CN114463269A - Chip defect detection method based on deep learning method - Google Patents
Chip defect detection method based on deep learning method Download PDFInfo
- Publication number
- CN114463269A CN114463269A CN202111647594.1A CN202111647594A CN114463269A CN 114463269 A CN114463269 A CN 114463269A CN 202111647594 A CN202111647594 A CN 202111647594A CN 114463269 A CN114463269 A CN 114463269A
- Authority
- CN
- China
- Prior art keywords
- network
- chip
- deep learning
- defect detection
- loss function
- 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.)
- Withdrawn
Links
- 230000007547 defect Effects 0.000 title claims abstract description 99
- 238000001514 detection method Methods 0.000 title claims abstract description 78
- 238000013135 deep learning Methods 0.000 title claims abstract description 38
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 11
- 238000004519 manufacturing process Methods 0.000 claims abstract description 3
- 230000006870 function Effects 0.000 claims description 56
- 238000012549 training Methods 0.000 claims description 34
- 238000013138 pruning Methods 0.000 claims description 17
- 238000012360 testing method Methods 0.000 claims description 14
- 238000010206 sensitivity analysis Methods 0.000 claims description 7
- 238000013145 classification model Methods 0.000 claims description 4
- 238000009499 grossing Methods 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 2
- 238000013459 approach Methods 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 230000001537 neural effect Effects 0.000 claims description 2
- 230000008569 process Effects 0.000 claims description 2
- 238000010187 selection method Methods 0.000 claims description 2
- 230000003213 activating effect Effects 0.000 claims 1
- 238000013461 design Methods 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000005034 decoration Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 210000003128 head Anatomy 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000005856 abnormality Effects 0.000 description 1
- 206010000496 acne Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 208000003464 asthenopia Diseases 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000004566 building material Substances 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
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
- G06T7/0006—Industrial image inspection using a design-rule based approach
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30148—Semiconductor; IC; Wafer
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a chip defect detection method based on a deep learning method, which comprises the following steps: 1) marking the chip defect pictures according to the actually applied industrial chip defect classification, and respectively manufacturing a chip defect classification data set and a chip defect positioning detection data set; 2) respectively carrying out classification detection on whether the chip has defects or not based on a VGG16 network and a ResNet50 network; 3) positioning and detecting the chip defects based on a deep learning network; 4) the invention adopts a VGG16 network and a ResNet50 network to classify whether a chip has defects or not respectively, and adopts a deep learning algorithm to detect a chip defect target, thereby improving the defect detection efficiency and quality of an industrial chip.
Description
Technical Field
The invention belongs to the fields of computer artificial intelligence, deep neural networks and target identification technologies, and particularly relates to a chip defect detection method based on a deep learning method.
Background
According to the traditional artificial chip defect detection method, the human eyes are used for observing a workpiece photo, defects are found, and then the workpiece is screened, the mode is influenced by continuous eye fatigue, human eye resolution and subjective factors, the efficiency is low, and wrong detection and omission often occur, so that the quality rate of products is influenced, and the requirement of high-precision defect detection is difficult to adapt.
According to the traditional machine learning defect detection method, according to the manual design characteristics of defects, such as geometric characteristics, color characteristics, texture characteristics, projection characteristics and the like, a shallow feature classifier, such as an SVM (support Vector machine), KNN (k-Nearest Neighbor), a decision tree and the like, is used for defect classification. The method depends on manual design characteristics, has weak expression capability, low flexibility and weak adaptability to new tasks, and the traditional machine vision method is difficult to meet the industrial defect detection requirements on high flexibility and high accuracy.
Therefore, it is desirable to provide a chip defect detection method based on deep learning, which can implement an end-to-end detection method and has high flexibility and high accuracy.
Disclosure of Invention
The invention aims to provide a defect positioning detection method based on deep learning to solve the existing technical problems. The method can replace the traditional artificial chip defect detection method and the traditional machine learning defect detection method, has high precision and flexibility, strong network expression capability, does not need manual design characteristics, and is easy to apply and transfer.
Based on the purpose, the invention adopts the following technical scheme: a chip defect detection method based on a deep learning method comprises the following steps:
step 1), marking the chip defect pictures according to the actually applied industrial chip defect classification, and respectively manufacturing a chip defect classification data set and a chip defect positioning detection data set. The chip defect classification data is concentrated, and the classification is divided into two types according to whether the chip defects exist in the pictures: defect free, labeled smile; defective, marked cry.
The chip defect positioning detection data are concentrated and are divided into eight types according to the types of industrial defects: friction defects, scratch defects, edge chipping defects, pimple defects, material defects, grid defects or abnormalities, dirt defects, foreign matter defects. According to the following steps of 6: 2: 2, dividing the pictures with the labels into a training set, a verification set and a test set.
And step 2), respectively carrying out classification detection on whether the chip has defects or not based on the VGG16 network and the ResNet50 network. Due to the fact that the number of detection types is small, the network is appropriately simplified, the number of neurons in a full connection layer is reduced, and the neurons are randomly discarded by adopting a Dropout algorithm after the full connection layer.
The network output is processed by a Softmax function, and the loss function is a sum-variance (SSE) loss function. After training on the training set is completed, testing the trained model on the training set and the testing set respectively to obtain the classification accuracy of the VGG16 model and the ResNet50 model on the training set and the testing set.
And 3) carrying out positioning detection on the chip defects based on the deep learning network. The method of introducing label smoothing increases the noise of the sample, thereby enhancing the generalization capability of the network.
The CSPNet network idea is applied to the deep learning network, the FPN network is improved, and the targets with different sizes are detected by adopting multiple sizes, so that the detection of the targets with larger size difference in the complex multi-target image is facilitated. And optimizing the model by adopting IoU, GIoU, CIoU, Gaussian loss function and focal loss function, and improving the model precision.
And 4) respectively adopting two modes of fixed cutting rate pruning and sensitivity analysis-based pruning to carry out channel pruning on the model in the step 3. When calculating sensitivity information of the convolutional layer, the final increase condition of the loss function under the corresponding clipping rate is calculated respectively by setting a group of channel number proportions sequentially clipped by the network to be clipped.
After the clipping rate is determined, the network is clipped, and unimportant convolution kernels are removed, so that the network parameter quantity and the operation quantity are reduced.
The invention has the beneficial effects that: the invention adopts the VGG16 network and the ResNet50 network to classify whether the chip has defects or not respectively, and adopts a deep learning algorithm to detect the defect target of the chip, thereby improving the defect detection efficiency and quality of the industrial chip; the method can replace the traditional artificial chip defect detection method and the traditional machine learning defect detection method, has high precision and flexibility, strong network expression capability, does not need manual design characteristics, and is easy to apply and transfer.
Drawings
FIG. 1 is a schematic diagram of chip defect types.
Fig. 2 is a graph of accuracy of classification detection on training and test sets based on VGG16 networks.
Fig. 3 is a graph of accuracy of classification detection on training and test sets based on ResNet 50.
Fig. 4 is a schematic diagram of a deep learning network structure.
FIG. 5 is a diagram illustrating a chip defect detection result.
Fig. 6 is a PR graph of different defect types.
Fig. 7 is a graph of the head Loss as a function of cut-out rate.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples.
The chip defect detection method based on the deep learning method comprises the following steps:
(1) the construction of the chip defect classification detection data set and the chip defect positioning detection data set specifically comprises the following steps:
(1.1) selecting 1000 chip pictures, wherein 600 chip pictures contain defects and are marked as cry; 400 pictures contain no defects and are marked smile. Disorganizing 1000 pictures according to the following ratio of 6: 2: the 2-scale division is into a training set, a validation set and a test set.
(1.2) selecting at least one picture reporting a defect from the industrially shot chip pictures 7940, and marking the type (classes) of the defect and a bounding box (bounding box) of the position of the defect by using a labelImg tool, wherein 2312 pictures are used in total. Randomly ordering the markup files, and according to the following steps of 6: 2: and 2, dividing a training set, a verification set and a test set according to the proportion. The type, number and description of the defect labels are shown in Table 1.
TABLE 1 chip Defect labeling Categories and quantities
(2) The chip defect classification detection based on the VGG16 and ResNet50 networks specifically comprises the following steps:
and (2.1) adopting a VGG16 network to establish a defect detection classification model. The embodiment of the invention has less detection types, moderately reduces the network, reduces the number of the fully connected neurons from 4096 of the original network to 512, and randomly discards the neurons by adopting a Dropout algorithm after full connection, wherein the discarding probability is 0.5. And processing the network output result by adopting a softmax function, wherein the loss function adopts a sum-variance (SSE) loss function, as shown in formula (1).
The resolution of the input image is 608X608, the image enhancement parameters are saturation adjusting range +/-1, exposure adjusting range +/-1 and angle adjusting range +/-7, and mirror image overturning on an X axis is adopted randomly. 128 pictures were randomly drawn for each training. The training method adopts a momentum gradient descent method, the momentum parameter is 0.9, a weight attenuation method is adopted for regularization, and the regularization coefficient is 0.0005. The training is carried out for 8000 turns, and the training is carried out by adopting a fixed learning rate of 0.001 due to a simpler network structure.
The accuracy of the classification detection based on VGG16 in the training set and the test set is shown in FIG. 2, the accuracy of the model in the training set is up to 97%, and the accuracy in the test set is up to 95.5%.
(2.2) a ResNet50 network is used to build a defect detection classification model. The number of convolution kernels of the previous layer of the output layer is reduced from 4096 of the original network to 1024, and the training parameters are consistent with the VGG16 classification detection.
The accuracy of the classification detection based on ResNet50 in the training set and test set is shown in FIG. 3. The classification detection model based on ResNet50 has better effect than that based on VGG16, can be converged only by training 100 turns, and the accuracy of the chip defect classification detection based on ResNet50 network reaches 97.88% in the training set and 96.5% in the test set.
(3) The chip positioning defect detection based on the deep learning algorithm specifically comprises the following steps:
(3.1) the deep learning algorithm first divides the picture into small regions with 13 x 13 grids, each with a size of 32 x 32 (taking 416x416 of the input picture as an example). Each small grid is responsible for predicting the category of the object with the central point in the grid and the position and size of the prediction box, as shown in fig. 4.
(3.2) the Darknet53 network is used in the feature extraction backbone network, the characteristics of ResNet and VGGnet are combined, the difficulty of deep network training is reduced by adopting a residual network structure, the feature extraction is carried out by adopting stacked small convolution kernels, and the parameter quantity of the network is reduced by adopting a convolution kernel of 1 multiplied by 1. And (3) adopting a 3 × 3 convolution kernel with the step size of 2 and the filling of 1 to zoom the image, wherein the output size is reduced by half after each convolution, and the output size is reduced by 5 times in total.
And (3.3) applying a Label smoothing (Label smoothing) method to add noise on the class labels to play a role of regularization. Traditional class labels are in the form of unique hot codes, making the network too confident and resulting in an overfitting. And the label is smooth and effective to soften the code, so that the predicted value of the network is not excessively concentrated on the category with higher probability any more, and the category with lower probability is considered.
And (3.4) the convolution operation adopts batch normalization to reduce the influence of data offset on training in forward propagation, and adopts a Leaky Relu activation function to activate the output of a neural unit.
And (3.5) detecting the targets with different sizes by adopting a characteristic pyramid FPN algorithm and adopting multiple scales in the target detection network, and enhancing the detection of small targets.
And (3.6) establishing a loss function according to the output result of the network prediction and the real output result, and continuously updating parameters through back propagation to enable the output result of the network prediction to approach the real output result. And establishing three types of loss functions for each prediction frame, namely a loss function for representing whether the target object is contained, a loss function for representing the target position and a loss function for representing the target class, wherein the total loss function is shown as a formula (2).
Loss＝Lossobj+Lossxy+Losswh+Lossclass (2)
The loss function of whether or not to include the target object uses a binary cross entropy loss function, as shown in equation (3), where
Is the true value of the label, and the value may be 0, 1, -1, PobjThe probability value of the real number output by the network after being processed by the sigmoid function is obtained.
The loss function of the target detection box center point value uses a binary cross entropy loss function, as shown in formula (4). Wherein the summation part represents the situation that the target object is in the j prediction box predicted in the ith grid and the target of regression
And
are all located in (0, 1), meaning that the center point can only be located in the grid.
The loss function of the target detection frame size directly adopts an absolute value loss function, as shown in formula (5).
The loss function of the target class adopts a binary cross entropy loss function, as shown in formula (6).
(3.7) comparing the predicted result with the labeled real box, and classifying the predicted result into the following types of TP (true Positive) to indicate that the prediction is positive and the classification is correct according to whether the predicted classification is correct and IoU reaches a threshold value, wherein IoU of the predicted box and the real box is smaller than the threshold value. Fp (false positive) indicates a prediction as positive sample but classification error. Fn (falsenegtive) indicates that the prediction is negative but actually positive. The accuracy (precision) and recall (recall) of each type are calculated, where accuracy refers to the proportion of the prediction that is classified correctly in the positive sample, as shown in equation (7). The recall ratio is the ratio of the predicted correct in all labeled boxes, as shown in equation (8).
And comprehensively considering the accuracy and the recall rate, a PR curve is required to be made according to the accuracy and the recall rate, and the average accuracy AP is obtained by calculating the area below the PR curve. For the detection of multiple categories, the average of the APs of each category is calculated to obtain the overall detection accuracy mapp (meanaverageprecision), as shown in formula (9).
(3.8) optimizing the loss function to improve network accuracy: the model is optimized using different loss functions: IoU, GIoU, CIoU, Gaussian loss function and focal loss function, which all improve the detection precision of the model, and the influence of different loss functions on the detection performance is shown in Table 2. Network mAP of IoU, GIoU and CIoU as loss functions reaches 82.71%, 82.75% and 82.53%, and is respectively increased by 1.2%, 1.6% and 1.02%, uncertainty of a position is introduced by adopting a Gaussian loss function, so that the uncertainty of the position can be predicted on the premise that the network basically does not change a deep learning network structure and calculated amount, and the detected mAP reaches 83.74% and is increased by 2.23%. The problem of unbalance of positive and negative samples is solved by adopting focal loss, the detected mAP reaches 84.09%, and the percentage is improved by 2.58.
TABLE 2 Effect of different loss functions on detection Performance
| Model | Size | Loss | Map(％) | Recall |
| Darknet53 | 416x416 | BaseLine | 81.51 | 84 |
| Darknet53 | 416x416 | IoU | 82.71 | 83 |
| Darknet53 | 416x416 | GIoU | 82.75 | 84 |
| Darknet53 | 416x416 | DIoU | 76.51 | 83 |
| Darknet53 | 416x416 | CIoU | 82.53 | 83 |
| Darknet53 | 416x416 | Gaussian | 83.74 | 84 |
| Darknet53 | 416x416 | Focal loss | 84.09 | 85 |
(3.9) network training is carried out under a GeForceGTX1080 hardware platform, and the training is carried out for 6500 rounds: the resolution of an input image is 416x416, the image enhancement parameters are a saturation adjusting range +/-1.5, an exposure adjusting range +/-1.5 and a hue adjusting range +/-1.0, and 64 pictures are randomly extracted in each training. In the training process, a training method of momentum gradient descent is adopted, the momentum parameter is set to be 0.9, a weight attenuation method is adopted to regularize the network, and the regularization coefficient is set to be 0.0005. Training was performed for a total of 6500 rounds, with the first 1000 rounds using warm-up (warmup) operation and the learning rate slowly increased. Then the learning rate is reduced to 0.2, 0.5 and 0.1 times in 3000, 4000 and 5500 rounds respectively by adopting a step attenuation method. In order to improve the adaptability to the defect size detection, a multi-scale training method is adopted in the training, and a group of resolutions in [320, 352, 384, 416, 448, 480, 512, 544 and 608] are randomly selected to perform the training in every 10 rounds.
(3.10) the chip defect detection method based on deep learning is shown in fig. 5, and it can be seen that the method can identify the defects more accurately and can also identify the defects under the condition of dense defects. A PR curve is made to calculate the accuracy of each defect type as shown in fig. 6. The APs for the 8 defect types were 86.83%, 86.41%, 63.61%, 84.04%, 96.76%, 72.64%, 67.71%, 94.05%, respectively, with a total accuracy of the maps of 81.5%.
And (3.11) considering the parameter quantity of the neural network, the total operation times and the detection speed of the network in the aspect of real-time performance. The detection speed of the network is usually expressed in frames of detected images per second, i.e. FPS. In actual work, 30FPS is generally used as a critical value, and the detection speed is larger than 30FPS, so that the requirement on detection instantaneity can be met. The parameter size of the chip defect detection method based on deep learning is 235.04MB, and the total operation amount is 65.355 BFLOPS. The time required for testing one picture under a GeForceGTX1080Ti platform is 14.998ms, namely 68FPS, and the requirement of real-time detection is met.
(4) Pruning the model by adopting two modes of fixing the cutting rate and selecting the building material rate based on sensitivity analysis, and specifically comprises the following steps:
(4.1) pruning the three output branches of the deep learning algorithm with empirical clipping rates of 0.5, 0.7, 0.8, respectively.
And (4.2) carrying out sensitivity analysis on the clipping rate of the output branch, setting a group of channels to be clipped by the network to be clipped, and respectively calculating the final increase condition of the loss function under the corresponding clipping rate. The increase condition of the loss function of the three detection heads of the deep learning network along with the change of the clipping rate is shown in fig. 7, and a group of proper clipping rates are calculated according to the sensitivity and the precision loss threshold value and are used for clipping the model. As the clipping rate increases, the number of channels to be clipped increases, at which time the expressive power of the network becomes weaker and the loss function for model prediction increases. The change is relatively gradual when the loss function increases below 0.03 and increases relatively quickly above 0.03. The clipping rate of each convolutional layer was determined by selecting 0.03 as the loss function threshold, and the calculation results are shown in table 3.
TABLE 3 pruning Rate selection for different clipping strategies
And (4.3) pruning the deep learning network model channel according to the selection method of the two cutting rates, and obtaining the table 4 through comparison. The pruning method with a fixed cutting rate is recorded as R578, and the sensitivity analysis method is recorded as Sensitive. After model cutting, the network parameter quantity and the operation quantity are greatly reduced, the parameter quantity of the R578 and the parameter quantity of the sensitive method are respectively reduced by 26.45 percent and 31.16 percent, the mAP is respectively improved by 3.74 percent and 3.27 percent, and the performance of the network is improved instead of being reduced by model cutting. The network pruning rate obtained through sensitivity analysis can realize greater pruning under the same precision loss.
TABLE 4 deep learning network model compression effect comparison
It should be noted that the above-mentioned contents only illustrate the technical idea of the present invention, and the protection scope of the present invention is not limited thereby, and it is obvious to those skilled in the art that several modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations fall within the protection scope of the claims of the present invention.
Claims (10)
1. A chip defect detection method based on a deep learning method is characterized by comprising the following steps:
1) marking the chip defect pictures according to the actually applied industrial chip defect classification, and respectively manufacturing a chip defect classification data set and a chip defect positioning detection data set;
2) respectively carrying out classification detection on whether the chip has defects or not based on a VGG16 network and a ResNet50 network;
3) positioning and detecting the chip defects based on a deep learning network;
4) and respectively carrying out channel pruning on the deep learning network model by adopting a fixed cutting rate pruning mode and a sensitivity analysis-based pruning mode.
2. The chip defect detection method based on the deep learning method as claimed in claim 1, wherein the chip positioning defect detection based on the deep learning algorithm in step 3) comprises the following steps:
1.1) dividing a chip defect picture into small areas, then using a Darknet53 network as a feature extraction backbone network, adopting a residual network structure to reduce the difficulty of deep network training, adopting stacked small convolution kernels to extract features, and adopting a 1 multiplied by 1 convolution kernel to reduce the parameter number of the network;
1.2) applying a label smoothing method, adding noise on a class label, and activating the output of a neural unit by adopting a Leaky Relu activation function;
1.3) the target detection network adopts a feature pyramid FPN algorithm and adopts multiple scales to detect targets with different sizes;
1.4) establishing a loss function according to the output result of the network prediction and the real output result, and continuously updating parameters through back propagation to enable the output result of the network prediction to approach the real output result;
1.5) comparing the predicted result with the labeled real frame, and optimizing the model by adopting different loss functions according to whether the predicted classification is correct and whether the loss function reaches a threshold value.
3. The chip defect detection method based on the deep learning method as claimed in claim 2, wherein in step 1.4), three types of loss functions are established for each prediction box, wherein the loss functions are a loss function for representing whether a target object is included, a loss function for representing a target position and a loss function for representing a target category.
4. The chip defect detection method based on the deep learning method as claimed in claim 1, wherein in step 4), the three output branches of the deep learning algorithm are pruned by respectively using empirical pruning rate, sensitivity analysis is performed on the pruning rate of the output branches, a group of channels to be pruned in sequence by the network to be pruned is set, and the final loss function growth condition under the corresponding pruning rate is respectively calculated.
5. The chip defect detection method based on the deep learning method as claimed in claim 4, wherein in step 4), a set of suitable clipping rates is calculated according to the sensitivity and the precision loss threshold value for clipping the model, and as the clipping rate increases, the number of clipped channels increases, the expressive power of the network becomes weaker, and the loss function for model prediction increases.
6. The chip defect detection method based on the deep learning method as claimed in claim 4, wherein in the step 4), the deep learning network module channel is pruned according to the selection method of two pruning rates.
7. The chip defect detection method based on the deep learning method as claimed in claim 1, wherein in step 1), the chip defects are classified into two categories according to whether the chip defects exist in the pictures, the chip defect positioning detection data set is classified into eight categories according to the industrial defect categories, and the pictures with labels are divided into a training set, a verification set and a test set.
8. The chip defect detection method based on the deep learning method as claimed in claim 1, wherein in step 2), a VGG16 network is used to establish a defect detection classification model, a momentum gradient descent method is used in the training method, a weight attenuation method is used to conduct regularization, a softmax function is used to process the network output result, and a sum variance loss function is used as the loss function.
9. The chip defect detection method based on the deep learning method as claimed in claim 2, wherein in step 1.5), the different loss functions include IoU, GIoU, CIoU, Gaussian loss function and focal loss function.
10. The chip defect detection method based on the deep learning method as claimed in claim 8, wherein in step 2), a ResNet50 network is used to establish a defect detection classification model, so as to reduce the number of convolution kernels in the previous layer of the output layer, and the training parameters are consistent with those in VGG16 classification detection.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111647594.1A CN114463269A (en) | 2021-12-29 | 2021-12-29 | Chip defect detection method based on deep learning method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111647594.1A CN114463269A (en) | 2021-12-29 | 2021-12-29 | Chip defect detection method based on deep learning method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114463269A true CN114463269A (en) | 2022-05-10 |
Family
ID=81408022
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111647594.1A Withdrawn CN114463269A (en) | 2021-12-29 | 2021-12-29 | Chip defect detection method based on deep learning method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114463269A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115965816A (en) * | 2023-01-05 | 2023-04-14 | 无锡职业技术学院 | Glass defect classification and detection method and system based on deep learning |
-
2021
- 2021-12-29 CN CN202111647594.1A patent/CN114463269A/en not_active Withdrawn
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115965816A (en) * | 2023-01-05 | 2023-04-14 | 无锡职业技术学院 | Glass defect classification and detection method and system based on deep learning |
| CN115965816B (en) * | 2023-01-05 | 2023-08-22 | 无锡职业技术学院 | Glass defect classification and detection method and system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhao et al. | Cloud shape classification system based on multi-channel cnn and improved fdm | |
| CN108345911B (en) | Steel plate surface defect detection method based on convolutional neural network multi-stage characteristics | |
| Al-Hiary et al. | Fast and accurate detection and classification of plant diseases | |
| CN110135503B (en) | Deep learning identification method for parts of assembly robot | |
| CN110032925B (en) | Gesture image segmentation and recognition method based on improved capsule network and algorithm | |
| CN112598713A (en) | Offshore submarine fish detection and tracking statistical method based on deep learning | |
| CN113111979B (en) | Model training method, image detection method and detection device | |
| Thakur et al. | An innovative approach for fruit ripeness classification | |
| CN112949704B (en) | Tobacco leaf maturity state identification method and device based on image analysis | |
| Travieso et al. | Pollen classification based on contour features | |
| CN105046202B (en) | Adaptive recognition of face lighting process method | |
| CN113222959B (en) | Fresh jujube wormhole detection method based on hyperspectral image convolutional neural network | |
| CN109815923B (en) | Needle mushroom head sorting and identifying method based on LBP (local binary pattern) features and deep learning | |
| CN111340019A (en) | Grain bin pest detection method based on Faster R-CNN | |
| CN108345835B (en) | Target identification method based on compound eye imitation perception | |
| CN113657294A (en) | Crop disease and insect pest detection method and system based on computer vision | |
| Lee | A Study on Classification and Detection of Small Moths Using CNN Model. | |
| CN114463269A (en) | Chip defect detection method based on deep learning method | |
| CN114663769A (en) | Fruit identification method based on YOLO v5 | |
| CN111178405A (en) | Similar object identification method fusing multiple neural networks | |
| Ni et al. | Convolution neural network based automatic corn kernel qualification | |
| CN117315380A (en) | Deep learning-based pneumonia CT image classification method and system | |
| CN113052091A (en) | Action recognition method based on convolutional neural network | |
| CN116703932A (en) | CBAM-HRNet model wheat spike grain segmentation and counting method based on convolution attention mechanism | |
| CN112507327B (en) | Weather detection and processing method based on machine learning |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20220510 |
|
| WW01 | Invention patent application withdrawn after publication |