CN115272746A - Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency - Google Patents

Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency Download PDF

Info

Publication number
CN115272746A
CN115272746A CN202210758465.8A CN202210758465A CN115272746A CN 115272746 A CN115272746 A CN 115272746A CN 202210758465 A CN202210758465 A CN 202210758465A CN 115272746 A CN115272746 A CN 115272746A
Authority
CN
China
Prior art keywords
bridge
damage
small sample
guided
consistency
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.)
Granted
Application number
CN202210758465.8A
Other languages
Chinese (zh)
Other versions
CN115272746B (en
Inventor
徐阳
李惠
范云蕾
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202210758465.8A priority Critical patent/CN115272746B/en
Publication of CN115272746A publication Critical patent/CN115272746A/en
Application granted granted Critical
Publication of CN115272746B publication Critical patent/CN115272746B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/761—Proximity, similarity or dissimilarity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

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

Abstract

The invention discloses a general identification method and a system for multiple types of damage of a bridge guided by small sample circulation consistency, wherein the method comprises the following steps: carrying out data alignment on preset bridge damage types, dividing a data set after the data alignment into a support set and a query set so as to establish a bridge multi-type damage general identification model guided by small sample cycle consistency; calculating a first loss value of a query set by using a universal identification model for multiple types of damage of the bridge guided by the cyclic consistency of the small sample; calculating a second loss value of the support set by utilizing a bridge multi-type damage general identification model and introducing cycle consistency; and calculating the comprehensive loss of the bridge multi-type damage general identification model guided by the small sample cycle consistency based on the first loss value and the second loss value, and updating the model through an error back propagation algorithm until the model converges. The method solves the problem of small samples and also solves the defect that the traditional method needs to establish different recognition models aiming at different tasks.

Description

Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency
Technical Field
The invention relates to the technical field of intelligent infrastructure and intelligent bridge detection, in particular to a universal identification method and system for multiple types of damage of a bridge guided by small sample circulation consistency.
Background
The bridge is an important component part for infrastructure construction and national economic development of China, and the number and scale of the bridge in China currently leap ahead in the world. The bridge structure inevitably suffers from the coupling effect of complex factors such as environmental erosion, material aging, fatigue load, disasters, emergencies and the like in a service period of hundreds of years to form various types of damages such as concrete peeling, concrete cracks, steel bar exposure, stay cable corrosion peeling, steel structure fatigue cracks and the like. The service safety situation of the bridge structure is more severe, and the tasks of detection, monitoring, operation, maintenance and management are increasingly aggravated.
The structure health monitoring and detecting technology becomes an advanced and effective method for guaranteeing the safety of the bridge. The traditional manual detection method depends heavily on subjective judgment of detection personnel, has low accuracy and stability, is very difficult to detect certain hard-to-reach areas, has high risk, high cost, lagged prediction and poor timeliness, and cannot meet the bridge safety management requirement in the current digital era.
Computer vision and deep learning techniques are currently undergoing rapid development. Particularly in the field of supervised learning, with the increase of available data sets, researchers at home and abroad successively put forward a series of image classification, target detection and semantic segmentation methods suitable for different computer vision recognition tasks, so as to respectively realize global scene understanding on an input image, positioning of a boundary box of the image containing a target object and classification and recognition of pixel levels. These methods have already begun to be applied to image recognition of bridge structure damage, and the advancement of image-based bridge structure damage recognition methods depends on the latest progress of computer vision field recognition models. Currently, the idea of improving different network models can be summarized in the following two forms:
(1) By improving a network architecture, such as ResNet/U-Net/DenseNet/PANET and the like, the relation between adjacent or different levels of feature maps and feature fusion or aggregation are increased, the depth and width of feature extraction are increased, and the feature extraction capability of a model is improved;
(2) By designing special function modules, such as a channel or spatial self-attention mechanism, a transform series architecture and the like, adding a new function module or replacing an original module to improve the perception capability of the model on some important features (such as correlation between local pixel areas or the whole image and the like), the performance of the model is improved.
At present, the mainstream bridge damage identification method based on computer vision generally adopts a supervised learning mode, and the identification effect of a model is very dependent on the supervision conditions: in actual engineering, a large amount of training data and labels are collected firstly, and the larger the number of training samples, the richer the categories and the higher the label accuracy, the better the recognition and generalization capability of the model. However, in a complex service scene of an actual bridge, a bridge damage image dataset often has the characteristics of incomplete information, insufficient samples, unbalanced data and inaccurate labels, that is, certain types of actual data often have the characteristics of small samples and strong interference, so that the feature expression capability of a trained model on a small sample image is not comprehensive, and the recognition accuracy and generalization capability of the model obtained by training on the small sample dataset in a real scene are poor.
Besides the problem of biased identification of small sample data, another difficulty faced by bridge multi-type damage identification is how to use a uniform model to realize universal identification of multi-type damage. The multi-class object semantic segmentation task in the traditional computer vision field often has the following characteristics: a single image in the data set includes substantially all or most of the object types; in contrast, the damage image obtained from an actual bridge often contains only one or a few types of damages, and does not contain all types or most types of damage forms, which results in that the model may be biased toward the damage types containing a larger number of samples, and the recognition effect is poor on the damage types with a smaller number.
Disclosure of Invention
The present invention is directed to solving, at least to some extent, one of the technical problems in the related art.
Therefore, the invention aims to provide a universal identification method for multiple types of damages of a bridge guided by small sample cycle consistency.
The invention also provides a bridge multi-type damage universal identification system guided by the small sample cycle consistency.
A third object of the invention is to propose a computer device.
A fourth object of the invention is to propose a non-transitory computer-readable storage medium.
In order to achieve the above object, an embodiment of the first aspect of the present invention provides a method for universally identifying multiple types of damages of a bridge guided by small sample cycle consistency, including the following steps: the method comprises the following steps that S1, data alignment is carried out on preset bridge damage types, and a data set after the data alignment is divided into a support set and a query set; s2, establishing a universal identification model of the multi-type damage of the bridge guided by the small sample cycle consistency by using the support set and the query set; s3, calculating a first loss value of the query set by using the bridge multi-type damage universal identification model guided by the small sample circulation consistency; s4, calculating a second loss value of the support set by using a universal identification model of the multi-type damage of the bridge guided by the small sample cycle consistency and introducing cycle consistency; and S5, calculating the comprehensive loss of the bridge multi-type damage general identification model guided by the small sample circulation consistency based on the first loss value and the second loss value, and updating the model through an error back propagation algorithm until the model converges.
According to the general identification method for the multi-type damage of the bridge guided by the small sample cycle consistency, disclosed by the embodiment of the invention, aiming at the difficult problem of the small sample of the bridge damage image identification, the image is divided into a support set and a query set, and the problem of the small sample is solved by designing the cycle consistency of a model on the query set and the support set; the universal identification of different types of bridge damages by adopting a unified model is realized, and the defect that different identification models need to be established for different tasks in the traditional method is overcome; only small sample images are used, a double supervision information utilization mode of a query set and a support set is created, and simultaneous learning of two supervision paths is achieved: firstly, taking an original image of a support set and original images of an annotation and query set as input, obtaining cosine similarity tensors of a multi-class damage high-dimensional feature space, taking the annotation of the query set as a true value, and calculating the loss of the query set; and then obtaining an approximate result of the query set annotation based on the query set original image and the cosine similarity tensor, taking the support set original image as input, taking the support set annotation as a true value, and calculating the loss of the support set.
In addition, the general identification method for the multi-type damage of the bridge guided by the small sample cycle consistency according to the above embodiment of the present invention may further have the following additional technical features:
further, in an embodiment of the present invention, the step S1 specifically includes:
step S101, defining the preset bridge damage type, and performing data alignment on the multi-type bridge damage image and the multi-channel pixel level mark corresponding to the multi-type bridge damage image to form a small sample data set;
step S102, the small sample data set is randomly divided into mutually exclusive support set and query set.
Further, in an embodiment of the present invention, the step S3 specifically includes:
step S301, inputting the original image of the support set into an encoder of the bridge multi-type damage general identification model guided by the small sample cycle consistency for feature extraction to obtain a high-level feature map, and simultaneously, passing the multi-channel pixel-level annotation image corresponding to the high-level feature map through a down-sampling module at the same level as the encoder to obtain a mask map;
step S302, performing element-by-element product operation on the mask graph and the high-level feature graph, and performing global average pooling to obtain a feature vector cluster;
step S303, inputting the original image of the query set into an encoder of the bridge multi-type damage general identification model guided by the small sample cycle consistency to obtain a corresponding high-level feature map, splitting the feature vector cluster into a plurality of feature vectors according to the damage type channel direction, and calculating cosine similarity tensors of each feature vector and the high-level feature map of the original image of the query set point by point;
step S304, carrying out channel-by-channel splitting and channel expansion on the cosine similarity tensor according to the channel direction of the damage types, and carrying out element-by-element multiplication on the cosine similarity tensor and a high-level feature map of an original image of the query set to obtain a plurality of corresponding damage types and feature maps; and then, sequentially passing through a decoder with an anti-symmetric structure with the encoder one by one according to the specified damage type order to obtain the prediction result of the original image of the query set, and calculating a first loss value with the multi-channel pixel-level label of the query set.
Further, in an embodiment of the present invention, the step S4 specifically includes:
performing softmax operation on the cosine similarity tensor along a damage type channel direction to obtain pixel-level labeled single-hot codes of the original image of the query set, then taking the original image of the query set, the multi-channel pixel-level labeled single-hot codes thereof and the original image of the support set as input, taking the multi-channel pixel-level labeled image corresponding to the original image in the support set as a true value, repeating the steps S302 to S304 to obtain a prediction result of the original image of the support set, and calculating a second loss value with the multi-channel pixel-level label of the support set.
Further, in an embodiment of the present invention, the cosine similarity tensor is calculated as:
Figure BDA0003723458120000041
wherein N is the total number of the damage type channels, i is the index sequence number, ViChannel i, F, which is a cluster of feature vectors(a,b)A vector corresponding to the high-level feature map at the plane position (a, b) obtained by the encoder for the original image of the query set, | |)2Is a vector modulo operation, D is a cosine similarity tensor, (H ', W') is H 'xW'Of the plane of (a).
Further, in an embodiment of the present invention, the calculation formula of the one-hot code is:
Figure BDA0003723458120000042
wherein, N is the total number of the damage type channels, i is the index sequence number, and Maxindicator represents that the following operations are executed: assigning the position with the maximum numerical value in each element of one vector as 1, and assigning the rest positions as 0; exp is an exponent fetching operation;
Figure BDA0003723458120000043
a cosine similarity value corresponding to the ith channel of the cosine similarity tensor at the plane position (a, b);
Figure BDA0003723458120000044
for the one-hot encoding corresponding to the ith channel at plane location (a, b), (H ', W') is the plane of H '× W'.
Further, in an embodiment of the present invention, the total loss in step S5 is:
L＝L1+λL2
wherein, L is the comprehensive loss, L1 is the first loss value of the query set, L2 is the second loss value of the support set after considering the cycle consistency, and lambda is the proportionality coefficient, and is set according to the ratio of the sample number of the query set to the sample number of the support set.
In order to achieve the above object, a second embodiment of the present invention provides a universal identification system for multiple types of damages of a bridge guided by small sample cycle consistency, including: the data set dividing module is used for carrying out data alignment on the preset bridge damage type and dividing the data set after the data alignment into a support set and a query set; the construction module is used for establishing a bridge multi-type damage universal identification model guided by small sample circulation consistency by utilizing the support set and the query set; the first loss calculation module is used for calculating a first loss value of the query set by utilizing the bridge multi-type damage universal identification model guided by the small sample cyclic consistency; the second loss calculation module is used for utilizing the bridge multi-type damage general identification model guided by the small sample cyclic consistency, introducing cyclic consistency and calculating a second loss value of the support set; and the updating convergence module is used for calculating the comprehensive loss of the bridge multi-type damage general identification model guided by the small sample cycle consistency based on the first loss value and the second loss value, and updating the model through an error back propagation algorithm until the model converges.
In order to achieve the above object, a third embodiment of the present invention provides a computer device, which includes a memory and a processor, where the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.
To achieve the above object, a fourth embodiment of the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored, where the computer program is used to implement the steps of the method described above when executed by a processor.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The foregoing and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a flowchart of a general identification method for multiple types of damage to a bridge guided by small sample loop consistency according to an embodiment of the present invention;
FIG. 2 is a schematic structural diagram of a small sample loop consistency guided bridge multi-type damage universal identification model according to an embodiment of the present invention;
FIG. 3 is a comparison of the semantic segmentation recognition results of multiple types of damage to a small sample bridge according to an embodiment of the present invention and the conventional method, (a) is concrete cracks, (b) is concrete spalling, and (c) is steel structure fatigue cracks;
FIG. 4 is a comparison graph of semantic segmentation recognition accuracy of a bridge multi-type damage in a small sample scene according to an embodiment of the present invention and a conventional method, (a) is a concrete crack, (b) is a concrete spalling, and (c) is a steel structure fatigue crack;
fig. 5 is a schematic structural diagram of a small sample loop consistency guided bridge multi-type damage universal identification system according to an embodiment of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.
The following describes a general identification method and system for multiple types of damages of a bridge guided by small sample cycle consistency according to an embodiment of the present invention with reference to the accompanying drawings, and first, a general identification method for multiple types of damages of a bridge guided by small sample cycle consistency according to an embodiment of the present invention will be described with reference to the accompanying drawings.
FIG. 1 is a flowchart of a general identification method for multiple types of damages of a bridge guided by small sample loop consistency according to an embodiment of the present invention.
As shown in fig. 1, the general identification method for multiple types of damages of a bridge guided by small sample circulation consistency comprises the following steps:
in step S1, data alignment is performed on a preset bridge damage type, and a data set after data alignment is divided into a support set and a query set.
Further, in an embodiment of the present invention, step S1 specifically includes:
step S101, defining preset bridge damage types, and carrying out data alignment on multi-type damage images of the bridge and multi-channel pixel-level marks corresponding to the multi-type damage images to form a small sample data set;
step S102, randomly dividing the small sample data set into a mutually exclusive support set and a query set.
That is, all bridge damage types considered are defined first, the total number of all damage types and image background categories is recorded as N, and data alignment is performed on the bridge multi-type damage image X and the multi-channel pixel level label Y corresponding to the bridge multi-type damage image X to form a small sample data set D. The data set D is then randomly divided into two mutually exclusive support set S and query set Q, and the number of samples included in the support set and the query set is represented by { S } and { Q } respectively.
Specifically, all bridge structure damage types to be identified and their order are specified. For example, considering three types of bridge structure damages of concrete cracks, concrete peeling and steel box girder fatigue cracks and image backgrounds, N is 4, and numerical codes 1-4 sequentially represent the backgrounds, the concrete cracks, the concrete peeling and the steel box girder fatigue cracks.
For any original image X (dimension is H multiplied by W) which possibly only contains one or a few types of bridge damage, according to the specified damage type channel sequence, splicing the pixel-level labels corresponding to the damage types along the channel direction to form a multi-channel pixel-level label image Y (dimension is H multiplied by W multiplied by N).
For a certain channel (dimension is H multiplied by W) of Y, if the channel contains a corresponding damage type, the pixel is in binary representation (1 represents that the pixel is the damage type, and 0 represents not yes); if the image does not contain the type corresponding to a certain channel, the values of the annotated image on the channel are all 0. Therefore, a data pair { X, Y } formed by any bridge damage original image and the corresponding multi-channel pixel-level annotation image can be obtained, and the set of all data pairs conforming to the form is a data set D. The data set D is randomly divided into two parts of a mutually exclusive support set S and a query set Q, and the number of samples included in the support set and the query set is respectively expressed by { S } and { Q }.
In step S2, as shown in fig. 2, a small sample cycle consistency guided bridge multi-type damage universal identification model is established by using the support set and the query set.
In step S3, a first loss value of the query set is calculated by using the bridge multi-type damage general identification model guided by the small sample cycle consistency.
That is, will branch offOriginal image XS in support set S and corresponding multi-channel pixel-level annotation image YSQuery original image X in set QQAs input, the query set Q is summed with XQCorresponding multi-channel pixel-level labeled image YQAs true values, a first fractional loss value L is calculated1。
Further, in an embodiment of the present invention, step S3 specifically includes:
step S301, inputting an original image of a support set into an encoder of a bridge multi-type damage general identification model guided by small sample circulation consistency for feature extraction to obtain a high-level feature map, and simultaneously passing a multi-channel pixel-level annotation image corresponding to the high-level feature map through a down-sampling module at the same level as the encoder to obtain a mask map;
step S302, the mask graph and the high-level feature graph are subjected to element-by-element product operation, and then global average pooling is carried out to obtain a feature vector cluster;
step S303, inputting the original image of the query set into an encoder of a bridge multi-type damage general identification model guided by small sample cycle consistency, obtaining a corresponding high-level feature map, splitting a feature vector cluster into a plurality of feature vectors according to the damage type channel direction, calculating the cosine similarity tensor of each feature vector and the high-level feature map of the original image of the query set point by point, wherein the calculation formula is as follows:
Figure BDA0003723458120000071
wherein N represents the total number of the damage type channels (including background), i is an index sequence number, and ViThe ith channel (dimension 1 XN') representing the feature vector cluster in step two(a,b)Representing a query set original image XQThe vector (dimension is 1 × N') corresponding to the high-level feature map at the plane position (a, b) obtained by the encoder, | | | | | tory2Represents a vector modulo operation, and D represents a cosine similarity tensor (dimension H '× W' × N).
Step S304, carrying out channel-by-channel splitting and channel expansion on the cosine similarity tensor according to the damage type channel direction, and carrying out element-by-element multiplication on the cosine similarity tensor and a high-level feature map of an original image of an inquiry set to obtain a plurality of corresponding damage types and feature maps; and then, sequentially passing through a decoder with an anti-symmetric structure with the encoder one by one according to the specified damage type sequence to obtain the prediction result of the original image of the query set, and calculating a first loss value with the multi-channel pixel-level label of the query set.
In step S4, a universal identification model of the bridge multi-type damage guided by the small sample cycle consistency is utilized, the cycle consistency is introduced, and a second loss value of the support set is calculated.
That is to say, the cosine similarity tensor of the multi-type damage of the bridge is obtained through the cycle consistency guide module, softmax operation is carried out in the channel direction of the damage type, and the original image X of the query set is obtainedQMarking one-hot coding at a multi-channel pixel level; then collecting the original image X by query setQAnd labeling the original image XS in the support set S as input by using the multi-channel pixel level label unique hot coding and the original image XS in the support set S as inputSCorresponding multi-channel pixel-level labeled image YSCalculating a second partial loss value L as a true value2。
Further, in an embodiment of the present invention, step S4 specifically includes:
and performing softmax operation on the cosine similarity tensor along the damage type channel direction to obtain the one-hot coding of the pixel level annotation of the original image of the query set, then taking the original image of the query set, the one-hot coding of the multi-channel pixel level annotation of the original image of the query set and the original image of the support set as input, taking the multi-channel pixel level annotation image corresponding to the original image of the support set as a real value, repeating the steps S302 to S304 to obtain a prediction result of the original image of the support set, and calculating a second loss value with the multi-channel pixel level annotation of the support set.
Wherein, the calculation formula of the one-hot code is as follows:
Figure BDA0003723458120000072
wherein N isThe total number of the damage type channels, i is an index sequence number, and the Maxindicator represents that the following operations are executed: assigning the position with the maximum value in each element of one vector as 1, and assigning the other positions as 0; exp is an exponent fetching operation;
Figure BDA0003723458120000073
a cosine similarity value corresponding to the ith channel of the cosine similarity tensor at the plane position (a, b);
Figure BDA0003723458120000074
for the one-hot encoding corresponding to the ith channel at plane location (a, b), (H ', W') is the plane of H '× W'.
Specifically, as shown in FIG. 2, the original image X of the support set SSExtracting features by an encoder to obtain a high-level feature map (dimension is H ' x W ' x N '), and obtaining a multi-channel pixel-level annotation image Y corresponding to the high-level feature mapSObtaining a mask map (dimension is H 'multiplied by W' multiplied by N) with the same plane size as the high-level feature map through a down-sampling module at the same level as an encoder;
performing element-by-element multiplication operation on the mask map (dimension is H ' × W ' × N) obtained in the previous step and the high-level feature map (dimension is H ' × W ' × N ') in the previous step one by one to remove background region information on the high-level feature map and only keep various types of damaged region information, and then performing global average pooling on an H ' × W ' plane to obtain a feature vector cluster (dimension is 1 × N ' × N) with N ' as a vector length and N as the number of channels;
query set Q raw image XQThrough the encoder (shared with the encoder parameters of the support set S), the corresponding high-level feature map (dimension H ' × W ' × N ') is obtained. Splitting the feature vector cluster (dimension is 1 XN ' × N) obtained in the previous step into N feature vectors of 1 XN ' according to the direction of the damage type channel, wherein each feature vector (dimension is 1 XN ') and the original image X of the query setQThe cosine similarity matrix (dimension is H ' × W ') is obtained by calculating the high-level characteristic diagram (dimension is H ' × W ' × N ') point by point on the H ' × W ' plane, and finally N damage classes are obtainedAnd splicing in the channel direction to obtain cosine similarity tensor (the dimension is H '× W' × N).
Splitting the cosine similarity tensor (dimension is H ' multiplied by W ' multiplied by N) obtained in the previous step channel by channel and expanding the channel (dimension is H ' multiplied by W ' multiplied by N ') according to the channel direction of the damage type, and comparing the cosine similarity tensor with the original image X of the query set in the previous stepQMultiplying the corresponding high-level feature map (with the dimension of H '× W' × N ') element by element to obtain a feature map with the dimension of H' × W '× N' corresponding to N damage types; then, the original images X of the query set are obtained through a decoder with an anti-symmetric structure with the encoder one by one according to the specified damage type sequenceQThe prediction result (dimension is H × W × N), and the multi-channel pixel level label Y of the query setQCalculating a first loss value L1。
The above process only utilizes the pixel-level labeling information of the query set to calculate the loss value, and the embodiment of the invention considers that if the model really learns the high-dimensional feature spaces of different damage types, the model can keep the cycle consistency, namely, the correct output can still be obtained after the positions of the support set and the query set are exchanged. Based on the idea, a cycle consistency module (as shown by a dashed arrow in fig. 2) is designed, which is specifically described as follows:
performing softmax operation on the cosine similarity tensor (the dimensionality is H '× W' × N) along the direction of the damage type channel to obtain an original image X of the query setQThe one-hot encoding (dimension H 'x W' x N) of the pixel level label corresponds to the mask map in step two. Then, the original image X is collected by the query setQAnd multi-channel pixel-level labeling one-hot coding and support set S original image X thereofSAs input, to support the integration of S with XSCorresponding multi-channel pixel-level labeled image YSAs a true value, repeating the previous steps to obtain a support set original image XSAnd the predicted result of support set, and support set multi-channel pixel level label YSCalculating a second loss value L2。
In step S5, the comprehensive loss of the bridge multi-type damage general identification model guided by the small sample cycle consistency is calculated based on the first loss value and the second loss value, and the model is updated through an error back propagation algorithm until the model converges.
Wherein, the comprehensive loss in the step S5 is as follows:
L＝L1+λL2
wherein, L is the comprehensive loss, L1 is the first loss value of the query set, L2 is the second loss value of the support set after considering the cycle consistency, and lambda is the proportionality coefficient and is set according to the ratio of the number of samples of the query set and the support set.
Wherein, the setting principle is as follows: if the number of samples of the support set is large and the number of samples of the query set is small, the function of the query set samples in the loss function needs to be improved, and lambda is correspondingly reduced; on the contrary, if there are fewer samples in the support set and more samples in the query set, the function of the support set samples in the loss function needs to be improved, and λ is increased accordingly.
It should be noted that, in the embodiment of the present invention, the specific forms of the encoder and the decoder in the model, the optimization algorithm and the hyper-parameter selection for model training, the specific form of the loss function, and the like are not limited. This is because the method proposed by the embodiment of the present invention can be applied to any encoder-decoder type deep convolutional neural network model architecture (e.g., U-Net, FCN, etc.), error back propagation-based optimization algorithm (e.g., RMSProp, SGD series, adam series, etc.), loss functions for different tasks (Dice loss, cross-entropy classification loss, ioU loss, focal loss, etc.). Any particular choice is but one specific form of implementation of the invention and is therefore not limited to a particular design. Different settings are adopted for an encoder, a decoder and an optimization algorithm of model training, as well as hyper-parameter selection, a loss function and the like in the model, and the method and the loss function are essentially still within the coverage range of the method and the core idea set forth in the embodiment of the invention.
The following describes a general identification of multiple types of damages of a bridge guided by small sample cycle consistency according to an embodiment of the present invention.
The general identification method for the multiple types of damage of the bridge guided by the small sample cycle consistency is applied to semantic segmentation tasks of three common bridge structure damage images, namely concrete cracks, concrete peeling and steel structure fatigue cracks. The data set comprises 200 bridge multi-type damage images, and the resolution is 512 x 512, wherein 100 images are used as a support set, and 100 images are used as a query set. The selected encoder architecture is a modified U-Net encoder and the decoder architecture is a modified U-Net decoder. Training the model by adopting an RMSProp optimization algorithm, setting the number of samples in a support set to be 3 by using hyper-parameters, setting the number of samples in a query set to be 1, setting the learning rate to be 0.01, setting the iteration number to be 20, selecting the Dice loss as loss functions in the query set and the support set, and setting the proportionality coefficient lambda to be 1/3.
The structure damage image data set is trained by using a traditional U-Net encoder and decoder framework, the same optimization algorithm and the same super-parameter setting are adopted, a loss function also adopts a Dice loss form, only a query set and a support set are not distinguished, and the method is compared with the method provided by the invention and is referred to as a traditional method in the following.
Fig. 3 is a comparison diagram of semantic segmentation recognition results of multiple types of damages (including concrete cracks, concrete peeling, steel structure fatigue cracks) of a small sample bridge by the method provided by the embodiment of the invention and the conventional method. Fig. 4 is a comparison diagram of semantic segmentation recognition accuracy of multiple types of damage to a bridge in a small sample scene by the method according to the embodiment of the present invention and the conventional method. The result shows that the general identification method for the multiple types of bridge damages guided by the small sample cycle consistency provided by the embodiment of the invention has the advantages that the identification effect is obviously better than that of the traditional method under the condition of less training samples, the false identification rate of the background is greatly reduced, the accurate identification of the multiple types of bridge damages is realized based on a unified model, and the average identification precision of the three types of bridge damages including concrete cracks, concrete peeling and steel structure fatigue cracks is respectively improved by 6.4%, 3.5% and 2.5%.
In summary, the general identification method for the multiple types of damages of the bridge guided by the small sample circulation consistency provided by the embodiment of the invention has the following beneficial effects:
(1) Aiming at the difficult problem of small samples of bridge damage image identification, images are divided into a support set and a query set, and the problem of small samples is solved by designing the cycle consistency of a model on the query set and the support set;
(2) The universal identification of different types of bridge damages by adopting a unified model is realized, and the defect that different identification models need to be established for different tasks in the traditional method is overcome;
(3) Only small sample images are used, a double supervision information utilization mode of a query set and a support set is created, and simultaneous learning of two supervision paths is achieved: firstly, taking an original image of a support set and original images of an annotation and query set as input to obtain cosine similarity tensors of high-dimensional feature spaces of multiple types of damages, and taking the annotation of the query set as a true value to calculate the loss of the query set; and then obtaining an approximate result of the query set annotation based on the query set original image and the cosine similarity tensor, taking the support set original image as input, taking the support set annotation as a true value, and calculating the loss of the support set.
The general identification system for the multi-type damage of the bridge guided by the circular consistency of the small samples is described with reference to the attached drawings.
FIG. 5 is a schematic structural diagram of a small sample loop consistency guided bridge multi-type damage universal identification system according to an embodiment of the present invention.
As shown in fig. 5, the system 10 includes: a data set partitioning module 100, a building module 200, a first loss calculation module 300, a second loss calculation module 400, and an update convergence module 500.
The data set partitioning module 100 is configured to perform data alignment on a preset bridge damage type, and partition a data set after the data alignment into a support set and a query set. The building module 200 is used for building a bridge multi-type damage universal identification model guided by small sample cycle consistency by using the support set and the query set. The first loss calculation module 300 is configured to calculate a first loss value of the query set using the small-sample cyclic consistency guided bridge multi-type damage universal identification model. The second loss calculation module 400 is configured to calculate a second loss value of the support set by using the universal identification model for multiple types of damages of the bridge guided by the small sample cycle consistency and introducing cycle consistency. The updating convergence module 500 is configured to calculate a comprehensive loss of the bridge multi-type damage general identification model guided by the small sample cycle consistency based on the first loss value and the second loss value, and update the model by using an error back propagation algorithm until the model converges.
It should be noted that the explanation of the foregoing embodiment of the universal identification method for multiple types of damages to a bridge guided by small sample cycle consistency is also applicable to the system of this embodiment, and is not repeated here.
The general identification system for the multiple types of damages of the bridge guided by the small sample circulation consistency provided by the embodiment of the invention has the following beneficial effects:
(1) Aiming at the difficult problem of small samples of bridge damage image identification, images are divided into a support set and a query set, and the problem of small samples is solved by designing the cycle consistency of a model on the query set and the support set;
(2) The universal identification of different types of bridge damages by adopting a unified model is realized, and the defect that different identification models need to be established for different tasks in the traditional method is overcome;
(3) Only small sample images are used, a double supervision information utilization mode of a query set and a support set is created, and simultaneous learning of two supervision paths is realized: firstly, taking an original image of a support set and original images of an annotation and query set as input, obtaining cosine similarity tensors of a multi-class damage high-dimensional feature space, taking the annotation of the query set as a true value, and calculating the loss of the query set; and then obtaining an approximate result of the query set annotation based on the query set original image and the cosine similarity tensor, taking the support set original image as input, taking the support set annotation as a true value, and calculating the loss of the support set.
In order to implement the foregoing embodiments, the present invention further provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the method for universally identifying multiple types of damages to a bridge guided by small sample loop consistency is implemented as in the foregoing embodiments.
In order to implement the foregoing embodiments, the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the method for universally identifying multiple types of damage to a bridge guided by small sample loop consistency as in the foregoing embodiments.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or N embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more N executable instructions for implementing steps of a custom logic function or process, and alternate implementations are included within the scope of the preferred embodiment of the present invention in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of implementing the embodiments of the present invention.
The logic and/or steps represented in the flowcharts or otherwise described herein, such as an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or N wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Further, the computer readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.
It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods may be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.
It will be understood by those skilled in the art that all or part of the steps carried out in the method for implementing the above embodiment may be implemented by hardware that is related to instructions of a program, and the program may be stored in a computer readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiment.
In addition, functional units in the embodiments of the present invention may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a separate product, may also be stored in a computer readable storage medium.
The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc. Although embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and not to be construed as limiting the present invention, and that changes, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims (10)

1. A general identification method for multiple types of damage of a bridge guided by small sample circulation consistency is characterized by comprising the following steps:
the method comprises the following steps that S1, data alignment is carried out on preset bridge damage types, and a data set after the data alignment is divided into a support set and a query set;
s2, establishing a universal identification model of the multi-type damage of the bridge guided by the small sample circulation consistency by using the support set and the query set;
s3, calculating a first loss value of the query set by using the bridge multi-type damage universal identification model guided by the small sample cycle consistency;
s4, utilizing the universal identification model of the multi-type damage of the bridge guided by the small sample cycle consistency, introducing cycle consistency, and calculating a second loss value of the support set;
and S5, calculating the comprehensive loss of the bridge multi-type damage general identification model guided by the small sample circulation consistency based on the first loss value and the second loss value, and updating the model through an error back propagation algorithm until the model converges.
2. The method for universally identifying multiple types of damages of a bridge guided by small sample cycle consistency according to claim 1, wherein the step S1 specifically comprises:
step S101, defining the preset bridge damage type, and performing data alignment on the multi-type bridge damage image and the multi-channel pixel level mark corresponding to the multi-type bridge damage image to form a small sample data set;
and step S102, randomly dividing the small sample data set into a mutually exclusive support set and a query set.
3. The method for universally identifying multiple types of damages of a bridge guided by small sample cycle consistency according to claim 1, wherein the step S3 specifically comprises:
step S301, inputting the original image of the support set into an encoder of the bridge multi-type damage general identification model guided by the small sample cycle consistency for feature extraction to obtain a high-level feature map, and simultaneously, passing a multi-channel pixel-level annotation image corresponding to the high-level feature map through a down-sampling module at the same level as the encoder to obtain a mask map;
step S302, performing element-by-element product operation on the mask graph and the high-level feature graph, and performing global average pooling to obtain a feature vector cluster;
step S303, inputting the original image of the query set into an encoder of the bridge multi-type damage general identification model guided by the small sample cycle consistency to obtain a corresponding high-level feature map, splitting the feature vector cluster into a plurality of feature vectors according to the damage type channel direction, and calculating the cosine similarity tensor of each feature vector and the high-level feature map of the original image of the query set point by point;
step S304, carrying out channel-by-channel splitting and channel expansion on the cosine similarity tensor according to the channel direction of the damage types, and carrying out element-by-element multiplication on the cosine similarity tensor and a high-level feature map of an original image of the query set to obtain a plurality of corresponding damage types and feature maps; and then, sequentially passing through a decoder with an anti-symmetric structure with the encoder one by one according to the specified damage type, obtaining the prediction result of the original image of the query set, and calculating a first loss value with the multi-channel pixel-level label of the query set.
4. The method for universally identifying multiple types of damages of a bridge guided by small sample cycle consistency according to claim 3, wherein the step S4 specifically comprises:
performing softmax operation on the cosine similarity tensor along a damage type channel direction to obtain pixel-level labeled single-hot codes of the original image of the query set, then taking the original image of the query set, the multi-channel pixel-level labeled single-hot codes thereof and the original image of the support set as input, taking the multi-channel pixel-level labeled image corresponding to the original image in the support set as a true value, repeating the steps S302 to S304 to obtain a prediction result of the original image of the support set, and calculating a second loss value with the multi-channel pixel-level label of the support set.
5. The method according to claim 3, wherein the cosine similarity tensor is computed as:
Figure FDA0003723458110000021
wherein N is the total number of the damage type channels, i is the index number, ViThe ith channel, F, being a cluster of feature vectors(a,b)Corresponding vector at plane position (a, b) for high-level feature map obtained from the original image of the query set after passing through the encoder2For vector modulo operation, D is the cosine similarity tensor, and (H ', W') is the plane of H '× W'.
6. The method for universally identifying multiple types of damages of a bridge guided by small sample cycle consistency according to claim 4, wherein the calculation formula of the one-hot code is as follows:
Figure FDA0003723458110000022
wherein, N is the total number of the damage type channels, i is the index sequence number, and Maxindicator represents that the following operations are executed: assigning the position with the maximum numerical value in each element of one vector as 1, and assigning the rest positions as 0; exp is the operation of fetching the exponent;
Figure FDA0003723458110000023
a cosine similarity value corresponding to the ith channel of the cosine similarity tensor at the plane position (a, b);
Figure FDA0003723458110000024
for the one-hot encoding corresponding to the ith channel at plane location (a, b), (H ', W') is the plane of H '× W'.
7. The method for universally identifying multiple types of damages of a bridge guided by cyclic consistency of small samples according to claim 4, wherein the comprehensive losses in the step S5 are as follows:
L＝L1+λL2
wherein, L is the comprehensive loss, L1 is the first loss value of the query set, L2 is the second loss value of the support set after considering the cycle consistency, and lambda is the proportionality coefficient, and is set according to the ratio of the sample number of the query set to the sample number of the support set.
8. A general identification system for multiple types of damages of a bridge guided by small sample circulation consistency is characterized by comprising the following steps:
the data set dividing module is used for carrying out data alignment on the preset bridge damage type and dividing the data set after the data alignment into a support set and a query set;
the construction module is used for establishing a bridge multi-type damage universal identification model guided by small sample circulation consistency by utilizing the support set and the query set;
the first loss calculation module is used for calculating a first loss value of the query set by utilizing the bridge multi-type damage universal identification model guided by the small sample cyclic consistency;
the second loss calculation module is used for calculating a second loss value of the support set by utilizing the bridge multi-type damage universal identification model guided by the small sample cycle consistency and introducing cycle consistency;
and the updating convergence module is used for calculating the comprehensive loss of the bridge multi-type damage general identification model guided by the small sample cycle consistency based on the first loss value and the second loss value, and updating the model through an error back propagation algorithm until the model converges.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the small sample loop consistency guided bridge multi-type damage universal identification method according to any one of claims 1 to 7.
10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the method for universally identifying multiple types of damage to a bridge guided by small sample loop consistency according to any one of claims 1 to 7.
CN202210758465.8A 2022-06-30 2022-06-30 Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency Active CN115272746B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210758465.8A CN115272746B (en) 2022-06-30 2022-06-30 Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210758465.8A CN115272746B (en) 2022-06-30 2022-06-30 Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency

Publications (2)

Publication Number Publication Date
CN115272746A true CN115272746A (en) 2022-11-01
CN115272746B CN115272746B (en) 2023-04-07

Family

ID=83764019

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210758465.8A Active CN115272746B (en) 2022-06-30 2022-06-30 Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency

Country Status (1)

Country Link
CN (1) CN115272746B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160352778A1 (en) * 2015-05-28 2016-12-01 International Business Machines Corporation Inferring Security Policies from Semantic Attributes
CN110490249A (en) * 2019-08-16 2019-11-22 哈尔滨工业大学 Structural Damage Identification based on attribute class relations and few sample meta learning
CN111667011A (en) * 2020-06-08 2020-09-15 平安科技（深圳）有限公司 Damage detection model training method, damage detection model training device, damage detection method, damage detection device, damage detection equipment and damage detection medium

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160352778A1 (en) * 2015-05-28 2016-12-01 International Business Machines Corporation Inferring Security Policies from Semantic Attributes
CN110490249A (en) * 2019-08-16 2019-11-22 哈尔滨工业大学 Structural Damage Identification based on attribute class relations and few sample meta learning
CN111667011A (en) * 2020-06-08 2020-09-15 平安科技（深圳）有限公司 Damage detection model training method, damage detection model training device, damage detection method, damage detection device, damage detection equipment and damage detection medium

Also Published As

Publication number Publication date
CN115272746B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
CN111696094B (en) Immunohistochemical PD-L1 membrane staining pathological section image processing method, device and equipment
Li et al. Automatic bridge crack identification from concrete surface using ResNeXt with postprocessing
CN112651406B (en) Depth perception and multi-mode automatic fusion RGB-D significance target detection method
CN114419323B (en) Cross-modal learning and domain self-adaptive RGBD image semantic segmentation method
CN115841004B (en) Strip steel hot rolling process mechanical property soft measurement method and device based on multidimensional data
CN113257357A (en) Method for predicting protein residue contact map
CN115170971A (en) Construction safety-oriented self-supervision monocular depth estimation transfer learning method and system
CN113436115B (en) Image shadow detection method based on depth unsupervised learning
Wen et al. PCDNet: Seed operation–based deep learning model for pavement crack detection on 3D asphalt surface
CN115272746B (en) Universal identification method and system for multiple types of damage of bridge guided by small sample circulation consistency
CN117171713A (en) Cross self-adaptive deep migration learning method and system based on bearing service life
CN114399901B (en) Method and equipment for controlling traffic system
CN115953578A (en) Method and device for evaluating inference result of fracture semantic segmentation model
CN115984640A (en) Target detection method, system and storage medium based on combined distillation technology
CN116007504A (en) Crack detection module, device and computer equipment based on image technology
CN112599195B (en) Gene sequence splicing method and application
CN115359484A (en) Image processing method, device, equipment and storage medium
CN114529794A (en) Infrared and visible light image fusion method, system and medium
CN111028245B (en) Multi-mode composite high-definition high-speed video background modeling method
CN113920311A (en) Remote sensing image segmentation method and system based on edge auxiliary information
Cai et al. Vonas: Network design in visual odometry using neural architecture search
CN113139332A (en) Automatic model construction method, device and equipment
CN113569660B (en) Learning rate optimization algorithm discount coefficient method for hyperspectral image classification
CN113469405B (en) Subimage segmentation-based cross-city tail gas migration prediction method, system and medium
CN114385831B (en) Knowledge-graph relation prediction method based on feature extraction

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