CN112233059A - Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network - Google Patents

Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network Download PDF

Info

Publication number
CN112233059A
CN112233059A CN202010814606.4A CN202010814606A CN112233059A CN 112233059 A CN112233059 A CN 112233059A CN 202010814606 A CN202010814606 A CN 202010814606A CN 112233059 A CN112233059 A CN 112233059A
Authority
CN
China
Prior art keywords
decision
network
segmentation
training
light guide
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
Application number
CN202010814606.4A
Other languages
Chinese (zh)
Inventor
李俊峰
李兆攀
王昊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN202010814606.4A priority Critical patent/CN112233059A/en
Publication of CN112233059A publication Critical patent/CN112233059A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • 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]

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a light guide plate defect detection method based on a 'segmentation + decision' two-stage residual error attention network, which comprises the following steps of: step 1, collecting images of a light guide plate; step 2, preprocessing the light guide plate image; step 3, establishing and training a two-stage segmentation + decision-making residual error attention network, including a segmentation sub-network and a decision sub-network, and then storing and outputting the trained two-stage segmentation + decision-making residual error attention network; and 4, performing defect detection by using the trained two-stage residual error attention network of segmentation and decision, and obtaining a result. The method can realize the purpose of quickly and accurately distinguishing the scratches or the defects on the light guide plate from the collected light guide plate image.

Description

Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network
Technical Field
The invention belongs to the field of image recognition of deep learning, and particularly relates to a light guide plate defect detection method based on a 'segmentation + decision' two-stage residual error attention network.
Background
The Light Guide Plate (Light Guide Plate, LGP) is mainly made of optical acrylic Plate, the chemical name of which is methyl methacrylate, and is formed by utilizing optical-grade acrylic Plate and then using high-tech materials with extremely high reflectivity and no Light absorption to print Light Guide points on the bottom surface of the optical-grade acrylic Plate by using laser engraving, cross grid engraving and screen printing technologies. The light guide plate has the advantages of being ultrathin, ultra-bright, uniform in light guide, energy-saving, environment-friendly, durable, simple and fast to install and maintain and the like, and therefore the light guide plate is widely applied to occasions such as liquid crystal display, advertising lamps, lights, panel lamp illumination and the like. In the production and manufacturing processes of silk-screen printing, chemical etching, laser processing, bump processing and the like of the light guide plate, due to the influence of factors such as raw material components, equipment use conditions, processing technology, worker operation and the like, processing defects such as bright spots, missing dots, screen surface ink, line scratches, mirror surface scratches and the like inevitably occur on the surface of the light guide plate. The existence of light guide plate defect can influence the use of relevant equipment, leads to the availability factor of equipment, and luminous homogeneity and life-span etc. all can receive the influence, and in addition, the credit of enterprise can seriously be harmd in the sales of defect light guide plate, causes great negative effect to the long-term development of enterprise, consequently, carries out quality testing to the light guide plate of production, rejects the inferior product and is especially important.
At present, domestic light guide plate defect detection is mainly completed by manual operation, but the manual detection has obvious limitation and numerous disadvantages. The following disadvantages exist in manual detection: 1. the detection quality is unstable due to subjective judgment and long-time eye fatigue; 2. the labor cost is high; 3. the labor intensity is high; 4. low labor efficiency and the like. Therefore, it is necessary to develop a device to replace the manual inspection of the light guide plate, and the development of the algorithm is particularly important.
In the conventional visual inspection method, features are generally designed manually through a conventional image processing method, such as morphological operation, a histogram and the like, and then defect detection is performed through a formulated decision rule or a learning method, such as a classifier of a KNN, a decision tree, an SVM and the like. Because the light guide points are changed from sparse to dense from the light incident side to the light emergent side of the light guide plate, the light guide plate defect detection method based on the traditional image processing needs to carry out partition processing, and each partition corresponds to a corresponding defect detection algorithm, so that the inconvenience is brought to algorithm maintenance; aiming at different types of defects, the traditional image processing algorithm needs to set different judgment conditions to detect the defects; in addition, when the classification network is used for defect detection, the network training depends on a large number of positive and negative samples, and under the condition that the sample amount is insufficient, the network is difficult to extract accurate features, so that the defect detection capability is greatly reduced; therefore, an efficient and accurate method based on deep learning is needed for detecting complex defects of the light guide plate.
Disclosure of Invention
The invention aims to provide a light guide plate defect detection method based on a 'segmentation + decision' two-stage residual error attention network, so as to realize the purpose of quickly and accurately distinguishing scratches or defects on a light guide plate from an acquired light guide plate image.
In order to solve the technical problem, the invention provides a light guide plate defect detection method based on a 'segmentation + decision' two-stage residual error attention network, which comprises the following steps:
step 1, light guide plate image acquisition: at the tail end of the light guide plate production line, carrying out light guide plate image acquisition by using a high-resolution industrial camera and transmitting the light guide plate image to an upper computer for processing;
step 2, preprocessing the light guide plate image:
in the upper computer, a black background contained in the light guide plate image acquired in the step 1 is removed by utilizing a threshold segmentation technology, and a regional image of the ROI of the light guide plate is obtained;
then cutting the regional image of the light guide plate ROI into a group of small images with the size of 1 × 224, wherein 1/10 image width overlaps between the adjacent images;
step 3, establishing and training a two-stage segmentation + decision-making residual error attention network, including a segmentation sub-network and a decision sub-network, and then storing and outputting the trained two-stage segmentation + decision-making residual error attention network;
step 4, using the trained two-stage residual error attention network of 'segmentation + decision' to detect defects, and obtaining a result:
and (3) sending a group of images with the size of 1 × 224 obtained in the step 2 into the two-stage residual attention network of segmentation + decision trained in the step 3, obtaining corresponding masks from the segmentation subnetworks, presenting the outlines and positions of the defects, obtaining decision results from the decision subnetworks, wherein the decision results are greater than 0.5 and less than 0.5, the defects are contained, and finally the corresponding masks and decision results are marked in the original image.
The invention relates to an improvement of a light guide plate defect detection method based on a 'segmentation + decision' two-stage residual error attention network, which comprises the following steps:
the step 3 of establishing, training and testing the two-stage residual error attention network of 'segmentation + decision' is as follows:
step 3-1, constructing a segmentation subnet:
the segmentation sub-network adopts a U-shaped network structure and comprises an encoder and a decoder, wherein the encoder comprises a standard convolution Conv 3+ 3+ batch normalization BN + activation function layer with the kernel size of 3 × 3 and three groups of residual error attention unit RAU + Average pooling Average-Pool layers from top to bottom, and feature maps with four resolutions are output in sequence; the decoder comprises three cascaded skip connection structures from bottom to top, Feature Maps with four resolutions are fused to obtain a final Feature map Last Feature Maps, and finally the final Feature map Last Feature Maps are integrated with features through a standard convolution Conv1 x 1+ batch normalization BN + activation function layer with the kernel size of 1 x 1 to obtain an Output Feature map Segmentation Output of a segmented subnet with the size of 1 x 224;
step 3-2, constructing a decision subnet:
the decision sub-network comprises an important decision node and an auxiliary decision node; taking the output characteristic graph of the sub-network segmented in the step 3-1 as the most important information, and directly and respectively obtaining two important decision nodes through a global average pooling GAP layer and a global maximum pooling GMP layer; taking a final feature graph output by a third skip layer connection structure in a split sub-network and an output feature graph of the split sub-network as input, sequentially passing through a group of standard convolution Conv3 x 3+ maximum pooling Max-Pool layer structures with a kernel size of 3 x 3 and two groups of residual error attention units RAU + maximum pooling Max-Pool layer structures, extracting auxiliary information from the output feature graph of the split sub-network and the final feature graph output by the third skip layer connection structure, and sequentially passing through a global average pooling GAP layer and a global maximum pooling GMP layer to obtain an auxiliary decision node; finally, outputting a final Decision node Decision Output through a fully connected full connected layer, wherein the Output of the node is the probability that the defect is true;
combining the divided subnetworks and the decision subnet into a whole network which is a two-stage residual error attention network of 'division + decision';
step 3-3, training and testing the network:
1. establishing a loss function
Loss function Loss adopted for split subnet trainingsegmentationAs shown in the following equation:
Losssegmentation＝0.5×Lossbce+0.5×Lossdice，
Figure BDA0002632212920000031
Figure BDA0002632212920000032
wherein H represents the height of the mask label, W represents the width of the mask label, yiFor a certain pixel value of the mask label,
Figure BDA0002632212920000033
for dividing up sub-networksOutput corresponding pixel value, label is mask label, label*To split subnet output, smooth is a fixed value to prevent divide by 0 errors;
loss function Loss adopted by decision sub-networkdecisionAs shown in the following equation:
Lossdecision＝ylog(y*)+(1-y)log(1-y*)，
wherein y is a classification label, y*Outputting for the decision subnet;
2. building a training data set
Collecting light guide plate images in the step 1, wherein the light guide plate images comprise 422 defective images and 400 normal images, and each image is manually provided with two labels: one is a mask label, the other is a classification label, the classification label corresponding to the light guide plate image containing the defect is set to be 1, and the classification label corresponding to the light guide plate image without the defect is set to be 0;
setting 75 defect images, 75 normal images and corresponding labels in the image set as a training set, setting the other 75 defect images, 75 normal images and corresponding labels as a verification set, and setting the rest 272 defect images, 250 normal images and corresponding labels as a test set;
3. network training
Training a two-stage residual error attention network of 'segmentation + decision' built by PyTorch according to the step 3-1 and the step 3-2, wherein the batch size is taken as 8 in the training, and an Adam optimizer is adopted in the optimizer;
first, independently training and dividing sub-network
Training by taking the training set as the input of the divided subnets, and passing the Loss function Loss corresponding to the output of the divided subnets and the mask labelssegmentationCalculating the loss of a training set of the subnet segmented by the current round, adjusting the network parameters through a back propagation algorithm and a gradient descent algorithm to enable the loss of the training set to be continuously descended, calculating the loss of a verification set on the verification set, and if the loss of the verification set in the current round is lower than the loss of the verification set in the previous round, storing the current model and the model parameters of the current round, including connection of each hop layerObtaining a feature graph output by the skip connection structure and dividing an output feature graph of the subnet; thus, the subnet is divided for 500 training rounds;
freezing and dividing the subnet, training decision subnet
In the process of training the decision subnet, parameters for segmenting the subnet cannot be updated; in the first step, the training set obtains the final characteristic diagram output by the third skip connection structure through the division of the subnet and obtains the output characteristic diagram of the division of the subnet, the final characteristic diagram and the output characteristic diagram are used as the input of the decision subnet for training, and the output and the classification label corresponding to the decision subnet pass through the Loss function LossdecisionCalculating the loss of a training set of a current round decision subnet, adjusting network parameters through a back propagation algorithm and a gradient descent algorithm to enable the loss of the training set to be continuously descended, then calculating the loss of a verification set on the verification set, if the loss of the verification set in the current training round is lower than the loss of the verification set in the previous round, saving a model and model parameters of the current round, and performing 100 rounds of total training of the decision subnet;
after training is finished, a trained two-stage residual error attention network of 'segmentation + decision' is obtained;
third, off-line testing
Detecting the trained two-stage residual error attention network of 'segmentation + decision' by using a test set, and counting the accuracy of detection:
the accuracy rate is (1- (number of missed detections + number of false detections)/total number of detections) 100%,
the test result reaches 99.808%;
the light guide plate defect detection method based on the two-stage residual error attention network of the invention is further improved as follows:
the residual attention unit RAU includes: a main branch and an attention branch, wherein the main branch adopts a depth separable convolution DSConv, and the attention branch comprises a channel attention module and a position attention module; in the channel attention module, after passing through a global average pooling GAP layer and a global maximum pooling GMP layer, the channel attention module is input into an add layer, and then, the channel attention module passes through two standard convolution Conv 1+ batch normalization BN + activation function layers with the kernel size of 1 × 1 in sequence; the position attention position attribution module respectively operates a Mean layer through averaging along the channel direction and a Max layer through maximum value computing along the channel direction, then splicing Concat operation is carried out along the channel direction, then a global convolution network GCN is used, and finally a standard convolution Conv 1+ batch normalization BN + activation function layer with the kernel size of 1 x 1 is used;
the implementation process of the residual attention unit RAU is shown as the following formula:
F(x)＝(1+A(x))×T(x)，
A(x)＝C(x)×P(x)，
where x represents input, a (x) represents output of attention branch, t (x) represents output of trunk branch, c (x) represents output of channel attention module, p (x) represents output of position attention module, and f (x) represents output of residual attention unit RAU.
The light guide plate defect detection method based on the two-stage residual error attention network of the invention is further improved as follows:
the Skip layer Connection Skip Connection structure adopts two feature maps with different resolutions as input, firstly, Up-sampling is carried out on a Low-resolution feature map to obtain a new Low-resolution feature map, then, a High-resolution feature map and the new Low-resolution feature map are spliced together through a splicing Concat operation, and finally, feature fusion is completed through a standard convolution Conv 3+ 3+ batch normalization BN + activation function layer with the kernel size of 3 to obtain the new High-resolution feature map.
The invention has the following advantages:
1. compared with the traditional image processing algorithm, the method does not need to consider the light guide point density problem and partition processing, and can solve the problems through learning; aiming at different types of defects, the defects are detected without setting different judgment conditions, and the generalization capability of the defect detection can be obtained through positive and negative sample training;
compared with a single classification network, the training of the method does not need to rely on a large number of positive and negative samples, the defect positioning and segmentation are carried out by utilizing the segmentation sub-network, and the decision is carried out by utilizing the decision sub-network on the basis, so that the decision blindness can be avoided, and the decision accuracy under the condition of training a model by a small number of samples can be ensured;
3. the invention utilizes the constructed residual error attention unit (RAU) to replace the standard convolution layer, can complete the accurate positioning of the meaningful feature, simultaneously inhibits the meaningless feature, and is beneficial to improving the segmentation performance of the small defect; the introduction of a residual error structure ensures that the network can still be easily optimized and learned when the units are superposed; meanwhile, the network performance is improved by comparing the network containing the RAU with the network without the RAU;
in general, the detection algorithm of the invention has strong universality, strong stability and high detection precision, meets the precision requirement of enterprises, and can be applied to actual production.
Drawings
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.
FIG. 1 is a flow chart of a light guide plate defect detection method based on a "segmentation + decision" two-stage residual error attention network according to the present invention;
FIG. 2 is a schematic structural diagram of a two-stage residual attention network of "segmentation + decision" according to the present invention;
FIG. 3 is a schematic diagram of the RAU structure of FIG. 2;
FIG. 4 is a schematic diagram of the layer jump connection structure shown in FIG. 2;
fig. 5 is a comparison graph of the input-output images of the two-stage residual attention network of "segmentation + decision" in the comparison example and the comparison network without the residual attention unit RAU.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto.
Embodiment 1, a method for detecting a defect of a light guide plate based on a "segmentation + decision" two-stage residual attention network, for example, in fig. 1, a dashed path is a network establishment and training process, and a solid path is an online production process, and specifically includes the following steps:
s01, light guide plate image acquisition
And at the tail end of the light guide plate production line, carrying out light guide plate image acquisition by using a high-resolution industrial camera, and then transmitting the acquired light guide plate image to an upper computer for processing.
S02 light guide plate image preprocessing
S0201, light guide plate ROI area extraction
The light guide plate image acquired in the S01 contains a black background, the background is removed by using a threshold segmentation technology in an upper computer, a light guide plate ROI area is extracted, and an area image of the light guide plate ROI is acquired;
s0202, image cropping
Cutting the region image of the light guide plate ROI acquired by the S0201 into a group of small images with the size of 1 × 224, wherein the adjacent images are overlapped by 1/10 image widths, and the overlapping of the adjacent images is to ensure the integrity of the defect and prevent the complete defect from being segmented due to image cutting; this set of small images 1 x 224 serves as input for the subsequent steps.
S03, establishing a segmentation and decision two-stage residual attention network
Constructing a 'segmentation + decision' two-stage residual attention network by using PyTorch, wherein the 'segmentation + decision' two-stage residual attention network comprises a segmentation subnet (segmentation subnet) and a decision subnet (decision subnet); as shown in fig. 2, on the basis of extracting the defect feature by dividing the subnet, the decision whether the defect exists in the subnet is made and the judgment result is output: the probability that the defect is true is greater as the probability approaches 1; the closer the probability is to 0, the smaller the probability of the existence of the defect;
s0301, construction of a split subnet
The segmentation sub-network adopts a U-shaped network structure comprising an encoder and a decoder, and is used for segmenting and positioning the defects of the input 1 × 224 small images (namely defect feature extraction); the encoder comprises a standard convolution (Conv3 x 3) + Batch Normalization (BN) + activation function layer and three groups of residual error attention unit (RAU) + Average pooling (Average-Pool) layers with a kernel size of 3 x 3 sequentially from top to bottom to respectively obtain feature maps with four resolutions; the decoder comprises a step of fusing Feature Maps with four resolutions respectively through three cascaded skip connection structures from bottom to top to obtain a final Feature map (Last Feature Maps), and a step of integrating the final Feature map (Last Feature Maps) through a standard convolution (Conv1 1) with a kernel size of 1 × 1 and Batch Normalization (BN) and an activation function layer to obtain an Output Feature map (Segmentation Output) with a size of 1 × 224, wherein the specific process is as follows:
s030101, set up Batch Normalization (BN)
Batch Normalization (BN), i.e. Batch Normalization, is the input of pixel points xiSubtract mean μ and divide by mean variance
Figure BDA0002632212920000071
To obtain a normalized value xiThen carrying out scale conversion and offset to obtain a value y after batch normalization processingiWherein:
Figure BDA0002632212920000072
n is the batch size, epsilon is a fixed value to prevent divide by 0 errors, and gamma and beta are network learned parameters;
in the structures of a residual error attention unit (RAU) and a skip connection (skip connection), batch normalization BN operation is performed by default after corresponding convolution layers, so that network optimization is facilitated;
s030102, establishing an activation function ReLU and a Sigmoid, wherein:
Figure BDA0002632212920000073
Figure BDA0002632212920000074
wherein x is a characteristic diagram;
in the structures of a Residual Attention Unit (RAU) and a skip connection (skip connection), an activation function is defaulted after a Batch Normalization (BN) operation, when the required output range is between 0 and 1, a Sigmoid is adopted, and in other cases, a ReLU is adopted;
s030103, construction of Residual Attention Unit (RAU)
Each Residual Attention Unit (RAU) includes two branches, such as fig. 3, a trunk branch and an attention branch, the trunk branch performs feature extraction, and the number of parameters can be reduced while ensuring performance by adopting depth separable convolution (DSConv), namely, depthwise partial constraint, and the attention branch includes a channel attention module and a position attention module;
in a channel attention (channel attention) module, global information is extracted through a Global Average Pooling (GAP) layer and a Global Maximum Pooling (GMP) layer, then the global information is input into an add layer, next, a new channel weight C (x) is obtained through two standard convolutions (Conv1 1) + Batch Normalization (BN) + activation function layers with the kernel size of 1 × 1 in sequence, and the channel attention (channel attention) module can enable a network to pay more attention to meaningful features and inhibit meaningless features; however, the new weight of each channel obtained by the channel attention module plays the same role for each position of the corresponding feature map, which means that the channel attention module can only focus on what features are meaningful, but cannot focus on the specific position of the meaningful features on the feature map; the method comprises the steps that a position attention module obtains two feature graphs through an averaging operation (Mean) layer along a channel direction and a maximum value operation (Max) layer along the channel direction, splicing (Concat) operation is conducted along the channel direction to synthesize one feature graph, then a Global Convolutional Network GCN is used for extracting position attention information, and finally a standard convolution (Conv1 1) + Batch Normalization (BN) + activation function layer with the kernel size of 1 x 1 is used for obtaining a new position weight P (x), so that the Network can pay more attention to the position information of features, and meanwhile interference information on the feature graphs is restrained;
the implementation of the Residual Attention Unit (RAU) is shown in the following formula:
F(x)＝(1+A(x))×T(x)，
A(x)＝C(x)×P(x)，
wherein x represents input, a (x) represents output of attention branch (branch), t (x) represents output of trunk branch, c (x) represents output of channel attention module, p (x) represents output of position attention module, f (x) represents output of Residual Attention Unit (RAU);
s030104, construction of Skip Connection structure
Taking two feature maps with different resolutions as input, as shown in fig. 4, the size of the High-resolution feature map (High-resolution feature maps) is C1 × H × W, the size of the Low-resolution feature map (Low-resolution feature maps) is C2 × H/2 (W/2), first, the Low-resolution feature map (Low-resolution feature maps) is Up-sampled (Up-sampling) to obtain a new Low-resolution feature map (size: C2 × H W), the High-resolution feature map (High-resolution feature maps) and the new Low-resolution feature map are spliced together by a splicing (Concat) operation to obtain an integrated feature map, the size is (C1+ C2) × W, and finally, the High-resolution feature map (High-resolution feature map) is obtained by a standard (constant 3 × 3) convolution with kernel size 3, and the High-resolution feature map (High-resolution feature map, normalized BN, the High-resolution feature map (High-resolution feature map) + fusion function is obtained, size C1 × H × W;
s030105, construction of U-shaped structure split subnet
Segmenting the subnetwork to include an encoder and a decoder, where the input 1 x 224 image is passed through a standard convolution with kernel size 3 x 3 (Conv3 x 3) + Batch Normalization (BN) + activation function layer to obtain a first resolution profile (size: 16 x 224); then obtaining a second resolution feature map (size: 32 × 112) by the first Residual Attention Unit (RAU) + Average pooling layer (Average-Pool); then, a third resolution feature map (size: 64 × 56) is obtained by a second Residual Attention Unit (RAU) + Average pooling layer (Average-Pool); finally, obtaining a fourth resolution feature map (size: 128 × 28) through a third Residual Attention Unit (RAU) + Average pooling layer (Average-Pool); fusing the fourth resolution feature map and the third resolution feature map through a first skip connection structure in a decoder part to generate a new third resolution feature map (size: 64 × 56); fusing the new third resolution feature and the second resolution feature map to generate a new second resolution feature map (size: 32 × 112) by using a second skip connection structure; then, fusing the new second resolution Feature and the first resolution Feature map through a third skip connection (skip connection) structure to generate a final Feature map (Last Feature Maps, size: 16 × 224); finally, integrating the final feature map by a standard convolution (Conv1 × 1) with the kernel size of 1 × 1, Batch Normalization (BN) and an activation function layer to obtain an Output feature map (Segmentation Output, size: 1 × 224) of the segmented subnet;
in the sub-network segmentation, the performance of a constructed residual error attention unit (RAU) is used for replacing a standard convolution layer to be improved, the precise positioning of meaningful features can be completed, simultaneously meaningless features are inhibited, the segmentation performance of small defects is favorably improved, the introduction of a residual error structure ensures that the network can still be easily optimized and learned when the units are overlapped;
s0302: constructing decision subnets
A decision sub-network comprising an important decision node and an auxiliary decision node; taking the output characteristic graph of the S0301 segmented sub-network as the most important information, and directly and respectively obtaining two important decision nodes through a Global Average Pooling (GAP) layer and a Global Maximum Pooling (GMP) layer; secondly, taking a final feature graph output by a third skip connection structure in S0301 and an output feature graph of an S0301 split subnet as input, sequentially passing through a group of standard convolution (Conv3 x 3) + maximum pooling (Max-Pool) layer structures with a kernel size of 3 x 3 and two groups of Residual Attention Units (RAU) + maximum pooling (Max-Pool) layer structures, extracting auxiliary information from the output feature graph of the split subnet and the final feature graph output by the third skip connection structure, and sequentially passing through a Global Average Pooling (GAP) layer and a Global Maximum Pooling (GMP) layer to obtain an auxiliary decision node; finally, outputting a final Decision node (Decision Output) through a full connected (full connected) layer, wherein the Output of the node is the probability that the defect is true;
combining the divided subnetworks and the decision subnet into a whole network which is a two-stage residual error attention network of 'division + decision';
s04: training and testing network
S0401 establishing a loss function
Loss function Loss adopted for split subnet trainingsegmentationAs shown in the following equation:
Losssegmentation＝0.5×Lossbce+0.5×Lossdice，
Figure BDA0002632212920000101
Figure BDA0002632212920000102
wherein H represents the height of the mask label, W represents the width of the mask label, yiFor a certain pixel value of the mask label,
Figure BDA0002632212920000103
for dividing the corresponding pixel values output by the sub-network, label is the mask label*To split subnet output, smooth is a fixed value to prevent divide by 0 errors;
loss function Loss adopted by decision sub-networkdecisionAs shown in the following equation:
Figure BDA0002632212920000104
wherein y is a classification label, y*Outputting for the decision subnet;
s0402, set up training data set
Constructing a light guide plate image data set comprising 422 defective images and 400 normal images through image preprocessing of S02 by using the images acquired in S01, and manually setting two types of labels for each image: one is a mask label, and the other is a classification label, wherein the classification label corresponding to the light guide plate image containing the defect is set to be 1, and the classification label corresponding to the light guide plate image not containing the defect is set to be 0;
setting 75 defect images, 75 normal images and corresponding labels in the data set as a training set, setting the other 75 defect images, 75 normal images and corresponding labels as a verification set, and setting the rest 272 defect images, 250 normal images and corresponding labels as a test set;
s0403, network training
Training a two-stage residual error attention network of 'segmentation + decision' built through S03, wherein the batch size is 8 in the training, and an Adam optimizer is adopted in the optimizer;
first, independently training and dividing sub-network
Training by taking the training set as the input of the divided subnets, and passing the Loss function Loss corresponding to the output of the divided subnets and the mask labelssegmentationCalculating the loss of a training set of the subnet segmented by the current round, adjusting network parameters through a back propagation algorithm and a gradient descent algorithm to enable the loss of the training set to be continuously descended, then calculating the loss of a verification set on the verification set, and if the loss of the verification set in the current round is lower than the loss of the verification set in the previous round, storing a current model and model parameters of the current round, wherein the output characteristic graph of the subnet and the output characteristic graph of the segmented subnet are obtained by including characteristic graphs output by each skip connection structure; thus, the subnet is divided for 500 training rounds;
freezing and dividing the subnet, training decision subnet
In the process of training the decision subnet, parameters for segmenting the subnet cannot be updated; in the first step, the training set obtains the final feature graph output by the third skip connection structure obtained by dividing the subnet and the output feature graph of the divided subnet, and the two feature graphs are used as the output feature graph of the third skip connection structureTraining the input of the decision subnet, and passing the Loss function Loss corresponding to the output and the classification label of the decision subnetdecisionCalculating the loss of a training set of a current round decision subnet, adjusting network parameters through a back propagation algorithm and a gradient descent algorithm to enable the loss of the training set to be continuously descended, then calculating the loss of a verification set on the verification set, if the loss of the verification set in the current training round is lower than the loss of the verification set in the previous round, saving a model and model parameters of the current round, and performing 100 rounds of total training of the decision subnet;
after training is finished, a trained two-stage residual error attention network of 'segmentation + decision' is obtained;
s0404, offline test
Detecting the trained two-stage residual error attention network of 'segmentation + decision' by using a test set, and counting the accuracy of detection:
the accuracy rate is (1- (number of missed detections + number of false detections)/total number of detections) 100%,
the test result reaches 99.808%; therefore, the trained two-stage residual error attention network of 'segmentation + decision' is verified to meet the requirement of on-line actual production.
S05, using the trained two-stage residual error attention network of 'segmentation + decision' to detect defects and output the detection result
The method comprises the steps of collecting light guide plate images through the step S01, inputting the light guide plate images into an upper computer, preprocessing the images through the step S02 to obtain a group of images to be detected (the size is 1 x 224), and obtaining a corresponding mask (namely, the output of a segmentation subnet can present the outline and the position of a defect) and a decision result (namely, the output of a decision subnet is greater than 0.5 to indicate that the defect is contained and less than 0.5 to indicate that the defect is not contained) through a trained two-stage residual error attention network of segmentation and decision, namely, if a group of images to be detected have a defect, marking the corresponding position of the defect in an original image of the light guide plate, thereby meeting the requirements of a light guide plate manufacturer, namely judging whether the whole light guide plate is qualified or not and carrying out the visualization result of defect positioning on the corresponding light guide.
Comparative example 1:
the two-phase residual attention network based on "split + decision" was compared to a single classification network (assuming the existing Resnet18 network): under the same conditions of using a training set, a verification set and a test set in S0402, respectively, a two-stage residual attention network based on "segmentation + decision" and a single classification network (adopting the existing Resnet18 network) are optimally trained, the training process is consistent with S0403, and then the accuracy obtained by respectively testing on the test set is shown in table 1:
table 1 comparison of the inventive network and the Resnet18 network
Model (model) Rate of accuracy
Our network 99.808％
Resnet18 77.969％
From the above experiment, it can be found that under the condition of few training samples, a single classification network cannot obtain good detection performance, and a two-stage residual error attention network of 'segmentation + decision' can obtain good detection performance under the assistance of defect segmentation and positioning, so that the superiority of the method is reflected; moreover, a single classification network can only obtain a single classification result, and the two-stage residual error attention network of 'segmentation + decision' can obtain a decision result and the position and the contour of the defect, so that visual display of the defect is facilitated.
Comparative example 2:
changing the Residual Attention Unit (RAU) in the two-stage residual attention network structure of 'segmentation + decision' into a standard convolution layer (Conv 3) with the core size of 3 x 3, and keeping the rest unchanged, thereby constructing a contrast network without the Residual Attention Unit (RAU), which is mainly used for proving the effectiveness of the Residual Attention Unit (RAU);
respectively inputting the training set and the verification set established by the S0402 into a two-stage residual attention network based on segmentation + decision and a contrast network without a Residual Attention Unit (RAU), respectively performing optimization training on the two-stage residual attention network based on segmentation + decision and the contrast network without the Residual Attention Unit (RAU), wherein the training process is consistent with S0403, and then respectively testing on a test set, wherein four tests are shown in FIG. 5, and the first column of Image in FIG. 5 is a light guide plate Image picture with a defect; the second column group Truth of FIG. 5 is the corresponding light guide plate image after the mask label is manually placed; the third column of fig. 5, ours (none RAU), is the mask (i.e. the image of the feature to be defect) output by the comparison network without Residual Attention Unit (RAU), and the number in the upper left corner of the image is the output result of the corresponding decision sub-network, indicating the probability that the defect is true; the fourth column of Ours in fig. 5 is the result of the corresponding output of the two-stage residual attention network "segmentation + decision", which includes the output mask of the segmentation sub-network (i.e. the image of the feature to be defected), and the number in the upper left corner of the image is the output result of the corresponding decision sub-network, which indicates the probability that the defect is true; the results obtained for the final statistical "split + decision" two-stage residual attention network and the comparative network without Residual Attention Unit (RAU) are shown in table 2:
TABLE 2 comparison of RAU-containing and RAU-free network models
Model (model) Amount of ginseng Average cross-over ratio Rate of accuracy
Not containing RAU 250.358K 91.515％ 99.234％
Containing RAU 197.751K 92.315％ 99.808％
From the experiment, the accuracy of the contrast network without the Residual Attention Unit (RAU) reaches 99.234%, the network structure provided by the invention has higher accuracy reaching 99.808%, and the defect detection method of the light guide plate based on the two-stage residual attention network of 'segmentation and decision' has very high detection performance; when the parameter quantity is reduced after the Residual Attention Unit (RAU) is used, the average cross-over ratio (note: the average cross-over ratio is an index for image semantic segmentation task investigation) and the accuracy are improved, and the effectiveness of the Residual Attention Unit (RAU) is proved.
In summary, the light guide plate defect detection method based on the two-stage residual error attention network of segmentation + decision provided by the invention can achieve excellent effect on the task of light guide plate defect detection.
Finally, it is noted that the above-mentioned lists merely illustrate some specific embodiments of the invention. It is obvious that the invention is not limited to the above embodiments, but that many variations are possible. All modifications which can be derived or suggested by a person skilled in the art from the disclosure of the present invention are to be considered within the scope of the invention.

Claims (4)

1. A light guide plate defect detection method based on a 'segmentation + decision' two-stage residual error attention network is characterized by comprising the following steps:
step 1, light guide plate image acquisition: at the tail end of the light guide plate production line, carrying out light guide plate image acquisition by using a high-resolution industrial camera and transmitting the light guide plate image to an upper computer for processing;
step 2, preprocessing the light guide plate image:
in the upper computer, a black background contained in the light guide plate image acquired in the step 1 is removed by utilizing a threshold segmentation technology, and a regional image of the ROI of the light guide plate is obtained;
then cutting the regional image of the light guide plate ROI into a group of small images with the size of 1 × 224, wherein 1/10 image width overlaps between the adjacent images;
step 3, establishing and training a two-stage segmentation + decision-making residual error attention network, including a segmentation sub-network and a decision sub-network, and then storing and outputting the trained two-stage segmentation + decision-making residual error attention network;
step 4, using the trained two-stage residual error attention network of 'segmentation + decision' to detect defects, and obtaining a result:
and (3) sending a group of images with the size of 1 × 224 obtained in the step 2 into the two-stage residual attention network of segmentation + decision trained in the step 3, obtaining corresponding masks from the segmentation subnetworks, presenting the outlines and positions of the defects, obtaining decision results from the decision subnetworks, wherein the decision results are greater than 0.5 and less than 0.5, the defects are contained, and finally the corresponding masks and decision results are marked in the original image.
2. The light guide plate defect detection method based on the "segmentation + decision" two-stage residual attention network according to claim 1, wherein the step 3 of establishing, training and testing the "segmentation + decision" two-stage residual attention network specifically comprises the following steps:
step 3-1, constructing a segmentation subnet:
the segmentation sub-network adopts a U-shaped network structure and comprises an encoder and a decoder, wherein the encoder comprises a standard convolution Conv 3+ 3+ batch normalization BN + activation function layer with the kernel size of 3 × 3 and three groups of residual error attention unit RAU + Average pooling Average-Pool layers from top to bottom, and feature maps with four resolutions are output in sequence; the decoder comprises three cascaded skip connection structures from bottom to top, Feature Maps with four resolutions are fused to obtain a final Feature map Last Feature Maps, and finally the final Feature map Last Feature Maps are integrated with features through a standard convolution Conv1 x 1+ batch normalization BN + activation function layer with the kernel size of 1 x 1 to obtain an Output Feature map Segmentation Output of a segmented subnet with the size of 1 x 224;
step 3-2, constructing a decision subnet:
the decision sub-network comprises an important decision node and an auxiliary decision node; taking the output characteristic graph of the sub-network segmented in the step 3-1 as the most important information, and directly and respectively obtaining two important decision nodes through a global average pooling GAP layer and a global maximum pooling GMP layer; taking a final feature graph output by a third skip layer connection structure in a split sub-network and an output feature graph of the split sub-network as input, sequentially passing through a group of standard convolution Conv3 x 3+ maximum pooling Max-Pool layer structures with a kernel size of 3 x 3 and two groups of residual error attention units RAU + maximum pooling Max-Pool layer structures, extracting auxiliary information from the output feature graph of the split sub-network and the final feature graph output by the third skip layer connection structure, and sequentially passing through a global average pooling GAP layer and a global maximum pooling GMP layer to obtain an auxiliary decision node; finally, outputting a final Decision node Decision Output through a fully connected full connected layer, wherein the Output of the node is the probability that the defect is true;
combining the divided subnetworks and the decision subnet into a whole network which is a two-stage residual error attention network of 'division + decision';
step 3-3, training and testing the network:
1. establishing a loss function
Loss function Loss adopted for split subnet trainingsegmentationThe following formulaShown in the figure:
Losssegmentation＝0.5×Lossbce+0.5×Lossdice，
Figure FDA0002632212910000021
Figure FDA0002632212910000022
wherein H represents the height of the mask label, W represents the width of the mask label, yiFor a certain pixel value of the mask label,
Figure FDA0002632212910000023
for dividing the corresponding pixel values output by the sub-network, label is the mask label*To split subnet output, smooth is a fixed value to prevent divide by 0 errors;
loss function Loss adopted by decision sub-networkdecisionAs shown in the following equation:
Lossdecision＝ylog(y*)+(1-y)log(1-y*)，
wherein y is a classification label, y*Outputting for the decision subnet;
2. building a training data set
Collecting light guide plate images in the step 1, wherein the light guide plate images comprise 422 defective images and 400 normal images, and each image is manually provided with two labels: one is a mask label, the other is a classification label, the classification label corresponding to the light guide plate image containing the defect is set to be 1, and the classification label corresponding to the light guide plate image without the defect is set to be 0;
setting 75 defect images, 75 normal images and corresponding labels in the image set as a training set, setting the other 75 defect images, 75 normal images and corresponding labels as a verification set, and setting the rest 272 defect images, 250 normal images and corresponding labels as a test set;
3. network training
Training a two-stage residual error attention network of 'segmentation + decision' built by PyTorch according to the step 3-1 and the step 3-2, wherein the batch size is taken as 8 in the training, and an Adam optimizer is adopted in the optimizer;
first, independently training and dividing sub-network
Training by taking the training set as the input of the divided subnets, and passing the Loss function Loss corresponding to the output of the divided subnets and the mask labelssegmentationCalculating the loss of a training set of the subnet segmented by the current round, adjusting network parameters through a back propagation algorithm and a gradient descent algorithm to enable the loss of the training set to be continuously descended, then calculating the loss of a verification set on the verification set, and if the loss of the verification set in the current round is lower than the loss of the verification set in the previous round, storing a current model and model parameters of the current round, wherein the current model and model parameters comprise characteristic graphs output by the skip connection structure of each hop layer and output characteristic graphs of the segmented subnet; thus, the subnet is divided for 500 training rounds;
freezing and dividing the subnet, training decision subnet
In the process of training the decision subnet, parameters for segmenting the subnet cannot be updated; in the first step, the training set obtains the final characteristic diagram output by the third skip connection structure through the division of the subnet and obtains the output characteristic diagram of the division of the subnet, the final characteristic diagram and the output characteristic diagram are used as the input of the decision subnet for training, and the output and the classification label corresponding to the decision subnet pass through the Loss function LossdecisionCalculating the loss of a training set of a current round decision subnet, adjusting network parameters through a back propagation algorithm and a gradient descent algorithm to enable the loss of the training set to be continuously descended, then calculating the loss of a verification set on the verification set, if the loss of the verification set in the current training round is lower than the loss of the verification set in the previous round, saving a model and model parameters of the current round, and performing 100 rounds of total training of the decision subnet;
after training is finished, a trained two-stage residual error attention network of 'segmentation + decision' is obtained;
third, off-line testing
And detecting the trained two-stage residual error attention network of 'segmentation + decision' by using a test set, and counting the accuracy of detection.
3. The light guide plate defect detection method based on the "segmentation + decision" two-stage residual attention network as claimed in claim 2, wherein:
the residual attention unit RAU includes: a main branch and an attention branch, wherein the main branch adopts a depth separable convolution DSConv, and the attention branch comprises a channel attention module and a position attention module; in the channel attention module, after passing through a global average pooling GAP layer and a global maximum pooling GMP layer, the channel attention module is input into an add layer, and then, the channel attention module passes through two standard convolution Conv 1+ batch normalization BN + activation function layers with the kernel size of 1 × 1 in sequence; the position attention position attribution module respectively operates a Mean layer through averaging along the channel direction and a Max layer through maximum value computing along the channel direction, then splicing Concat operation is carried out along the channel direction, then a global convolution network GCN is used, and finally a standard convolution Conv 1+ batch normalization BN + activation function layer with the kernel size of 1 x 1 is used;
the implementation process of the residual attention unit RAU is shown as the following formula:
F(x)＝(1+A(x))×T(x)，
A(x)＝C(x)×P(x)，
where x represents input, a (x) represents output of attention branch, t (x) represents output of trunk branch, c (x) represents output of channel attention module, p (x) represents output of position attention module, and f (x) represents output of residual attention unit RAU.
4. The method according to claim 3, wherein the method comprises the following steps:
the Skip layer Connection Skip Connection structure adopts two feature maps with different resolutions as input, firstly, Up-sampling is carried out on a Low-resolution feature map to obtain a new Low-resolution feature map, then, a High-resolution feature map and the new Low-resolution feature map are spliced together through a splicing Concat operation, and finally, feature fusion is completed through a standard convolution Conv 3+ 3+ batch normalization BN + activation function layer with the kernel size of 3 to obtain the new High-resolution feature map.
CN202010814606.4A 2020-08-13 2020-08-13 Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network Pending CN112233059A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010814606.4A CN112233059A (en) 2020-08-13 2020-08-13 Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010814606.4A CN112233059A (en) 2020-08-13 2020-08-13 Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network

Publications (1)

Publication Number Publication Date
CN112233059A true CN112233059A (en) 2021-01-15

Family

ID=74115392

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010814606.4A Pending CN112233059A (en) 2020-08-13 2020-08-13 Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network

Country Status (1)

Country Link
CN (1) CN112233059A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112767369A (en) * 2021-01-25 2021-05-07 佛山科学技术学院 Defect identification and detection method and device for small hardware and computer readable storage medium
CN112950606A (en) * 2021-03-15 2021-06-11 重庆邮电大学 Mobile phone screen defect segmentation method based on small samples
CN113763384A (en) * 2021-11-10 2021-12-07 常州微亿智造科技有限公司 Defect detection method and defect detection device in industrial quality inspection
CN113781479A (en) * 2021-11-09 2021-12-10 聚时科技(江苏)有限公司 Semiconductor appearance defect quantification method based on image segmentation algorithm
CN115578339A (en) * 2022-09-30 2023-01-06 湖北工业大学 Industrial product surface defect detection and positioning method, system and equipment

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104573740A (en) * 2014-12-22 2015-04-29 山东鲁能软件技术有限公司 SVM classification model-based equipment fault diagnosing method
CN111179250A (en) * 2019-12-30 2020-05-19 北京航空航天大学 Industrial product defect detection system based on multitask learning
CN111179229A (en) * 2019-12-17 2020-05-19 中信重工机械股份有限公司 Industrial CT defect detection method based on deep learning
CN111223083A (en) * 2020-01-06 2020-06-02 宜通世纪物联网研究院(广州)有限公司 Method, system, device and medium for constructing surface scratch detection neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104573740A (en) * 2014-12-22 2015-04-29 山东鲁能软件技术有限公司 SVM classification model-based equipment fault diagnosing method
CN111179229A (en) * 2019-12-17 2020-05-19 中信重工机械股份有限公司 Industrial CT defect detection method based on deep learning
CN111179250A (en) * 2019-12-30 2020-05-19 北京航空航天大学 Industrial product defect detection system based on multitask learning
CN111223083A (en) * 2020-01-06 2020-06-02 宜通世纪物联网研究院(广州)有限公司 Method, system, device and medium for constructing surface scratch detection neural network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
DOMEN TABERNIK等: "Segmentation-Based Deep-Learning Approach for Surface-Defect Detection", 《ARXIV》, pages 3 - 5 *
徐宏伟等: "基于残差双注意力 U-Net 模型的CT图像囊肿肾脏自动分割", 《计算机应用研究》, pages 1 *
李敏;崔树芹;陈佳;: "应用视觉显著性的小提花织物疵点检测", 纺织学报, no. 12 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112767369A (en) * 2021-01-25 2021-05-07 佛山科学技术学院 Defect identification and detection method and device for small hardware and computer readable storage medium
CN112950606A (en) * 2021-03-15 2021-06-11 重庆邮电大学 Mobile phone screen defect segmentation method based on small samples
CN112950606B (en) * 2021-03-15 2023-04-07 重庆邮电大学 Mobile phone screen defect segmentation method based on small samples
CN113781479A (en) * 2021-11-09 2021-12-10 聚时科技(江苏)有限公司 Semiconductor appearance defect quantification method based on image segmentation algorithm
CN113763384A (en) * 2021-11-10 2021-12-07 常州微亿智造科技有限公司 Defect detection method and defect detection device in industrial quality inspection
CN115578339A (en) * 2022-09-30 2023-01-06 湖北工业大学 Industrial product surface defect detection and positioning method, system and equipment

Similar Documents

Publication Publication Date Title
CN112233059A (en) Light guide plate defect detection method based on segmentation and decision-making two-stage residual error attention network
CN111612763B (en) Mobile phone screen defect detection method, device and system, computer equipment and medium
CN106952250B (en) Metal plate strip surface defect detection method and device based on fast R-CNN network
CN105894036B (en) A kind of characteristics of image template matching method applied to mobile phone screen defects detection
CN112037219B (en) Metal surface defect detection method based on two-stage convolutional neural network
CN111951249A (en) Mobile phone light guide plate defect visual detection method based on multitask learning network
CN109859163A (en) A kind of LCD defect inspection method based on feature pyramid convolutional neural networks
CN114140679A (en) Defect fusion method, device, recognition system and storage medium
CN110021005A (en) Circuit board flaw screening technique and its device and computer-readable recording medium
CN110544231B (en) Lithium battery electrode surface defect detection method based on background standardization and centralized compensation algorithm
CN106960208A (en) A kind of instrument liquid crystal digital automatic segmentation and the method and system of identification
CN115601355A (en) Method and device for detecting and classifying product surface defects and storage medium
CN113421230A (en) Vehicle-mounted liquid crystal display light guide plate defect visual detection method based on target detection network
CN110533660B (en) Method for detecting silk-screen defect of electronic product shell
CN112396575A (en) Image-based classification system
CN116990323A (en) High-precision printing plate visual detection system
CN115937517A (en) Photovoltaic fault automatic detection method based on multispectral fusion
CN112508857A (en) Aluminum product surface defect detection method based on improved Cascade R-CNN
CN114417993A (en) Scratch detection method based on deep convolutional neural network and image segmentation
CN111415330A (en) Copper foil appearance defect detection method based on deep learning
CN113393438A (en) Resin lens defect detection method based on convolutional neural network
CN115866502A (en) Microphone part surface defect online detection process
CN113327243B (en) PAD light guide plate defect visual detection method based on Ayolov3-Tiny new framework
CN116843615B (en) Lead frame intelligent total inspection method based on flexible light path
CN113469984B (en) Method for detecting appearance of display panel based on YOLO structure

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