CN115861281A - Anchor-frame-free surface defect detection method based on multi-scale features - Google Patents
Anchor-frame-free surface defect detection method based on multi-scale features Download PDFInfo
- Publication number
- CN115861281A CN115861281A CN202211686786.8A CN202211686786A CN115861281A CN 115861281 A CN115861281 A CN 115861281A CN 202211686786 A CN202211686786 A CN 202211686786A CN 115861281 A CN115861281 A CN 115861281A
- Authority
- CN
- China
- Prior art keywords
- frame
- surface defect
- stage
- defect detection
- fcos
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an anchor frame-free surface defect detection method based on multi-scale features, which comprises the following steps: establishing a surface defect detection model which comprises a Swin transform model, a first feature extraction unit, an N-layer feature pyramid network, a second feature extraction unit and N FCOS detection heads; acquiring part pictures as a data set to train a surface defect detection model; inputting the picture of the part to be detected into the trained surface defect detection model, predicting the center point of the target frame and the distances from the center point to the four edges of the real frame, and calculating the upper left corner coordinate and the lower right corner coordinate of the predicted frame as a surface defect detection result. The method can realize high-efficiency and multi-scale target surface defect detection, not only can balance positive and negative samples in a target frame, but also can better represent targets with larger shape change, and particularly has better detection precision and robustness for small-size defect targets and strip-shaped defect targets with extreme transverse-longitudinal ratios.
Description
Technical Field
The invention belongs to the technical field of target detection, and particularly relates to an anchor frame-free surface defect detection method based on multi-scale features.
Background
The rapid development of the manufacturing industry cannot leave a high-quality and digital new ecological mode, and in the production and manufacturing process of high and new industries such as integrated circuits and the like, the surface defects of the high and new industries can seriously affect the performance of electronic products, so that potential safety hazards are caused. Therefore, the defect detection is taken as a key step in industrial production and quality control, and the production quality of products can be effectively ensured by timely detecting and identifying the industrial defects.
With the development of deep learning, more and more defect detection methods are introduced for defect detection, which are generally classified into two types: one is based on anchor frames and the other is based on anchor-free frames. Due to the fact that the introduction of the Anchor frame often brings some adverse effects, for example, in the existing implementation based on an Anchor-based model, a group of target frames need to be preset for a detection network model based on priori knowledge, then the model outputs target frame fine tuning parameters, and finally the target frames output by the model are calculated through the fine tuning parameters and the target frame preset parameters, so that the scale and the shape of the preset target frames directly influence the model training effect, and the model design depends on the knowledge of a designer on the priori knowledge, the detection scale is too rigid, and the generalization capability is limited. In addition, imbalance of positive and negative samples is generated, more hyper-parameters are introduced, and the design difficulty and the calculation amount are greatly increased. Which greatly limits the accuracy and robustness of detection problems such as frequent occurrences of linear scale defects on the surface of microelectronic device Integrated Circuits (ICs). Therefore, in consideration of the limitations of the traditional model based on the anchor frame method, such as low detection speed, limited detection scale, difficult parameter adjustment and the like, the invention provides the anchor frame-free surface defect detection method based on the multi-scale features.
Disclosure of Invention
The invention aims to solve the problems, provides an anchor frame-free surface defect detection method with multi-scale characteristics, can realize high-efficiency and multi-scale target surface defect detection, not only balances positive and negative samples in a target frame, but also can better represent a target with larger shape change, and particularly has better detection precision and robustness for small-size defect targets and strip-shaped defect targets with extreme transverse-longitudinal ratios.
In order to realize the purpose, the technical scheme adopted by the invention is as follows:
the invention provides an anchor frame surface defect-free detection method based on multi-scale features, which comprises the following steps:
s1, establishing a surface defect detection model, wherein the surface defect detection model comprises a Swin transform model, a first feature extraction unit, an N-layer feature pyramid network, a second feature extraction unit and N FCOS detection heads, and the method comprises the following steps:
the first feature extraction unit comprises a plurality of parallel first convolution layers, each stage of the Swin transform model is respectively connected with the first N-1 layers of feature pyramid networks in a one-to-one corresponding sequence through the first convolution layers, the second feature extraction unit comprises a plurality of parallel second convolution layers, the feature pyramid networks are connected with the FCOS detection head in a one-to-one corresponding mode through the second convolution layers, the feature pyramid network corresponding to the last stage of the Swin transform model is taken as a reference layer, input feature graphs of the corresponding feature pyramid networks are fused from top to bottom from the reference layer through gradual up-sampling operation, and output feature graphs of the reference layer are input into the feature pyramid network of the Nth layer after passing through the third convolution layer;
the FCOS detection head comprises a classification branch and a regression branch, the classification branch comprises a category prediction branch and a center-less branch, regression and classification of the output characteristic diagram corresponding to the second convolution layer are carried out through the FCOS detection heads, and the FCOS detection heads update the target frame based on an FCOS center sampling optimization method;
s2, collecting part pictures as a data set to train a surface defect detection model;
s3, inputting the picture of the part to be detected into the trained surface defect detection model, and predicting the central point (c) of the target frame x ,c y ) And the distances (l) from the center point to the four sides of the real frame * ,t * ,r * ,b * ) Calculating the coordinates (c) of the upper left corner of the prediction box x +l * ,c y +t * ) And the coordinates of the lower right corner (c) x +r * ,c y +b * ) As a result of surface defect detection, wherein l * Is the center point (c) of the target frame x ,c y ) Distance to the left border of the real frame, t * Is the center point (c) of the target frame x ,c y ) Distance to the upper border of the real frame, r * Is the center point (c) of the target frame x ,c y ) Distance to the right frame of the real frame, b * Is the center point (c) of the target frame x ,c y ) Distance to the lower border of the real border.
Preferably, the FCOS-centric sampling optimization method specifically includes the following:
taking out a circle which is smaller than the target frame and is externally connected with a sub-rectangle by taking the central point of the target frame as the circle center, updating the target frame into the sub-rectangle, regarding the sampling points which fall into the sub-rectangle as positive samples, regarding other sampling points as negative samples, and regarding the position parameter of the sub-rectangle as (c) x -rs,c y -rs,c x +rs,c y + rs), wherein (c) x ,c y ) And the coordinate of the central point of the target frame is s, the downsampling multiplying power of the input feature map corresponding to the FCOS detection head is s, and r is a hyper-parameter.
Preferably, the Swin Transformer model comprises a batch Partition operation module, a first stage, a second stage, a third stage and a fourth stage which are connected in sequence, wherein the first stage comprises a Linear Embedding operation module and a plurality of Swin Transformer modules which are connected in sequence, and the second stage, the third stage and the fourth stage comprise a batch Merging operation module and a plurality of Swin Transformer modules which are connected in sequence.
Preferably, the number of Swin transducer modules in the first, second and fourth stages is 2, and the number of Swin transducer modules in the third stage is 6.
Preferably, the convolution kernel size of the first convolution layer is 1 × 1 and the channel is 256-dimensional.
Preferably, the convolution kernel size of the second convolution layer is 3 × 3.
Preferably, the convolution kernel size of the third convolution layer is 3 × 3 with a step size of 2.
Compared with the prior art, the invention has the following beneficial effects:
the method is characterized in that a Swin transform model based on a vision field self-attention mechanism and a characteristic pyramid network (FPN model) are fused to encode, fuse and express image characteristics, an anchor frame-free pixel-by-pixel regression coding mode FCOS is used for redefining a target detection head to obtain a result of a target detection downstream task, a central sampling strategy is adopted to optimize the FCOS detection head, efficient and multi-scale target surface defect detection is achieved, positive and negative samples in a target frame are balanced, targets with large shape changes can be well expressed, particularly, good detection accuracy and robustness are achieved for small-size defect targets and strip defect targets with extreme transverse-longitudinal ratios, and the method can be applied to detection in the fields of integrated circuit chips, industrial parts and the like.
Drawings
FIG. 1 is a flow chart of an anchor frame-free surface defect detection method based on multi-scale features according to the present invention;
FIG. 2 is a schematic structural diagram of a surface defect inspection model according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only some embodiments of the present application, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
It will be understood that when an element is referred to as being "connected" to another element, it can be directly connected to the other element or intervening elements may also be present. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the present application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.
As shown in fig. 1-2, an anchor-frame-free surface defect detection method based on multi-scale features includes the following steps:
s1, establishing a surface defect detection model, wherein the surface defect detection model comprises a Swin transform model, a first feature extraction unit, an N-layer feature pyramid network, a second feature extraction unit and N FCOS detection heads, and the method comprises the following steps:
the first feature extraction unit comprises a plurality of parallel first convolution layers, each stage of the Swin transform model is respectively connected with the first N-1 layers of feature pyramid networks in a one-to-one corresponding sequence through the first convolution layers, the second feature extraction unit comprises a plurality of parallel second convolution layers, the feature pyramid networks are connected with the FCOS detection head in a one-to-one corresponding mode through the second convolution layers, the feature pyramid network corresponding to the last stage of the Swin transform model is taken as a reference layer, input feature graphs of the corresponding feature pyramid networks are fused from top to bottom from the reference layer through gradual up-sampling operation, and output feature graphs of the reference layer are input into the feature pyramid network of the Nth layer after passing through the third convolution layer;
the FCOS detection heads comprise classification branches and regression branches, the classification branches comprise class prediction branches and center-less branches, regression and classification of output feature graphs corresponding to the second convolution layers are carried out through the FCOS detection heads, and the FCOS detection heads update the target frames based on an FCOS center sampling optimization method.
In an embodiment, the FCOS-centric sampling optimization method specifically includes the following steps:
taking out a circle which is smaller than the target frame and is externally connected with a sub-rectangle by taking the central point of the target frame as the circle center, updating the target frame into the sub-rectangle, regarding the sampling points which fall into the sub-rectangle as positive samples, regarding other sampling points as negative samples, and regarding the position parameter of the sub-rectangle as (c) x -rs,c y -rs,c x +rs,c y + rs) wherein (c) x ,c y ) The coordinate of the center point of the target frame is shown, s is the down-sampling multiplying factor of the input feature map corresponding to the FCOS detection head, and the lower sampling multiplying factor stride is 2,r is an over-parameter, such as r =1.5.
In one embodiment, the switch transform model comprises a batch Partition operation module, a first stage, a second stage, a third stage and a fourth stage which are connected in sequence, wherein the first stage comprises a Linear Embedding operation module and a plurality of switch transform modules which are connected in sequence, and the second stage, the third stage and the fourth stage comprise a batch gathering operation module and a plurality of switch transform modules which are connected in sequence.
In an embodiment, the number of Swin transformer modules in the first stage, the second stage and the fourth stage is 2, and the number of Swin transformer modules in the third stage is 6.
In one embodiment, the convolution kernel size of the first convolution layer is 1 × 1 and the channel is 256-dimensional.
In one embodiment, the convolution kernel size of the second convolution layer is 3 × 3.
In one embodiment, the convolution kernel size of the third convolution layer is 3 × 3 with a step size of 2.
As shown in fig. 2, the feature extraction backbone network adopted in the method is a Swin-Tiny version in the swintformer model, so that the detection speed of the model is greatly improved, the first Stage to the fourth Stage correspond to Swin transformer Stage1 to Swin transformer Stage4 in sequence, 2,2,6,2 Swin transformer modules are used in 4 stages respectively, input pictures (Images) are divided into 8 × 8 windows, a small image block tchpad is divided for 7 × 7 pixels of each window and is input into the Swin transformer module for self-attention calculation, parameters calculated by the windows are shared, and therefore the sequence length of the input Swin transformer modules is unified to 49. Specifically, the method comprises the following steps:
in the forward propagation flow, an image (Images) with a size of H × W × 3 is first input and split into 4 × 4 blocks by a patch partition operation, and the blocks are H/4 × W/4 small image blocks patch.
Following encoding by the stacked Swin transformer module, the first stage changes the vector dimension to the sequence length C accepted by the Swin transformer module (e.g., 96) and flattens it by the patch embedding operation module.
The first stage, the second stage and the third stage enter through a patch clustering operation module which is similar to the pooling downsampling in the convolutional neural network, the combined length-width halving channel number of each 2 × 2 adjacent small image block patch feature is expanded to four times, and simultaneously the 1 × 1 convolution is used for uniformly doubling the channel for the convolution form.
And finally outputting 4 sampling feature maps with dimensions of down-sampling 4, 8, 16 and 32 times and channel numbers 96, 192, 384 and 768 by the backbone network, and ending the feature representation learning process of the backbone network.
In order to solve the problem of detecting the large-scale change of the target scale, feature information of different scales needs to be fused to feature maps of different scaling ratios so as to adapt to the regression size information of the target frame in advance. Therefore, in this embodiment, a 5-layer Feature Pyramid Network (FPN) is used to dock the trunk Network swinttransformer model and the FCOS detection head, and the 1 st to 5 th layers sequentially correspond to Feature pyramids P3 to P7, which are specifically as follows:
firstly, 4 convolutions (first convolution layers) with 1 × 1 channels of 256 dimensions are constructed at the input side of the feature pyramid network to change the number of channels of the sampling feature map, and 4 sampling feature maps of different levels extracted by the backbone network are respectively input into the corresponding feature pyramid network through the first convolution layers.
And taking the feature pyramid network corresponding to the fourth stage of the backbone network as a reference layer, fusing the 32-time sampling feature map extracted in the fourth stage from the third stage to the first stage from top to bottom by gradually performing up-sampling operation three times from top to bottom, and enabling the 32-time sampling feature map extracted in the fourth stage to obtain a smaller feature map through a convolution layer (third convolution layer) with the length of 3 x 3 and the step length of 2 so as to generate stronger semantic information.
By the method, the target frames with different sizes and scales can be naturally distributed into the 5 feature maps output by the 5-layer feature pyramid network, and the targets in the different feature maps are further regressed and classified by the FCOS detection head. Meanwhile, the FPN network and the FCOS center sampling optimization adopted can be well adapted to the detection field with multi-scale transformation, and the parameter adjusting difficulty can be reduced based on the anchor-free frame.
The FCOS detection Head (FCOS Head) takes the output feature map of the corresponding second convolution layer as input, and divides the output feature map into two paths (classification branch and regression branch) to further regress and classify the corresponding feature map. And continuously adopting 4 convolutions of 3 multiplied by 3 to carry out further coding in two paths so as to extract classification and regression feature representations, not changing the length and width scales of the feature map in the coding process of the convolutions, and keeping the number of the channels of the feature map to be 256.5 FCOS test header feature level sharing.
The first path is a classification branch and comprises a classification prediction branch and a center-less branch, the classification prediction branch outputs H multiplied by W multiplied by C, namely, a C-dimensional vector is predicted for each point on the characteristic diagram to determine the classification of the point, and the center-less branch outputs H multiplied by W multiplied by 1, namely, only one value is predicted for each point on the characteristic diagram to represent the distance weight between the point and the target center.
The second path is a regression branch, which is responsible for performing regression (regression) on the position of the target frame, and the regression branch outputs H × W × 4, that is, each point on the feature map is regarded as a detection point and 4 distance parameters from the point to the target frame are regressed.
The FCOS detection head does not need a priori frames, the sample size and the parameter number are greatly reduced, the calculated amount is reduced, the effect of small target detection is good, and the number of the predictable frames is increased through multi-scale detection.
The predicted points within the target box are pre-screened and optimized by target box Center sampling (Center Sample). The core idea of the Center Sample is to screen a prediction point with higher probability falling on a real positive Sample in advance, firstly find a central point in a target frame, take out a circle circumscribed rectangle smaller than the target frame by taking the central point as a circle Center, redefine a sub-rectangle as a positive Sample area, only the sampling point falling in the sub-rectangle is a positive Sample, and other sampling points are regarded as negative samples. After screening, most of sampling points which are at the inner edge of the target frame but actually fall on the background part are reasonably removed, so that the convergence rate and the training result of the FCOS detection head can be optimized. Specifically, the center sample defines the center region as a sub-rectangle of the labeled frame, and the position parameter of the sub-rectangle is (c) x -rs,c y -rs,c x +rs,c y + rs) wherein (c) x ,c y ) The coordinate of the center point of the target frame is set as s, the downsampling multiplying factor of the input feature map corresponding to the FCOS detection head is set as s, the lower sampling multiplying factor stride is 2,r is used as an over-parameter, and r =1.5 can be adjusted according to actual requirements.
And S2, collecting part pictures as a data set to train the surface defect detection model.
S3, inputting the picture of the part to be detected into the trained surface defect detection model, and predicting the central point (c) of the target frame x ,c y ) And the distances (l) from the center point to the four sides of the real frame * ,t * ,r * ,b * ) Calculating the coordinates (c) of the upper left corner of the prediction box x +l * ,c y +t * ) And the coordinates of the lower right corner (c) x +r * ,c y +b * ) As a result of surface defect detection, wherein l * Is the center point (c) of the target frame x ,c y ) Distance to the left border of the real frame, t * Is the center point (c) of the target frame x ,c y ) Distance, r, to the upper border of the real frame * Is the center point (c) of the target frame x ,c y ) Distance to the right frame of the real frame, b * Is the center point (c) of the target frame x ,c y ) ToThe distance of the lower border of the real frame. By adopting the anchor-free frame mode, the detection result is obtained by directly outputting the coordinates of the prediction frame instead of outputting a plurality of candidate frames and then carrying out post-processing, and the surface defect detection result is obtained.
The method is characterized in that a Swin transform model based on a vision field self-attention mechanism and a characteristic pyramid network (FPN model) are fused to encode, fuse and express image characteristics, an anchor frame-free pixel-by-pixel regression coding mode FCOS is used for redefining a target detection head to obtain a result of a target detection downstream task, a central sampling strategy is adopted to optimize the FCOS detection head, efficient and multi-scale target surface defect detection is achieved, positive and negative samples in a target frame are balanced, targets with large shape changes can be well expressed, particularly, good detection accuracy and robustness are achieved for small-size defect targets and strip defect targets with extreme transverse-longitudinal ratios, and the method can be applied to detection in the fields of integrated circuit chips, industrial parts and the like.
The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express the more specific and detailed embodiments described in the present application, but not be construed as limiting the claims. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (7)
1. An anchor-frame-free surface defect detection method based on multi-scale features is characterized by comprising the following steps: the method for detecting the surface defects of the anchor-free frame based on the multi-scale features comprises the following steps:
s1, establishing a surface defect detection model, wherein the surface defect detection model comprises a Swin transform model, a first feature extraction unit, an N-layer feature pyramid network, a second feature extraction unit and N FCOS detection heads, and the method comprises the following steps:
the first feature extraction unit comprises a plurality of parallel first convolution layers, each stage of the Swin transform model is respectively connected with the first N-1 layers of feature pyramid networks in a one-to-one corresponding sequence through the first convolution layers, the second feature extraction unit comprises a plurality of parallel second convolution layers, the feature pyramid networks are connected with the FCOS detection head in a one-to-one corresponding mode through the second convolution layers, the feature pyramid network corresponding to the last stage of the Swin transform model is taken as a reference layer, input feature graphs of the corresponding feature pyramid networks are fused from top to bottom from the reference layer through gradual up-sampling operation, and output feature graphs of the reference layer are input into the feature pyramid network of the Nth layer after passing through the third convolution layer;
the FCOS detection head comprises a classification branch and a regression branch, the classification branch comprises a category prediction branch and a center-less branch, regression and classification of an output characteristic diagram corresponding to the second convolution layer are carried out through each FCOS detection head, and the FCOS detection head updates a target frame based on an FCOS center sampling optimization method;
s2, collecting a part picture as a data set to train the surface defect detection model;
s3, inputting the picture of the part to be detected into the trained surface defect detection model, and predicting the central point (c) of the target frame x ,c y ) And the distance (l) from the center point to the four edges of the real frame * ,t * ,r * ,b * ) Calculating the coordinates (c) of the upper left corner of the prediction box x +l * ,c y +t * ) And the coordinates of the lower right corner (c) x +r * ,c y +b * ) As a result of surface defect detection, wherein l * Is the center point (c) of the target frame x ,x y ) Distance to the left border of the real frame, t * Is the center point (c) of the target frame x ,c y ) Distance, r, to the upper border of the real frame * Is the center point (c) of the target frame x ,c y ) Distance to the right frame of the real frame, b * Is the center point (c) of the target frame x ,c y ) Distance to the lower border of the real border.
2. The method for detecting the surface defect of the anchor-free frame based on the multi-scale features as claimed in claim 1, wherein: the FCOS center sampling optimization method specifically comprises the following steps:
taking out a circle external sub-rectangle smaller than the target frame by taking the central point of the target frame as the circle center, updating the target frame into the sub-rectangle, regarding the sampling point falling into the sub-rectangle as a positive sample, regarding other sampling points as negative samples, and regarding the position parameter of the sub-rectangle as (c) x -rs,c y -rs,c x +rs,c y + rs), wherein (c) x ,c y ) The coordinate of the central point of the target frame is s, the downsampling multiplying power of the input feature map corresponding to the FCOS detection head is s, and r is a hyper-parameter.
3. The method for detecting the surface defect of the anchor-free frame based on the multi-scale features as claimed in claim 1, wherein: the Swin Transformer model comprises a Patch Partition operation module, a first stage, a second stage, a third stage and a fourth stage which are sequentially connected, wherein the first stage comprises a Linear Embedding operation module and a plurality of Swin Transformer modules which are sequentially connected, and the second stage, the third stage and the fourth stage comprise a Patch gathering operation module and a plurality of Swin Transformer modules which are sequentially connected.
4. The multi-scale feature-based anchor-frame-free surface defect detection method of claim 3, wherein: the number of Swin transducer modules in the first stage, the second stage and the fourth stage is 2, and the number of Swin transducer modules in the third stage is 6.
5. The method for detecting the surface defect of the anchor-free frame based on the multi-scale features as claimed in claim 1, wherein: the convolution kernel size of the first convolution layer is 1 × 1, and the channel has 256 dimensions.
6. The method for detecting the surface defect of the anchor-free frame based on the multi-scale features as claimed in claim 1, wherein: the convolution kernel size of the second convolution layer is 3 x 3.
7. The method for detecting the surface defect of the anchor-free frame based on the multi-scale features as claimed in claim 1, wherein: the convolution kernel size of the third convolution layer is 3 × 3 with a step size of 2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211686786.8A CN115861281A (en) | 2022-12-26 | 2022-12-26 | Anchor-frame-free surface defect detection method based on multi-scale features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211686786.8A CN115861281A (en) | 2022-12-26 | 2022-12-26 | Anchor-frame-free surface defect detection method based on multi-scale features |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115861281A true CN115861281A (en) | 2023-03-28 |
Family
ID=85655126
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211686786.8A Pending CN115861281A (en) | 2022-12-26 | 2022-12-26 | Anchor-frame-free surface defect detection method based on multi-scale features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115861281A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117094999A (en) * | 2023-10-19 | 2023-11-21 | 南京航空航天大学 | Cross-scale defect detection method |
| CN117274253A (en) * | 2023-11-20 | 2023-12-22 | 华侨大学 | Part detection method and device based on multimode transducer and readable medium |
| CN117496323A (en) * | 2023-12-27 | 2024-02-02 | 泰山学院 | Multi-scale second-order pathological image classification method and system based on transducer |
-
2022
- 2022-12-26 CN CN202211686786.8A patent/CN115861281A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117094999A (en) * | 2023-10-19 | 2023-11-21 | 南京航空航天大学 | Cross-scale defect detection method |
| CN117094999B (en) * | 2023-10-19 | 2023-12-22 | 南京航空航天大学 | Cross-scale defect detection method |
| CN117274253A (en) * | 2023-11-20 | 2023-12-22 | 华侨大学 | Part detection method and device based on multimode transducer and readable medium |
| CN117274253B (en) * | 2023-11-20 | 2024-02-27 | 华侨大学 | Part detection method and device based on multimode transducer and readable medium |
| CN117496323A (en) * | 2023-12-27 | 2024-02-02 | 泰山学院 | Multi-scale second-order pathological image classification method and system based on transducer |
| CN117496323B (en) * | 2023-12-27 | 2024-03-29 | 泰山学院 | Multi-scale second-order pathological image classification method and system based on transducer |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112396002B (en) | SE-YOLOv 3-based lightweight remote sensing target detection method | |
| CN110321923B (en) | Target detection method, system and medium for fusion of different-scale receptive field characteristic layers | |
| CN115861281A (en) | Anchor-frame-free surface defect detection method based on multi-scale features | |
| CN109902600B (en) | Road area detection method | |
| CN112561910A (en) | Industrial surface defect detection method based on multi-scale feature fusion | |
| CN111680655A (en) | Video target detection method for aerial images of unmanned aerial vehicle | |
| CN111161244B (en) | Industrial product surface defect detection method based on FCN + FC-WXGboost | |
| CN113313706B (en) | Power equipment defect image detection method based on detection reference point offset analysis | |
| CN111738114B (en) | Vehicle target detection method based on anchor-free accurate sampling remote sensing image | |
| CN111027539A (en) | License plate character segmentation method based on spatial position information | |
| CN114359245A (en) | Method for detecting surface defects of products in industrial scene | |
| CN114820579A (en) | Semantic segmentation based image composite defect detection method and system | |
| CN112633149A (en) | Domain-adaptive foggy-day image target detection method and device | |
| CN116152498A (en) | Metal surface defect semantic segmentation network and training method based on data driving | |
| CN115731400A (en) | X-ray image foreign matter detection method based on self-supervision learning | |
| CN109919215B (en) | Target detection method for improving characteristic pyramid network based on clustering algorithm | |
| CN114092467A (en) | Scratch detection method and system based on lightweight convolutional neural network | |
| CN115631186B (en) | Industrial element surface defect detection method based on double-branch neural network | |
| CN110033443B (en) | Display panel defect detection method | |
| CN116958073A (en) | Small sample steel defect detection method based on attention feature pyramid mechanism | |
| CN115578339A (en) | Industrial product surface defect detection and positioning method, system and equipment | |
| CN114419078B (en) | Surface defect region segmentation method and device based on convolutional neural network | |
| CN116052149A (en) | CS-ABCNet-based electric power tower plate detection and identification method | |
| CN115830302A (en) | Multi-scale feature extraction and fusion power distribution network equipment positioning identification method | |
| 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 |