CN112489001B - Tunnel water seepage detection method based on improved deep learning - Google Patents
Tunnel water seepage detection method based on improved deep learning Download PDFInfo
- Publication number
- CN112489001B CN112489001B CN202011319025.XA CN202011319025A CN112489001B CN 112489001 B CN112489001 B CN 112489001B CN 202011319025 A CN202011319025 A CN 202011319025A CN 112489001 B CN112489001 B CN 112489001B
- Authority
- CN
- China
- Prior art keywords
- feature
- feature map
- channel
- images
- attention
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
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
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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
- G06T7/00—Image analysis
- G06T7/97—Determining parameters from multiple pictures
-
- 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]
-
- 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
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a tunnel water seepage detection method based on improved deep learning, which comprises the following steps: (1) residual network-based image feature extraction; (2) attention mechanism: adding an attention module after each stage of the residual error network, strengthening the feature map containing more semantic information, and inhibiting the feature map containing useless information; (3) propagating the high-level semantic information: the high-level semantic information is transmitted to the low-level feature map in a top-down mode; (4) obtaining context information: channel splicing is carried out on the feature images with different scales, and a channel attention mechanism module is followed to correct the feature images; and (5) water seepage prediction. According to the invention, aiming at tunnel lining surface image features, different semantic information features are extracted by using a deep learning method, meanwhile, a attention mechanism is adopted to correct the feature map, and the different semantic information features are fused, so that the low-order feature map has better noise resistance, and accurate tunnel water seepage detection is realized.
Description
Technical Field
The invention relates to an image detection method, in particular to a tunnel water seepage image, and particularly relates to a tunnel water seepage detection method based on improved deep learning.
Background
The tunnel water seepage detection is to identify the water seepage on the surface of the tunnel lining. And the tunnel seepage damage is detected in time, so that the seepage damage is prevented from causing larger damage and threatening the operation safety of railway lines. At present, the tunnel seepage detection mostly adopts the mode of manual inspection, and the manual mode can receive subjective factor's influence, causes to distinguish the deviation, seriously threatens inspection personnel's safety in addition. With the development of computer vision technology, especially the development of deep learning technology, nondestructive testing based on deep learning becomes a research hotspot for detecting tunnel diseases at home and abroad.
And processing and analyzing the tunnel surface image through a deep learning algorithm to obtain a label graph containing the position and the size of the water seepage area. The label graph can provide accurate information for tunnel maintenance work. At present, the tunnel water seepage detection based on a deep learning method mostly adopts the existing deep learning model, and the water seepage detection is carried out through the surface image of the tunnel lining of the deep learning model.
Deep learning comprises a plurality of convolution layers, feature maps of different semantic information are obtained in different convolution stages, and low-level feature maps comprise more detail information and position information but are sensitive to noise; the high-level feature map contains more high-level semantic information and shows better noise immunity. And the feature images with different levels are fused, and the fused feature images contain detail and position information and simultaneously show good noise resistance. Deep learning obtains the capability of extracting rich features through data self-training, but due to the complexity of the surface of the tunnel lining and the variability of the size of the water seepage area, useless feature patterns are also included in the deep learning, and the deep learning is unfavorable for detecting the water seepage area to finish accurate water seepage identification.
Disclosure of Invention
The invention aims at the tunnel lining surface image characteristics, extracts different semantic information characteristics by using a deep learning method, corrects a characteristic diagram by adopting a attention mechanism, and provides a tunnel water seepage detection method based on improved deep learning by using different semantic information characteristic fusion.
The invention adopts the following technical scheme:
a tunnel water seepage detection method based on improved deep learning comprises the following steps:
(1) Extracting image features based on a residual error network;
(2) Attention mechanism: adding an attention module after each stage of the residual error network, strengthening the feature map containing more semantic information, and inhibiting the feature map containing useless information;
(3) Propagating the high-level semantic information: the high-level semantic information is transmitted to the low-level feature map in a top-down mode;
(4) Obtaining context information: channel splicing is carried out on the feature images with different scales, the feature images contain more water seepage area context information, and a channel attention mechanism module is arranged behind the feature images to correct the feature images;
(5) And (3) water seepage prediction: and (3) inputting the corrected feature map obtained in the step (4) into a classification layer to classify the image pixels.
Wherein, the step (1) is as follows: and inputting the tunnel lining image into a residual network to obtain feature images with different scales.
Wherein the residual network is Resnet101.
Wherein, the step (2) is as follows: and (3) adding an attention mechanism module between every two stages in the residual error network of the step (1), and correcting the feature graphs with different scales obtained in the step (1).
Wherein the attention mechanism module comprises:
(a) Channel attention correction: carrying out global maximum pooling and global average pooling on an input image to obtain different channel attention vectors, obtaining two channel attention feature vectors Vm and Va by two different channel attention vectors through two full-connection layers, adding Vm and Va, and then sending the Vm and Va into a sigmoid activation function to obtain a final channel attention feature vector Vc; multiplying Vc with the input feature map;
(b) Spatial attention correction: carrying out global maximum pooling and global average pooling on an input image to obtain two different spatial feature images, then splicing the two spatial feature images, inputting the two spatial feature images into a 7*7 convolution layer, and obtaining a final spatial attention feature image Fc by a sigmoid activation function; multiplying Fc with the input signature;
(c) Correcting the characteristics extracted by the residual network in the serial mode through the steps (a) and (b): the input feature map is first channel-corrected, and then channel-corrected.
When the high-level semantic information is transmitted from top to bottom to the low-level feature map for feature fusion in the step (3), a bilinear interpolation method is adopted, after dimensions and sizes are unified, a channel splicing mode is adopted, and finally final feature fusion is realized through 3*3 convolution.
Wherein, the step (4) is as follows: firstly, carrying out alignment and bilinear interpolation for enlarging scale, adjusting the respective scale to be the same as the scale of an input tunnel image, then carrying out channel splicing on feature images of different grades, and finally inputting the feature images after channel splicing into a channel attention module.
Wherein, the step (5) is as follows: the corrected feature map obtained in the step (4) is consistent with the size of the network input image, then the feature map is input into a 1*1 convolution layer, the number of the finally obtained channels is 1, and finally classification is carried out through a sigmoid activation function; and obtaining a final prediction graph P.
The invention has the beneficial effects that: the invention utilizes the residual network (Resnet 101) to extract the image characteristics; and then, sending the feature images with different scales output by the network into an attention module, finally, fusing the high-order feature images with the low-order feature images in a top-down mode, so that the high-order semantic information is transmitted to the low-order feature images, the low-order feature images show better noise resistance, and finally, splicing the feature images with different scales in a channel, thereby realizing accurate tunnel water seepage detection through an attention mechanism.
Drawings
Fig. 1 is a diagram of a network model structure according to the present invention.
Detailed Description
The technical solution of the present application will be clearly and completely described below with reference to fig. 1.
Step 1: image feature extraction method based on residual network (Resnet 101). And inputting the tunnel lining image into a residual network to obtain feature images with different scales.
The deep learning model is a published Resnet101 model, the scale of an input image is 1024 x 3, and a tunnel water seepage image training set is constructed, wherein the tunnel water seepage image training set comprises 1000 tunnel water seepage images and 1000 tunnel non-water seepage images. Finally, 800 water seepage images and 800 water non-seepage images are selected for training, and the rest 400 images are used for testing. Image feature extraction is performed using a network of Resnet101. The extracted feature images are the following 5 feature images with dimensions of 64, 128, 256, 512 and 1024, respectively, and dimensions of 1/2,1/4,1/8,1/16 and 1/32 of the input image, respectively.
Step 2: attention mechanism, adding an attention module after each stage of the residual network (Resnet 101) strengthens the feature map containing more semantic information and suppresses the feature map containing useless information (as shown in FIG. 1).
The network of the resnet101 in the step 1 is divided into five stages, each stage comprises a plurality of convolution layers and a maximum pooling layer, so that image features with different scales are obtained in different stages, and in addition, feature maps with different scales also comprise different semantic information. For better feature extraction, a Convolutional Block Attention Module (CBAM) attention mechanism module is added in the middle of every two stages.
CBAM is divided into three parts:
(1) Channel attention correction. Firstly, carrying out global maximum pooling and global average pooling on an input image to obtain different channel attention vectors, changing the scale from H to W to C to 1 to C, obtaining two channel attention feature vectors Vm and Va by two different channel attention vectors through two full connection layers, adding Vm and Va, and then sending the sum to a sigmoid activation function to obtain a final channel attention feature vector Vc. By multiplying Vc with the input feature map, the response of the feature map containing the water immersion semantic information is improved, and the response of the feature map not containing the water immersion semantic information is reduced.
(2) Spatial attention correction. Firstly, carrying out global maximum pooling and global average pooling on an input image to obtain two different space feature images, changing the scale from H to W to 1, then splicing the two space feature images to obtain a space feature image with the scale of H to W to 2, inputting the space feature image into a 7*7 convolution layer, which can contain more semantic information of a receptive field, obtaining a better space feature image, and obtaining a final space attention feature image Fc by following a sigmoid activation function. And multiplying Fc by the input feature map to obtain the response of the spatial position containing the water immersion semantic information, and reducing the response of the noise area.
(3) Finally, the channel attention and the space attention are connected in series, and the channel attention is corrected for the input feature images, so that the influence of useless feature images can be reduced, and then the channel attention is corrected for the input feature images.
Step 3: and (3) propagating the high-level semantic information, and propagating the high-level semantic information into the low-level feature map in a top-down mode.
Five image feature maps with different sizes are obtained through the steps 1 and 2, namely C1, C2, C3, C4 and C5. The scale is reduced in turn, but the noise immunity is increased in turn, and the semantic information extracting capability is increased in turn. When the feature fusion is carried out, firstly, the top-down feature fusion is carried out, C5 is selected to amplify the scale by 2 times through a bilinear interpolation method, then the dimension reduction operation is carried out through a 3*3 convolution layer, the dimension reduction operation is marked as F5, the F5 and the C4 channels are spliced, and finally, the feature fusion is carried out through 3*3 convolution layers one by one, so that F4 is obtained. Then C3, C2 and C1 are fused in sequence to obtain fusion characteristic diagrams F3, F2 and F1. When the feature fusion is carried out, firstly dimension unification is realized, and dimension unification is also realized, wherein the dimension unification is realized by adopting a bilinear interpolation method; after the dimensions and the sizes are unified, a channel splicing mode is adopted, and finally, final feature fusion is achieved through 3*3 convolution.
Step 4: and obtaining context information, performing channel splicing on the feature images with different scales, including context information of more water seepage areas, and correcting the feature images by a channel attention mechanism module.
The feature maps F4, F3, F2, F1 are obtained through the above steps. The dimensions are different, and the semantic information contained in the memory is also different, so that in order to obtain different context semantic information, uniform dimensions are needed to fuse the context semantic information. Bilinear interpolation of C5, F4, F3, F2 restores the scale to the same scale as the original. Channel splicing is carried out on the amplified C5, F4, F3, F2 and F1 to obtain a feature map F containing more context semantic information f， Finally F is arranged f Input a Channel Attention Module (channel attention module), first F f And carrying out global maximum pooling operation to obtain a compressed feature vector, then obtaining a final channel correction vector V through two full-connection layers, and finally multiplying the V with an input feature map to highlight more important feature channels.
After top-down feature fusion is carried out, the low-order feature images keep the details and the position information, and meanwhile, the high-order feature images contain high-level voice information, and although the detail information is lost, the high-order feature images extract better semantic information from a large-scale water seepage area, so that feature images of different levels are fused, firstly, the feature images are aligned and amplified by a bilinear interpolation method, the respective scales of the feature images are adjusted to be the same as the scales of an input tunnel image, and then, channel splicing is carried out on the feature images of different levels, so that more context semantic information is kept. Finally, through a Squeeze-and-specification module, detection of water seepage areas with different scales is realized.
Step 5: and (3) water seepage prediction, namely inputting the corrected feature map into a classification layer, and classifying image pixels.
The method comprises the steps of obtaining a feature map which is finally consistent with the size of a network input image through the steps, inputting the feature map into a 1*1 convolution layer, obtaining the number of channels which is 1 finally, classifying through a sigmoid activation function, and finally outputting a prediction map P which is identical with the input image in scale.
The above embodiments are only preferred examples of the invention and are not exhaustive of the possible implementations of the invention. Any obvious modifications thereof, which would be apparent to those skilled in the art without departing from the principles and spirit of the present invention, should be considered to be included within the scope of the appended claims.
Claims (4)
1. The tunnel water seepage detection method based on improved deep learning is characterized by comprising the following steps of:
(1) Extracting image features based on a residual error network; inputting tunnel lining images into a residual error network to obtain feature images with different scales; the residual network is Resnet101;
(2) Attention mechanism: in the residual error network of the step (1), an attention mechanism module is added between every two stages, and the feature graphs with different scales obtained in the step (1) are corrected;
the attention mechanism module includes:
(a) Channel attention correction: carrying out global maximum pooling and global average pooling on an input image to obtain different channel attention vectors, obtaining two channel attention feature vectors Vm and Va by two different channel attention vectors through two full-connection layers, adding Vm and Va, and then sending the Vm and Va into a sigmoid activation function to obtain a final channel attention feature vector Vc; multiplying Vc with the input feature map;
(b) Spatial attention correction: carrying out global maximum pooling and global average pooling on an input image to obtain two different spatial feature images, then splicing the two spatial feature images, inputting the two spatial feature images into a 7*7 convolution layer, and obtaining a final spatial attention feature image Fc by a sigmoid activation function; multiplying Fc with the input signature;
(c) Correcting the characteristics extracted by the residual network in the serial mode through the steps (a) and (b): firstly, carrying out channel attention correction on an input feature map, and then carrying out channel attention correction on the input feature map;
(3) Propagating the high-level semantic information: the high-level semantic information is transmitted to the low-level feature map in a top-down mode;
(4) Obtaining context information: channel splicing is carried out on the feature images with different scales, the feature images contain more water seepage area context information, and a channel attention mechanism module is arranged behind the feature images to correct the feature images;
(5) And (3) water seepage prediction: and (3) inputting the corrected feature map obtained in the step (4) into a classification layer to classify the image pixels.
2. The tunnel water seepage detection method based on improved deep learning according to claim 1, wherein in the step (3), when the high-level semantic information is propagated from top to bottom to the low-level feature map for feature fusion, a bilinear interpolation method is adopted, after dimensions and sizes are unified, a channel splicing mode is adopted, and finally final feature fusion is achieved through 3*3 convolution.
3. The tunnel penetration detection method based on improved deep learning of claim 2, wherein the step (4) is: firstly, carrying out alignment and bilinear interpolation for enlarging scale, adjusting the respective scale to be the same as the scale of an input tunnel image, then carrying out channel splicing on feature images of different grades, and finally inputting the feature images after channel splicing into a channel attention module.
4. The tunnel penetration detection method based on improved deep learning of claim 3, wherein the step (5) is: the corrected feature map obtained in the step (4) is consistent with the size of the network input image, then the feature map is input into a 1*1 convolution layer, the number of the finally obtained channels is 1, and finally classification is carried out through a sigmoid activation function; and obtaining a final prediction graph P.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011319025.XA CN112489001B (en) | 2020-11-23 | 2020-11-23 | Tunnel water seepage detection method based on improved deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011319025.XA CN112489001B (en) | 2020-11-23 | 2020-11-23 | Tunnel water seepage detection method based on improved deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112489001A CN112489001A (en) | 2021-03-12 |
| CN112489001B true CN112489001B (en) | 2023-07-25 |
Family
ID=74932927
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011319025.XA Active CN112489001B (en) | 2020-11-23 | 2020-11-23 | Tunnel water seepage detection method based on improved deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112489001B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197468A (en) * | 2019-06-06 | 2019-09-03 | 天津工业大学 | A kind of single image Super-resolution Reconstruction algorithm based on multiple dimensioned residual error learning network |
| CN110544264A (en) * | 2019-08-28 | 2019-12-06 | 北京工业大学 | Temporal bone key anatomical structure small target segmentation method based on 3D deep supervision mechanism |
| CN110599502A (en) * | 2019-09-06 | 2019-12-20 | 江南大学 | Skin lesion segmentation method based on deep learning |
| CN110781923A (en) * | 2019-09-27 | 2020-02-11 | 重庆特斯联智慧科技股份有限公司 | Feature extraction method and device |
| CN111160311A (en) * | 2020-01-02 | 2020-05-15 | 西北工业大学 | Yellow river ice semantic segmentation method based on multi-attention machine system double-flow fusion network |
| CN111161195A (en) * | 2020-01-02 | 2020-05-15 | 重庆特斯联智慧科技股份有限公司 | Feature map processing method and device, storage medium and terminal |
| CN111242288A (en) * | 2020-01-16 | 2020-06-05 | 浙江工业大学 | Multi-scale parallel deep neural network model construction method for lesion image segmentation |
| CN111353396A (en) * | 2020-02-20 | 2020-06-30 | 江苏东印智慧工程技术研究院有限公司 | Concrete crack segmentation method based on SCSEOCUnet |
| CN111738110A (en) * | 2020-06-10 | 2020-10-02 | 杭州电子科技大学 | Remote sensing image vehicle target detection method based on multi-scale attention mechanism |
-
2020
- 2020-11-23 CN CN202011319025.XA patent/CN112489001B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197468A (en) * | 2019-06-06 | 2019-09-03 | 天津工业大学 | A kind of single image Super-resolution Reconstruction algorithm based on multiple dimensioned residual error learning network |
| CN110544264A (en) * | 2019-08-28 | 2019-12-06 | 北京工业大学 | Temporal bone key anatomical structure small target segmentation method based on 3D deep supervision mechanism |
| CN110599502A (en) * | 2019-09-06 | 2019-12-20 | 江南大学 | Skin lesion segmentation method based on deep learning |
| CN110781923A (en) * | 2019-09-27 | 2020-02-11 | 重庆特斯联智慧科技股份有限公司 | Feature extraction method and device |
| CN111160311A (en) * | 2020-01-02 | 2020-05-15 | 西北工业大学 | Yellow river ice semantic segmentation method based on multi-attention machine system double-flow fusion network |
| CN111161195A (en) * | 2020-01-02 | 2020-05-15 | 重庆特斯联智慧科技股份有限公司 | Feature map processing method and device, storage medium and terminal |
| CN111242288A (en) * | 2020-01-16 | 2020-06-05 | 浙江工业大学 | Multi-scale parallel deep neural network model construction method for lesion image segmentation |
| CN111353396A (en) * | 2020-02-20 | 2020-06-30 | 江苏东印智慧工程技术研究院有限公司 | Concrete crack segmentation method based on SCSEOCUnet |
| CN111738110A (en) * | 2020-06-10 | 2020-10-02 | 杭州电子科技大学 | Remote sensing image vehicle target detection method based on multi-scale attention mechanism |
Non-Patent Citations (2)
| Title |
|---|
| "CBAM: Convolutional Block Attention Module";Sanghyun Woo等;《arXiv》;20180718;第1-17页 * |
| "融合残差注意力机制的 UNet 视盘分割";侯向丹等;《中国图象图形学报》;20200930;第25卷(第9期);第1915-1929页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112489001A (en) | 2021-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109299274B (en) | Natural scene text detection method based on full convolution neural network | |
| CN107330453B (en) | Pornographic image identification method based on step-by-step identification and fusion key part detection | |
| CN109977921B (en) | Method for detecting hidden danger of power transmission line | |
| CN111681273B (en) | Image segmentation method and device, electronic equipment and readable storage medium | |
| CN110111334B (en) | Crack segmentation method and device, electronic equipment and storage medium | |
| CN112164038A (en) | Photovoltaic hot spot detection method based on deep convolutional neural network | |
| US20170200274A1 (en) | Human-Shape Image Segmentation Method | |
| CN111950453A (en) | Optional-shape text recognition method based on selective attention mechanism | |
| CN111257341B (en) | Underwater building crack detection method based on multi-scale features and stacked full convolution network | |
| JP2020123330A (en) | Method for acquiring sample image for label acceptance inspection from among auto-labeled images utilized for neural network learning, and sample image acquisition device utilizing the same | |
| CN112488025B (en) | Double-temporal remote sensing image semantic change detection method based on multi-modal feature fusion | |
| CN115131797B (en) | Scene text detection method based on feature enhancement pyramid network | |
| CN111914654B (en) | Text layout analysis method, device, equipment and medium | |
| CN110163208A (en) | A kind of scene character detecting method and system based on deep learning | |
| CN112381763A (en) | Surface defect detection method | |
| CN112991364A (en) | Road scene semantic segmentation method based on convolution neural network cross-modal fusion | |
| CN112488229A (en) | Domain self-adaptive unsupervised target detection method based on feature separation and alignment | |
| WO2024066035A1 (en) | Defect detection method and system based on battery surface image, and related device | |
| CN113344826A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| CN115830471A (en) | Multi-scale feature fusion and alignment domain self-adaptive cloud detection method | |
| CN111539924A (en) | Defect detection method, device and equipment for suspension clamp and storage medium | |
| CN104899821A (en) | Method for erasing visible watermark of document image | |
| CN115471754A (en) | Remote sensing image road extraction method based on multi-dimensional and multi-scale U-net network | |
| CN113780423A (en) | Single-stage target detection neural network based on multi-scale fusion and industrial product surface defect detection model | |
| CN116823793A (en) | Device defect detection method, device, electronic device and readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |