hierarchy can increase computational cost and training cost, but network performance may not be significantly improved, and not as deep as possible.
Step 1.2: constructing a training data set:
step 1.2.1: acquiring a chip surface defect-free image, carrying out image size normalization and Gaussian smooth denoising pretreatment, and obtaining a preprocessed chip surface image set which is recorded as I ═ I
1,I
2,...,I
i,...,I
n}，I
iRepresenting the ith chip surface picture in a chip surface image set I, wherein n represents the total number of images in a chip surface image data set I;
step 1.2.2: adopting a segmentation algorithm to carry out surface picture I on the ith chip
iDividing to obtain corresponding sub-image set
Picture I showing the surface of the ith chip
iA j sub-image block after being divided, m represents the number of the divided image blocks, the divided image blocks should be of uniform size, such as 28 × 28, and the network structure in the method inputs one to one image;
step 1.3: training a network model:
step 1.3.1: adding Gaussian noise to jth block sub-image
Carrying out damage processing to obtain corresponding data containing Gaussian noise
Thereby obtaining the ith sub-image set containing Gaussian noise
Further obtaining a chip surface data set containing Gaussian noise
Step 1.3.2: using a chip surface data set containing Gaussian noise
Training the network model to obtain a network model for reconstructing the chip surface image, wherein the Loss function adopts a mean square method and adopts a self-adaptive gradient descent algorithm to optimize the network model;
step 2: constructing a residual image based on the overlapping region:
step 2.1: carrying out size normalization and Gaussian smooth denoising pretreatment on a chip surface image to be detected to obtain a pretreated image Img; dividing step s
wAnd s
hWill have a significant effect on the overlap area of neighboring blocks. Step length s
wAnd s
hThe larger the value, the less the overlapping area of the neighboring blocks, so that the number of times the pixel is reconstructed in Img is less, and thus the effect of constructing the residual map based on the overlapping area is poor, as shown in fig. 3. Therefore, in order to enhance the contrast between the defective area and the non-defective area in the residual map, the division step size in the horizontal direction is s
wDivision step size s in vertical direction is 1
h1, carrying out segmentation processing on the preprocessed image Img to obtain a segmented image set to be detected, and recording the segmented image set as a mark
Wherein, a
i,jRepresenting image blocks in ith row and jth column, and image block a in ith row and jth column
i,jCorresponds to the position of the ith row and jth column of the image Img after preprocessing; i is more than or equal to 1 and less than or equal to h; j is more than or equal to 1 and less than or equal to w; h and w represent the height and width of the image Img after preprocessing, respectively.
Step 2.2: inputting the image block set A to be detected into a trained network model to obtain a reconstructed image block set A', wherein
Wherein, a'
i,jRepresenting the reconstructed block a of the ith row and the jth column
i,j；
Step 2.3: and (3) residual map synthesis:
step 2.3.1: obtaining an image block a of the ith row and the jth column by using the formula (1)
i,jAnd its corresponding reconstructed image block a'
i,jResidual theta between
i,jThus obtaining the residuals of all image blocks:
θ
i,j＝a′
i,j-a
i,j(1)
step 2.3.2: synthesizing the residual errors of all the image blocks in the formula (2) to obtain a synthesized residual error image R
imgResidual matrix R (M, N):
in the formula (2), (m, n) represents the position coordinates of the pixel point in the m-th row and n-th column of any image block in the reconstructed image block set a', and θ
i,j(m, n) denotes the residual θ
i,jThe pixel points of the M-th row and the N-th column in the middle (M, N) represent a residual image R after synthesis
imgThe position coordinates of the pixel points in the M-th row and the N-th column, the initial value of R (M, N) is an all-zero matrix, N
pRepresents the number of image blocks, and n
pI × j; and comprises the following components:
M＝i×s
w+m (3)
N＝j×s
h+n (4)
and step 3: based on a residual map R
imgThe defect detection of (2):
step 3.1: by the residual error map synthesis method based on the overlapped area, the contrast between the defect area and the non-defect area in the residual error map is enhanced, so that the defect in the residual error map is easier to divide, and therefore the residual error map R is obtained by adopting the maximum inter-class variance method
imgA segmentation threshold value T of (1);
step 3.2: according to the divided threshold value T, making a residual error map R
imgThe pixel points with the middle gray pixel value higher than T are defect areas, the pixel points with the middle gray pixel value lower than or equal to T are non-defect areas, and the detection result is shown in figure 4, so that the residual error graph R is subjected to equation (5)
imgPerforming segmentation treatment to obtain a segmented residual map R'
img：
In the formula (5), R
img(i, j) denotes a residual map R
imgThe gray value at the ith row and jth column pixel point in the bit.
Claims (1)
1. A chip surface detection method based on a convolution denoising autoencoder is characterized by comprising the following steps:
step 1: defect-free image reconstruction based on a convolution denoising auto-encoder:
step 1.1: constructing a convolution denoising autoencoder and taking the autoencoder as a network model:
the network model is composed of an encoder, a full connection layer and a decoder; the encoder consists of n convolutional layers and n pooling layers; the decoder is composed of n deconvolution layers; and the encoder and decoder are connected by a full connection layer; the n deconvolution layers realize a deconvolution function by utilizing a nearest neighbor interpolation method and convolution;
step 1.2: constructing a training data set:
step 1.2.1: acquiring a chip surface defect-free image, carrying out image size normalization and Gaussian smooth denoising pretreatment, and obtaining a preprocessed chip surface image set which is recorded as I ═ I
1,I
2,...,I
p,...,I
u}，I
pRepresenting the p-th chip surface picture in a chip surface image set I, and u represents the total number of images in a chip surface image data set I;
step 1.2.2: adopting a segmentation algorithm to carry out surface picture I on the p-th chip
pDividing to obtain corresponding sub-image set
Picture I showing the surface of the p-th chip
pA q-th divided sub-image block, m representing the number of divided image blocks;
step 1.3: training a network model:
step 1.3.1: adding Gaussian noise to sub-image of q block
Carrying out damage processing to obtain corresponding data containing Gaussian noise
Thereby obtaining the p sub-image set containing Gaussian noise
Further obtaining a chip surface data set containing Gaussian noise
Step 1.3.2: using a chip surface data set containing Gaussian noise
Training the network model to obtain a network model for reconstructing the chip surface image:
step 2: constructing a residual image based on the overlapping region:
step 2.1: carrying out size normalization and Gaussian smooth denoising pretreatment on a chip surface image to be detected to obtain a pretreated image Img; dividing the step length in the horizontal direction into s
wDivision step length in vertical direction is s
hThe preprocessed image Img is segmented to obtain a segmented image set to be detected, and the segmented image set is recorded as
Wherein, a
i,jRepresenting the image block in the ith row and the jth column, and the image block a in the ith row and the jth column
i,jCorresponds to the position of the ith row and jth column of the image Img after preprocessing; i is more than or equal to 1 and less than or equal to h; j is more than or equal to 1 and less than or equal to w; h and w represent the height and width of the image Im g after preprocessing, respectively;
step 2.2: inputting the image block set A to be detected into a trained network model to obtain a reconstructed image block set A', wherein
Wherein, a'
i,jRepresenting the reconstructed image block of the ith row and the jth column;
step 2.3: and (3) residual map synthesis:
step 2.3.1: obtaining an image block a of the ith row and the jth column by using the formula (1)
i,jAnd its corresponding reconstructed image block a'
i,jResidual theta between
i,jThus obtaining the residuals of all image blocks:
θ
i,j＝a′
i,j-a
i,j(1)
step 2.3.2: synthesizing the residual errors of all the image blocks in the formula (2) to obtain a synthesized residual error image R
imgResidual matrix R (M, N):
in the formula (2), (e, v) represents the position coordinates of the pixel points in the e-th row and the v-th column of any image block in the reconstructed image block set A', and θ
i,j(e, v) denotes the residualθ
i,jThe position coordinates of the pixel points in the ith row and the vth column in the middle row, and (M, N) represent a residual image R after synthesis
imgThe position coordinates of the pixel points in the M-th row and the N-th column, the initial value of R (M, N) is an all-zero matrix, N
pRepresents the number of image blocks, and n
pI × j; and comprises the following components:
M＝i×s
w+e (3)
N＝j×s
h+v (4)
and step 3: based on a residual map R
imgThe defect detection of (2):
step 3.1: obtaining residual error image R by maximum inter-class variance method
imgA segmentation threshold value T of (1);
step 3.2: according to the segmentation threshold T, making a residual error map R
imgThe pixel points with the middle gray pixel value higher than T are defect areas, and the pixel points with the middle gray pixel value lower than or equal to T are non-defect areas, so that the residual image R is subjected to equation (5)
imgPerforming segmentation treatment to obtain a segmented residual map R'
img：
In the formula (5), R
img(i, j) denotes a residual map R
imgThe gray value at the ith row and jth column pixel point in the bit.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910078674.6A CN109829903B (en) | 2019-01-28 | 2019-01-28 | Chip surface defect detection method based on convolution denoising autoencoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910078674.6A CN109829903B (en) | 2019-01-28 | 2019-01-28 | Chip surface defect detection method based on convolution denoising autoencoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109829903A CN109829903A (en) | 2019-05-31 |
| CN109829903B true CN109829903B (en) | 2020-02-11 |
Family
ID=66862568
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910078674.6A Active CN109829903B (en) | 2019-01-28 | 2019-01-28 | Chip surface defect detection method based on convolution denoising autoencoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109829903B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111046080A (en) * | 2019-12-05 | 2020-04-21 | 武汉理工大学 | Carbon fiber precursor monitoring data preprocessing method based on convolution denoising autoencoder |
| CN111127412B (en) * | 2019-12-18 | 2021-05-28 | 广东工业大学 | Pathological image recognition device based on generation countermeasure network |
| CN111402197B (en) * | 2020-02-09 | 2023-06-16 | 西安工程大学 | Detection method for colored fabric cut-parts defect area |
| CN111598857A (en) * | 2020-05-11 | 2020-08-28 | 北京阿丘机器人科技有限公司 | Method and device for detecting surface defects of product, terminal equipment and medium |
| CN112116557B (en) * | 2020-08-12 | 2022-02-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
| CN111968095B (en) * | 2020-08-19 | 2022-08-02 | 成都数之联科技股份有限公司 | Product surface defect detection method, system, device and medium |
| CN114079776A (en) * | 2020-08-21 | 2022-02-22 | Oppo广东移动通信有限公司 | Image processing method, device, equipment, computer storage medium and system |
| CN113838041B (en) * | 2021-09-29 | 2023-09-08 | 西安工程大学 | Method for detecting defect area of color texture fabric based on self-encoder |
| CN114022409B (en) * | 2021-09-30 | 2023-04-18 | 电子科技大学 | Coated medicine surface defect detection algorithm based on deep learning |
| CN113989228A (en) * | 2021-10-27 | 2022-01-28 | 西安工程大学 | Method for detecting defect area of color texture fabric based on self-attention |
| CN114022442B (en) * | 2021-11-03 | 2022-11-29 | 武汉智目智能技术合伙企业(有限合伙) | Unsupervised learning-based fabric defect detection algorithm |
| CN114972327B (en) * | 2022-07-12 | 2023-01-06 | 爱尔达电气有限公司 | Semiconductor package test system and test method thereof |
| CN115358998B (en) * | 2022-08-22 | 2023-06-16 | 法博思(宁波)半导体设备有限公司 | Method and system for acquiring point coordinates in random array picture |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105976381A (en) * | 2016-05-10 | 2016-09-28 | 电子科技大学 | Quantitative evaluation method for large-area defect of pipeline based on far-field eddy false peak removal |
| CN107563433A (en) * | 2017-08-29 | 2018-01-09 | 电子科技大学 | A kind of infrared small target detection method based on convolutional neural networks |
| CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
| CN108345911A (en) * | 2018-04-16 | 2018-07-31 | 东北大学 | Surface Defects in Steel Plate detection method based on convolutional neural networks multi-stage characteristics |
| CN108389181A (en) * | 2018-01-23 | 2018-08-10 | 南京大学 | It is a kind of based on the part defect of deep learning and normal direction figure detection and localization method |
| WO2018165753A1 (en) * | 2017-03-14 | 2018-09-20 | University Of Manitoba | Structure defect detection using machine learning algorithms |
| CN108876737A (en) * | 2018-06-06 | 2018-11-23 | 武汉大学 | A kind of image de-noising method of joint residual error study and structural similarity |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106650770B (en) * | 2016-09-29 | 2019-12-17 | 南京大学 | Mura defect detection method based on sample learning and human eye visual characteristics |
| CN108061735A (en) * | 2017-12-01 | 2018-05-22 | 工业互联网创新中心（上海）有限公司 | The recognition methods of component surface defect and device |
| CN108334907B (en) * | 2018-02-09 | 2019-05-17 | 哈尔滨工业大学（威海） | A kind of complex equipment point method for detecting abnormality and system based on deep learning |
-
2019
- 2019-01-28 CN CN201910078674.6A patent/CN109829903B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105976381A (en) * | 2016-05-10 | 2016-09-28 | 电子科技大学 | Quantitative evaluation method for large-area defect of pipeline based on far-field eddy false peak removal |
| WO2018165753A1 (en) * | 2017-03-14 | 2018-09-20 | University Of Manitoba | Structure defect detection using machine learning algorithms |
| CN107563433A (en) * | 2017-08-29 | 2018-01-09 | 电子科技大学 | A kind of infrared small target detection method based on convolutional neural networks |
| CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
| CN108389181A (en) * | 2018-01-23 | 2018-08-10 | 南京大学 | It is a kind of based on the part defect of deep learning and normal direction figure detection and localization method |
| CN108345911A (en) * | 2018-04-16 | 2018-07-31 | 东北大学 | Surface Defects in Steel Plate detection method based on convolutional neural networks multi-stage characteristics |
| CN108876737A (en) * | 2018-06-06 | 2018-11-23 | 武汉大学 | A kind of image de-noising method of joint residual error study and structural similarity |
Non-Patent Citations (3)
| Title |
|---|
| Detection of a casting defect tracked by deep convolution neural network;Jinhua Lin 等;《The International Journal of Advanced Manufacturting Technology》;20180407(第97期);第573-581页 * |
| 基于卷积神经网络的工件缺陷检测研究;于科为;《信息与电脑》;20181231(第21期);第7-9,12页 * |
| 基于深度学习的轮胎缺陷无损检测与分类技术研究;崔雪红;《中国博士学位论文全文数据库 工程科技II辑》;20181015;第2018年卷(第10期);第C035-4页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109829903A (en) | 2019-05-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109829903B (en) | Chip surface defect detection method based on convolution denoising autoencoder | |
| Ren et al. | State of the art in defect detection based on machine vision | |
| CN108230339B (en) | Stomach cancer pathological section labeling completion method based on pseudo label iterative labeling | |
| CN111553929B (en) | Mobile phone screen defect segmentation method, device and equipment based on converged network | |
| CN107561738B (en) | Fast TFT-LCD surface defect detection method based on FCN | |
| Mei et al. | An unsupervised-learning-based approach for automated defect inspection on textured surfaces | |
| CN106875381B (en) | Mobile phone shell defect detection method based on deep learning | |
| CN109447979B (en) | Target detection method based on deep learning and image processing algorithm | |
| CN113643268B (en) | Industrial product defect quality inspection method and device based on deep learning and storage medium | |
| CN112037219A (en) | Metal surface defect detection method based on two-stage convolution neural network | |
| CN113706490B (en) | Wafer defect detection method | |
| CN112329588A (en) | Pipeline fault detection method based on fast R-CNN | |
| CN111932511A (en) | Electronic component quality detection method and system based on deep learning | |
| Wang et al. | Attention-based deep learning for chip-surface-defect detection | |
| CN116612106A (en) | Method for detecting surface defects of optical element based on YOLOX algorithm | |
| CN113393426A (en) | Method for detecting surface defects of rolled steel plate | |
| CN115829942A (en) | Electronic circuit defect detection method based on non-negative constraint sparse self-encoder | |
| CN117422696A (en) | Belt wear state detection method based on improved YOLOv8-Efficient Net | |
| CN114820541A (en) | Defect detection method based on reconstructed network | |
| CN115713480A (en) | Integrated circuit metal packaging deep learning defect detection method | |
| CN111325724B (en) | Tunnel crack region detection method and device | |
| CN115619799B (en) | Grain image segmentation method and system based on transfer learning | |
| Kögel et al. | Machine learning assisted signal analysis in Acoustic Microscopy for non-destructive defect identification | |
| CN116843611A (en) | Hardware surface flaw detection system based on difference and graph neural network | |
| CN113850335B (en) | Data augmentation method for bathroom ceramic defect detection |
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 |