CN115457006B - Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation - Google Patents

Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation Download PDF

Info

Publication number
CN115457006B
CN115457006B CN202211164829.6A CN202211164829A CN115457006B CN 115457006 B CN115457006 B CN 115457006B CN 202211164829 A CN202211164829 A CN 202211164829A CN 115457006 B CN115457006 B CN 115457006B
Authority
CN
China
Prior art keywords
similarity
defect
self
consistency
defect image
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
Application number
CN202211164829.6A
Other languages
Chinese (zh)
Other versions
CN115457006A (en
Inventor
张洪涛
迟福东
郭有安
毛莺池
陈豪
万旭
字陈波
王龙宝
彭欣欣
余记远
李洪波
赵欢
余意
吴光耀
翟笠
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.)
Hohai University HHU
Huaneng Group Technology Innovation Center Co Ltd
Huaneng Lancang River Hydropower Co Ltd
Original Assignee
Hohai University HHU
Huaneng Group Technology Innovation Center Co Ltd
Huaneng Lancang River Hydropower Co Ltd
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 Hohai University HHU, Huaneng Group Technology Innovation Center Co Ltd, Huaneng Lancang River Hydropower Co Ltd filed Critical Hohai University HHU
Priority to CN202211164829.6A priority Critical patent/CN115457006B/en
Publication of CN115457006A publication Critical patent/CN115457006A/en
Application granted granted Critical
Publication of CN115457006B publication Critical patent/CN115457006B/en
Priority to DE102023125558.4A priority patent/DE102023125558A1/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/17—Terrestrial scenes taken from planes or by drones
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • 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
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30132—Masonry; Concrete
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30168—Image quality inspection
    • 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
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
    • Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Quality & Reliability (AREA)
  • Remote Sensing (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)
  • Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)

Abstract

The application discloses an unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation, which designs a self-distillation defect image classification model, adopts a self-distillation training strategy compression model to facilitate the realization of high-precision image classification work on small-sized application equipment, and mainly comprises two parts: and constructing and transferring based on the similar consistency knowledge. And obtaining a similarity matrix by calculating the correlation between the instances in the Mini-batch based on the similarity consistency knowledge construction part. And the similarity consistency knowledge transfer part transfers a similarity matrix between the self-distilled defect image classification model layers, refines the bottom-level similarity, and captures rich context scenes and local characteristic information. Aiming at the problems that an image classification model is large and complex in capacity and cannot be used on small unmanned carrier equipment, the self-distillation model is designed, and the classification efficiency and the classification precision of the defect images are improved.

Description

Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation
Technical Field
The application relates to an unmanned aerial vehicle inspection defect classification method and device based on similarity and consistency self-distillation, and belongs to the technical field of image classification.
Background
The reservoir is an important facility for flood control and disaster reduction, water supply guarantee and agricultural irrigation. The reservoir dam is used as an important water conservancy infrastructure, not only bears the difficult tasks of daily life and industrial and agricultural production of masses of people, but also makes great contribution to flood control and flood control scheduling and ensuring safety of one party, and also gives attention to irrigation, power generation and water storage. Timely and accurate identification of dam defects is a precondition for ensuring safe operation of the dam, and ensuring the safety of the reservoir dam is just the life and property safety of people. In the safety guarantee work of the reservoir dam, regular inspection and safety monitoring are important links. At present, the traditional dam defect detection method has a plurality of inconveniences, and cannot meet the development requirement of intelligent water conservancy at present.
The existing dam concrete surface defect detection task mainly relies on manual construction of an observation platform and operation of an observation instrument, and is high in subjectivity, low in efficiency, multiple in interference factors and high in labor cost. And because the observation instrument is far away from the dam body, high-precision dam body images are difficult to obtain, and cracks and defects on the dam body cannot be clearly identified. And the dam face inspection is carried out by adopting the traditional high-altitude suspension operation mode, so that the dam face inspection device has the advantages of long period, low efficiency, poor accuracy, high cost, long distance from the dam face and high safety risk.
In data classification, the existing self-distillation method has high data preprocessing cost, and in order to fully mine and utilize the data trained by the supervised student model, the training sample is subjected to operations such as rotation, cutting, overturning and the like, a large amount of data preprocessing work is needed, and the training is complex, time and calculation cost are high. In addition, in the self-distillation process, as the information of interest of each layer of the deep neural network is different, particularly under the condition of large difference of positions of the layers, a similar matrix is simulated across different layers, the characteristic information of the middle layer is ignored, the local characteristic information among the examples is easily lost, the learning target is single and one-sided, the capability of detecting the local characteristic is damaged, and the model classification precision is low.
Disclosure of Invention
The application aims to: aiming at the problems of large model capacity, high data preprocessing cost and missing local feature detection in the existing defect image classification work, the application provides an unmanned aerial vehicle inspection defect classification method and device based on similarity and consistency self-distillation, which improve the image classification performance under the scene of small-sized application equipment, assist the dam three-dimensional inspection defect detection and improve the image classification precision.
The technical scheme is as follows: a unmanned aerial vehicle inspection defect classification method based on similarity self-distillation comprises the following steps:
step 1) obtaining defect images in the three-dimensional inspection process of the dam, collecting different types of defect images on the surface of concrete, performing image preprocessing, classifying and labeling the defect images by combining the knowledge of a dam defect library, and constructing a special defect classification data set;
step 2) taking ResNet-32 as a reference network, flattening the three-dimensional feature map on feature dimensions by utilizing a soft attention mechanism, acquiring feature weight distribution of an image, forming a two-dimensional attention map, calculating correlation among the attention maps by utilizing Euler distances, obtaining a similarity matrix, and acquiring similarity consistency knowledge as a similarity consistency knowledge-based construction part in a self-distilling defect image classification model;
step 3) transferring the similarity matrix obtained in the step 2) layer by layer between the self-distilling defect image classification model layers to serve as a similarity consistency knowledge transfer part in the self-distilling defect image classification model to obtain similarity consistency loss, representing similarity consistency knowledge supervision, combining real label supervision and correcting the training direction of the self-distilling defect image classification model;
step 4) combining the similarity-based consistency knowledge construction part in the step 2) with the similarity consistency knowledge transfer part in the step 3) to construct a self-distilling defect image classification model;
step 5) inputting the defect image data in the constructed defect classification special data set into a constructed self-distilling defect image classification model, training the self-distilling defect image classification model, and performing accuracy test;
and 6) transferring the trained self-distilled defect image classification model to application equipment, and acquiring images in real time based on the application equipment to realize online defect image classification.
And migrating the trained self-distilled defect image classification model to unmanned aerial vehicle equipment, and acquiring the dam concrete surface defect image in real time based on an unmanned aerial vehicle system to realize online defect image classification.
Wherein, the construction of the defect classification special data set in the step 1) comprises the following procedures:
1-1) collecting different types of defect images on the surface of the dam concrete from the network and the data obtained by field shooting, and classifying and labeling the defect images.
1-2) screening the classified defect images, selecting the defect images meeting the definition requirement, uniformly setting the image format into jpg format, and obtaining a defect image data set without requiring image pixels and camera shooting distance.
1-3) dividing the defect image data set into a support set, a query set and a test set by adopting a random distribution mode.
In the step 2), resNet-32 is used as a reference network, a soft attention mechanism is utilized to flatten a three-dimensional feature map in a feature dimension, feature weight distribution of an image is obtained, a two-dimensional attention map is formed, correlation among the attention maps is calculated by using Euler distances, a similarity matrix is obtained, and the similarity matrix is used as a similarity-based consistency knowledge construction part in a self-distilling defect image classification model to obtain similarity consistency knowledge. The method specifically comprises the following steps:
2-1) adopting a residual network ResNet-32 as a reference network, inputting a single defect image, generating a three-dimensional feature map after each ResBlock convolution operation, flattening the three-dimensional feature map on a channel by utilizing a soft attention mechanism, and acquiring feature weight distribution of the image to obtain a two-dimensional attention map of the single image input output after each ResBlock processing, wherein the two-dimensional attention map is specifically as follows:
wherein ,for a layer of the neural network and its corresponding activation tensor, h×w represents the height and width of the tensor, and C represents the number of tensors, i.e. the number of convolution kernels. The mapping function based on the activation tensor takes the three-dimensional activation tensor as input and outputs a two-dimensional spatial attention map, i.e. the three-dimensional activation tensor is compressed into a flat two-dimensional vector in the spatial dimension. The procedure for obtaining the two-dimensional attention map is as follows: the implicit assumption is that the absolute value of a neuron activation can represent the importance of the activation, and in order to obtain the absolute value statistic of the element in the tensor a over the channel dimension C, the sum of the powers of p of the absolute values of the channel C is used, as follows:
wherein ,Ai =a (i, i), maximum, power and absolute value operations are calculated in terms of elements, a i Is the ith plane of the three-dimensional space matrix.
2-2) different areas of the network model focus on the attention map at different levels, different feature expressions are obtained, and when the Mini-batch size is b, each ResBlock outputs b attention maps with the same size.
2-3) calculating similarity of attention diagram in the same hierarchy Mini-batch by using Euler distance, inputting the Mini-batch with the size of b into a network model ResNet-32, and generating a three-dimensional feature diagram after the ith ResBlock convolution operation to beWhere b is the batch size, c i Is the ith ResBlock filterNumber of devices, h i and wi The height and width of the i-th ResBlock output feature map are shown, respectively. Flattening the feature map on the number dimension of the filter to obtain b pieces of h i ×w i Is as follows:
calculating the correlation between b attention attempts to obtain the similarity matrix between the attention attemptsM i Element->The method comprises the following steps:
wherein ,representing attention patterns of the p-th and q-th images in Mini-batch, respectively.
The obtained similarity matrix in the step 3) is transferred layer by layer between the self-distilling defect image classification model layers, so that the shallow similarity matrix is similar to the deep similarity matrix, and is used as a similarity consistency knowledge transfer part in the self-distilling defect image classification model to obtain similarity consistency loss, represent similarity consistency knowledge supervision, combine with real label supervision and correct the training direction of the self-distilling defect image classification model. The method specifically comprises the following steps:
3-1) taking ResBlock of the deep layer part as a teacher, taking Resblock of the shallow layer part as a student, guiding Mini-batch shallow layer Block to obtain attention correlation of deep layer Block, enabling shallow layer Resblock to imitate a similarity matrix of deep layer Resblock, namely enabling shallow layer similarity matrix to be similar to deep layer similarity matrix, and representing similarity consistency knowledge supervisionObtaining the similarity consistency Loss mimic The method is characterized by comprising the following steps:
wherein ,Mi Is the similarity matrix obtained by the output of the ith ResBlock.
3-2) besides similarity consistency knowledge supervision, the real label is required to be used as traditional supervision, and the training direction of the model is corrected to obtain Loss from the real label GT The method is characterized by comprising the following steps:
L(W R ，x)＝CrossEntropy(y predict ，p frue )
wherein ,WR Is a model parameter, y predict Output of softmax layer, y, which is input x true Is the true label of x.
3-3) Loss function Total Loss of the entire model is lost by similarity consistency Loss mimic And Loss from real tags GT The composition is as follows:
wherein alpha and beta are super parameters.
The self-distilling defect image classification model comprises a similarity-based consistency knowledge construction part and a similarity-based consistency knowledge transmission part, and the problems of high data preprocessing cost and missing local feature detection are respectively solved. The method comprises the steps of learning different layers of sample images based on a similarity consistency knowledge construction part to obtain feature images, obtaining attention force diagram through feature weight distribution, calculating similarity of attention force diagram among sample images in Mini-batch to obtain a similarity matrix, constructing knowledge based on similarity consistency, and obtaining additional knowledge among defect images without carrying out distortion treatment on defect image data or extracting defect image data of the same category, so that the problems of high training cost and complex training caused by a large number of defect image data preprocessing works are avoided; the similarity knowledge transfer part transfers the similarity matrix unidirectionally between the model middle layers, so that the shallow-level similarity matrix imitates the deep-level similarity matrix, the low-level similarity is refined, the more abundant context scenes and local features are captured, the problem of missing local feature detection is solved, and the classification precision is improved.
Step 5) inputting the defect image data in the constructed defect classification special data set (the defect classification special data set is divided into a support set, a query set and a test set) into a constructed self-distilling defect image classification model, training the self-distilling defect image classification model, and performing precision test, wherein the method specifically comprises the following steps:
in the self-distillation defect image classification model training process, random gradient descent SGD with momentum is used for parameter learning and updating operation, the initial learning rate is set to 0.1, the momentum is set to 0.9, the weight attenuation is 5e-4, the batch size is set to 64, the epoch is set to 100, and the average classification error rate (Top-1 and Top-5) of 5 independent experiments is adopted as a judgment basis due to randomness of a single experiment. And selecting a loss function total of the whole model as a loss function for calculating a network training error, inputting defect image data in a test set into the model, continuously adjusting parameters of the model, and finishing training of the model when the loss function value of the model is minimum.
And 6) migrating the trained self-distillation defect image classification model to unmanned aerial vehicle equipment, and acquiring the dam concrete surface defect image in real time based on an unmanned aerial vehicle system to realize online defect image classification. The unmanned aerial vehicle system includes:
6-1) a data acquisition module, which is used for acquiring and storing the defect images of the dam in the shooting range through a binocular camera carried on the unmanned aerial vehicle;
6-2) an image recognition module for receiving a defect image shot by the unmanned aerial vehicle in real time and analyzing the type information of the defect image in real time by using the constructed self-distilled defect image classification model;
6-3) a data transmission module for transmitting the image information shot by the image recognition module and the classification information thereof;
6-4) the ground station system receives the defect images shot by the unmanned aerial vehicle and the classification information after the defect images are identified, and then a flight route is established;
6-5) the flight control system receives the instruction sent by the ground station and controls the unmanned aerial vehicle to shoot, identify and transmit the area to be identified.
Unmanned aerial vehicle inspection defect classification device based on similarity uniformity is from distillation, includes following content:
constructing a special defect classification data set module, acquiring common defect images in the three-dimensional inspection process of the dam, acquiring different types of defect images on the surface of the concrete, performing image preprocessing, marking the defect images by combining the knowledge of a dam defect library, and constructing a special defect classification data set;
based on a similarity consistency knowledge construction module, resNet-32 is used as a reference network, a soft attention mechanism is utilized to flatten a three-dimensional feature map on feature dimensions, feature weight distribution of images is obtained, a two-dimensional attention map is formed, correlation among the attention maps is calculated by using Euler distances, a similarity matrix is obtained, and the similarity matrix is used as a similarity consistency knowledge construction part in a self-distilling defect image classification model, so that similarity consistency knowledge is obtained;
the similarity consistency knowledge transfer module transfers the similarity matrix obtained in the similarity consistency knowledge construction module layer by layer between the self-distilling defect image classification model layers to be used as a similarity consistency knowledge transfer part in the self-distilling defect image classification model to obtain similarity consistency loss, represents similarity consistency knowledge supervision, combines with real label supervision and corrects the training direction of the self-distilling defect image classification model;
the self-distilling defect image classification model module is constructed, and a similarity-based consistency knowledge construction part in the similarity-based consistency knowledge construction module is combined with a similarity consistency knowledge transmission part in the similarity consistency knowledge transmission module to construct a self-distilling defect image classification model;
the self-distilling defect image classification model training module is used for inputting the defect image data in the constructed defect classification special data set into the constructed self-distilling defect image classification model, training the self-distilling defect image classification model and performing accuracy test;
the application equipment loads the trained self-distilling defect image classification model, acquires the dam concrete surface defect image in real time, and inputs the self-distilling defect image classification model to realize online defect image classification.
The application equipment is an unmanned aerial vehicle system, the unmanned aerial vehicle loads a trained self-distillation defect image classification model, and the unmanned aerial vehicle collects the defect images in real time to realize online image classification.
The unmanned aerial vehicle system includes:
the data acquisition module is used for carrying the binocular camera on the unmanned aerial vehicle through the cradle head, acquiring the defect image of the dam in the shooting range by the camera and temporarily storing the defect image;
the image recognition module is used for receiving the defect image shot by the unmanned aerial vehicle in real time and analyzing the type information of the defect image in real time by utilizing the constructed self-distilled defect image classification model;
the data transmission module is used for transmitting the image information shot by the image recognition module and the classification information thereof;
the ground station system receives the defect images shot by the unmanned aerial vehicle and the classification information identified by the defect images, and then establishes a flight route;
and the flight control system receives the instruction sent by the ground station and controls the unmanned aerial vehicle to shoot, identify and transmit the area to be identified.
The implementation process and method of the device module are the same.
A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the unmanned aerial vehicle inspection defect classification method based on similar consistency self-distillation as described above when executing the computer program.
A computer-readable storage medium storing a computer program for executing the unmanned aerial vehicle inspection defect classification method based on similarity-consistency self-distillation as described above.
The beneficial effects are that: compared with the prior art, the self-distilling defect image classification model is designed aiming at the characteristics of complex model, large capacity, high image data preprocessing cost and partial feature detection missing in the image classification process in the prior art, the model is compressed so as to be conveniently used on small application equipment, the problems of high training cost, complex training and partial feature detection missing caused by a large amount of data preprocessing work in the traditional self-distilling method are solved, and the classification precision of the defect image classification model can be improved.
Drawings
FIG. 1 is a flow chart of a method according to an embodiment of the present application;
FIG. 2 is a general frame diagram of a self-distillation based on similarity consistency in accordance with an embodiment of the present application.
Detailed Description
The present application is further illustrated below in conjunction with specific embodiments, it being understood that these embodiments are meant to be illustrative of the application and not limiting the scope of the application, and that modifications of the application, which are equivalent to those skilled in the art to which the application pertains, fall within the scope of the application defined in the appended claims after reading the application.
As shown in fig. 1, an unmanned aerial vehicle inspection defect classification method based on similarity and consistency self-distillation comprises the following specific implementation processes:
step 1: collecting defect images of the surface of the dam concrete by different means such as photos, defect databases, network searching and the like, mainly selecting 8 types of defect images of cracks, seepage, pits, landslide, collapse, concrete damage, subsidence and deformation, taking the collected defect images of the surface of the dam concrete as sample data, screening a sample data set one by one, eliminating blurred and unclear-focus defect images, uniformly setting an image format into a jpg format, and classifying and marking the defect images according to the 8 types without image pixels and camera shooting distances to obtain a defect image data set; dividing the 8-type defect image set into 3 non-overlapped type subsets in a random distribution mode, wherein the 3-type defect image data set is used as a support set, the 2-type defect image data set is used as a query set, and the 3-type defect image data set is used as a test set.
Step 2: knowledge based on similarity consistency is constructed for reducing data preprocessing costs.
Specifically, step 2 includes:
adopting a residual network ResNet-32 as a reference network, inputting a single defect image, generating a three-dimensional feature map after each ResBlock convolution operation, flattening the three-dimensional feature map on a channel by using a soft attention mechanism, acquiring feature weight distribution of the image, and obtaining a two-dimensional attention map, wherein the mechanism passes through a mapping functionPerforming an operation in which->For a layer of the neural network and its corresponding activation tensor, h×w represents the height and width of the tensor, and C represents the number of tensors, i.e. the number of convolution kernels. The absolute value of the activation tensor may represent the importance of the activation, the absolute value statistic of the elements in tensor A over the channel dimension C, using the sum of the p powers of the absolute values of channel C, through the attention mapping function->Operate with A i =a (i, i). When the Mini-batch size is b, after the first ResBlock convolution operation, each ResBlock outputs b attention patterns with the same size, and the Euler distance is utilized to calculate the similarity between the b attention patterns in the same level Mini-batch to obtain a similarity matrix between the attention patternsM i Element->By function->Obtained by->Representing attention patterns of the p-th and q-th images in Mini-batch, respectively.
Step 3: and transmitting the similarity consistency knowledge, transmitting the obtained similarity matrix layer by layer between the self-distilled defect image classification model layers to obtain similarity consistency loss, representing similarity consistency knowledge supervision, and correcting the training direction of the self-distilled defect image classification model by combining with real label supervision.
Specifically, step 3 includes:
the shallow similarity matrix is similar to the deep similarity matrix, and the similarity consistency knowledge supervision is represented, and the similarity consistency Loss is calculated mimic By a function ofCalculated, where M i Is the similarity matrix obtained by the output of the ith ResBlock. Besides similarity consistency knowledge supervision, a real label is required to be used as a traditional supervision, the training direction of a model is corrected, and Loss of the real label is reduced GT By a function L (W R ，x)＝CrossEntropy(y predict ，y true ) Calculated, where W R Is a model parameter, y predict Output of softmax layer, y, which is input x true Is the true label of x. Loss function total pass function of whole modelAnd calculating to obtain alpha and beta which are super parameters.
Step 4: and (3) transmitting the similarity matrix constructed in the step (2) as similarity consistency knowledge in the step (3) and constructing a self-distilling defect image classification model. The self-distilling defect image classification model comprises a similarity-based consistency knowledge construction part and a similarity consistency knowledge transfer part. Wherein step 2 belongs to a knowledge construction part based on similarity consistency, and step 3 belongs to a knowledge transmission part based on similarity consistency. Based on the similarity consistency knowledge construction part, learning different layers of the sample image to obtain a feature map, acquiring an attention map through feature weight distribution, and calculating similarity of the attention map to obtain a similarity matrix, so that the data preprocessing cost is reduced; the similarity consistency knowledge transfer part transfers the similarity matrix in one way between model middle layers, captures rich context scenes and local features, solves the problem of missing local feature detection, and improves the classification precision of the defect images.
Step 5: inputting the defect image data in the constructed defect classification special data set into a constructed self-distilled defect image classification model, training the self-distilled defect image classification model, and performing accuracy test, wherein the method specifically comprises the following steps of:
in the network and model training process, random gradient descent SGD with momentum is used for parameter learning and updating operation, the initial learning rate is set to 0.1, the momentum is set to 0.9, the weight attenuation is set to 64, the epoch is set to 100, and the average classification error rate (Top-1 and Top-5) of 5 independent experiments is adopted as a judgment basis due to randomness of a single experiment. And selecting a loss function Totalloss as a loss function for calculating a network training error, inputting the defect image data in the test set into the model, continuously adjusting parameters of the model according to the classification accuracy, and training the whole network until convergence to obtain the final self-distilling defect image classification model.
Step 6: migrating the trained self-distilled defect image classification model to unmanned aerial vehicle equipment, collecting defect images in real time based on an unmanned aerial vehicle system, and realizing online defect image classification, wherein the method specifically comprises the following steps of:
the ground station system and the unmanned aerial vehicle flight control system are connected in a control mode;
the command input module of the ground station system inputs a control command corresponding to the unmanned aerial vehicle;
the unmanned aerial vehicle flight control system receives the instruction of the ground station system to trigger the corresponding function and control the unmanned aerial vehicle to fly;
in the flight process of the unmanned aerial vehicle, a binocular camera carried on the unmanned aerial vehicle through a cradle head collects a defect image in a shooting range, and the collected defect image is transmitted to a flight control system;
the flight control system inputs the collected defect images into a self-distilled defect image classification model, outputs a defect image classification result from the self-distilled defect image classification model, and transmits the classification result and an original image to the ground station system;
the data processing model of the ground station system receives the classification result information, analyzes the classification result information and displays the classification result information on a user interface.
Unmanned aerial vehicle inspection defect classification device based on similarity uniformity is from distillation, includes following content:
constructing a special defect classification data set module, acquiring common defect images in the three-dimensional inspection process of the dam, acquiring different types of defect images on the surface of the concrete, performing image preprocessing, marking the defect images by combining the knowledge of a dam defect library, and constructing a special defect classification data set;
based on a similarity consistency knowledge construction module, resNet-32 is used as a reference network, a soft attention mechanism is utilized to flatten a three-dimensional feature map on feature dimensions, feature weight distribution of images is obtained, a two-dimensional attention map is formed, correlation among the attention maps is calculated by using Euler distances, a similarity matrix is obtained, and the similarity matrix is used as a similarity consistency knowledge construction part in a self-distilling defect image classification model, so that similarity consistency knowledge is obtained;
the similarity consistency knowledge transfer module transfers the similarity matrix obtained based on the similarity consistency knowledge construction module layer by layer between the self-distilled defect image classification model layers, so that the shallow similarity matrix is similar to the deep similarity matrix, and is used as a similarity consistency knowledge transfer part in the self-distilled defect image classification model to obtain similarity consistency loss, represent similarity consistency knowledge supervision, combine with real label supervision and correct the training direction of the self-distilled defect image classification model;
the self-distilling defect image classification model module is constructed, and a similarity-based consistency knowledge construction part in the similarity-based consistency knowledge construction module is combined with a similarity consistency knowledge transmission part in the similarity consistency knowledge transmission module to construct a self-distilling defect image classification model;
the self-distilling defect image classification model training module is used for inputting the defect image data in the constructed defect classification special data set into the constructed self-distilling defect image classification model, training the self-distilling defect image classification model and performing accuracy test;
and the unmanned aerial vehicle system loads a trained self-distilled defect image classification model, and the unmanned aerial vehicle acquires the defect images in real time to realize online defect image classification.
The unmanned aerial vehicle system further comprises:
the data acquisition module is used for carrying the binocular camera on the unmanned aerial vehicle through the cradle head, acquiring the defect image in the shooting range by the camera and temporarily storing the defect image;
the image recognition module is used for receiving the defect image shot by the unmanned aerial vehicle in real time and analyzing the type information of the defect image in real time by utilizing the constructed self-distilled defect image classification model;
the data transmission module is used for transmitting the image information shot by the image recognition module and the classification information thereof;
the ground station system receives the defect images shot by the unmanned aerial vehicle and the classification information identified by the defect images, and then establishes a flight route;
and the flight control system receives the instruction sent by the ground station and controls the unmanned aerial vehicle to shoot, identify and transmit the area to be identified.
It will be apparent to those skilled in the art that the steps of the method for classifying a unmanned aerial vehicle inspection defect based on self-distillation of similar consistency or the modules of the unmanned aerial vehicle inspection defect classification device based on self-distillation of similar consistency may be implemented by a general purpose computing device, they may be concentrated on a single computing device or distributed over a network of computing devices, or they may alternatively be implemented with program code executable by a computing device, so that they may be stored in a storage device for execution by the computing device and, in some cases, the steps shown or described may be performed in a different order than herein, or they may be fabricated separately as individual integrated circuit modules, or a plurality of modules or steps of them may be fabricated as a single integrated circuit module. Thus, embodiments of the application are not limited to any specific combination of hardware and software.

Claims (8)

1. The unmanned aerial vehicle inspection defect classification method based on similarity and consistency self-distillation is characterized by comprising the following steps of:
step 1) obtaining a defect image in the three-dimensional inspection process of the dam, collecting different types of defect images on the surface of the concrete, performing image preprocessing, classifying and labeling the defect images, and constructing a special data set for defect classification;
step 2) taking ResNet-32 as a reference network, flattening the three-dimensional feature map on feature dimensions by utilizing a soft attention mechanism, acquiring feature weight distribution of an image, forming a two-dimensional attention map, calculating correlation among the attention maps by utilizing Euler distances, obtaining a similarity matrix, and acquiring similarity consistency knowledge as a similarity consistency knowledge-based construction part in a self-distilling defect image classification model;
step 3) transferring the similarity matrix obtained in the step 2) layer by layer between the self-distilling defect image classification model layers to serve as a similarity consistency knowledge transfer part in the self-distilling defect image classification model to obtain similarity consistency loss, representing similarity consistency knowledge supervision, combining real label supervision and correcting the training direction of the self-distilling defect image classification model;
the similarity consistency loss is specifically as follows:
wherein ,Mi Is the similarity matrix obtained by outputting the ith ResBlock;
step 4) combining the similarity-based consistency knowledge construction part in the step 2) with the similarity consistency knowledge transfer part in the step 3) to construct a self-distilling defect image classification model;
step 5) inputting the defect image data in the constructed defect classification special data set into a constructed self-distilling defect image classification model, training the self-distilling defect image classification model, and performing accuracy test;
step 6) transferring the trained self-distilled defect image classification model to application equipment, and acquiring images in real time based on the application equipment to realize online defect image classification;
step 2) comprises the following steps:
2-1) adopting a residual network ResNet-32 as a reference network, inputting a single defect image, generating a three-dimensional feature map after each ResBlock convolution operation, flattening the three-dimensional feature map on a channel by utilizing a soft attention mechanism, and acquiring feature weight distribution of the image to obtain a two-dimensional attention map of the single image input output after each ResBlock processing, wherein the two-dimensional attention map is specifically as follows:
wherein ,for a certain layer of the neural network and the corresponding activation tensor, H multiplied by W represents the height and width of the tensor, and C represents the number of tensors, namely the number of convolution kernels; based on the mapping function of the activation tensor, taking the three-dimensional activation tensor as input, outputting a two-dimensional space attention map, namely compressing the three-dimensional activation tensor into a flat two-dimensional vector in the space dimension; the procedure for obtaining the two-dimensional attention map is as follows: the implicit assumption is that the absolute value of a neuron activation can represent the importance of the activation, and in order to obtain the absolute value statistic of the element in the tensor a over the channel dimension C, the sum of the powers of p of the absolute values of the channel C is used, as follows:
wherein ,Ai =a (i, i), maximum, power and absolute value operations are in terms of elementsCalculated, A i Is the ith plane of the three-dimensional space matrix;
2-2) different areas of the network model focusing attention patterns at different levels, obtaining different feature expressions, and when the Mini-batch size is b, outputting b attention patterns with the same size by each ResBlock;
2-3) calculating similarity of attention diagram in the same hierarchy Mini-batch by using Euler distance, inputting the Mini-batch with the size of b into a network model ResNet-32, and generating a three-dimensional feature diagram after the ith ResBlock convolution operation to beWhere b is the batch size, c i Is the number of the ith ResBlock filter, h i and wi Respectively representing the height and width of an ith ResBlock output characteristic diagram; flattening the feature map on the number dimension of the filter to obtain b pieces of h i ×w i Is as follows:
calculating the correlation between b attention attempts to obtain the similarity matrix between the attention attemptsM i Elements of (a)The method comprises the following steps:
wherein ,respectively represent the p-th and q-th images in Mini-batchAttention is paid to the force diagram.
2. The unmanned aerial vehicle inspection defect classification method based on similarity self-distillation according to claim 1, wherein the trained self-distillation defect image classification model is migrated to unmanned aerial vehicle equipment, and dam concrete surface defect images are acquired in real time based on an unmanned aerial vehicle system, so that online defect image classification is realized.
3. The unmanned aerial vehicle inspection defect classification method based on the similarity self-distillation according to claim 1, wherein the defect classification dedicated data set construction of step 1) comprises the following procedures:
1-1) collecting different types of defect images on the surface of dam concrete from network and data obtained by field shooting, and classifying and labeling the defect images;
1-2) screening the classified defect images, selecting the defect images meeting the definition requirement, and unifying image formats to obtain a defect image data set;
1-3) dividing the defect image data set into a support set, a query set and a test set by adopting a random distribution mode.
4. The unmanned aerial vehicle inspection defect classification method based on the similarity self-distillation according to claim 1, wherein the obtained similarity matrix in the step 3) is transferred layer by layer between the self-distillation defect image classification model layers, so that the shallow similarity matrix is similar to the deep similarity matrix, and is used as a similarity consistency knowledge transfer part in the self-distillation defect image classification model to obtain similarity consistency loss, represent similarity consistency knowledge supervision, and combine with real label supervision to correct the training direction of the self-distillation defect image classification model; the method specifically comprises the following steps:
3-1) taking ResBlock of a deep layer part as a teacher, taking ResBlock of a shallow layer part as a student, guiding Mini-batch shallow layer Block to obtain attention correlation of deep layer Block, enabling shallow layer Resblock to imitate a similarity matrix of the deep layer Resblock, and obtaining similarity consistency Loss mimic ；
3-2) taking the real label as the traditional supervision, and correcting the training direction of the model to obtain Loss from the real label GT ；
3-3) Loss function Total Loss of the entire model is lost by similarity consistency Loss mimic And Loss from real tags GT Composition is prepared.
5. The unmanned aerial vehicle inspection defect classification method based on similar consistency self-distillation according to claim 1, wherein the step 6) is to migrate the trained self-distillation defect image classification model to unmanned aerial vehicle equipment, collect dam concrete surface defect images in real time based on an unmanned aerial vehicle system, and realize online defect image classification; the unmanned aerial vehicle system includes:
6-1) a data acquisition module, which is used for acquiring and storing the defect images of the dam in the shooting range through a binocular camera carried on the unmanned aerial vehicle;
6-2) an image recognition module for receiving a defect image shot by the unmanned aerial vehicle in real time and analyzing the type information of the defect image in real time by using the constructed self-distilled defect image classification model;
6-3) a data transmission module for transmitting the image information shot by the image recognition module and the classification information thereof;
6-4) the ground station system receives the defect images shot by the unmanned aerial vehicle and the classification information after the defect images are identified, and then a flight route is established;
6-5) the flight control system receives the instruction sent by the ground station and controls the unmanned aerial vehicle to shoot, identify and transmit the area to be identified.
6. Unmanned aerial vehicle inspection defect classification device based on similarity uniformity is from distillation, which is characterized by comprising the following contents:
constructing a special defect classification data set module, acquiring common defect images in the three-dimensional inspection process of the dam, acquiring different types of defect images on the surface of the concrete, performing image preprocessing, marking the defect images by combining the knowledge of a dam defect library, and constructing a special defect classification data set;
based on a similarity consistency knowledge construction module, resNet-32 is used as a reference network, a soft attention mechanism is utilized to flatten a three-dimensional feature map on feature dimensions, feature weight distribution of images is obtained, a two-dimensional attention map is formed, correlation among the attention maps is calculated by using Euler distances, a similarity matrix is obtained, and the similarity matrix is used as a similarity consistency knowledge construction part in a self-distilling defect image classification model, so that similarity consistency knowledge is obtained; the specific implementation process is as follows:
2-1) adopting a residual network ResNet-32 as a reference network, inputting a single defect image, generating a three-dimensional feature map after each ResBlock convolution operation, flattening the three-dimensional feature map on a channel by using a soft attention mechanism, and acquiring feature weight distribution of the image to obtain a two-dimensional attention map output by the single image input after each ResBlock processing;
2-2) different areas of the network model focusing attention diagram at different levels to obtain different feature expressions;
2-3) calculating the similarity of attention force diagrams in the same level Mini-batch by using Euler distances to obtain a similarity matrix between the attention force diagrams; similarity matrix between attention patternsM i Element->The method comprises the following steps:
wherein ,representing attention patterns of the p-th and q-th images in the Mini-batch respectively;
the similarity consistency knowledge transfer module transfers the similarity matrix obtained in the similarity consistency knowledge construction module layer by layer between the self-distilling defect image classification model layers to be used as a similarity consistency knowledge transfer part in the self-distilling defect image classification model to obtain similarity consistency loss, represents similarity consistency knowledge supervision, combines with real label supervision and corrects the training direction of the self-distilling defect image classification model; the similarity consistency loss is specifically as follows:
wherein ,Mi Is the similarity matrix obtained by outputting the ith ResBlock;
the self-distilling defect image classification model module is constructed, and a similarity-based consistency knowledge construction part in the similarity-based consistency knowledge construction module is combined with a similarity consistency knowledge transmission part in the similarity consistency knowledge transmission module to construct a self-distilling defect image classification model;
and the self-distilling defect image classification model training module is used for inputting the defect image data in the constructed defect classification special data set into the constructed self-distilling defect image classification model, training the self-distilling defect image classification model and performing accuracy test.
7. A computer device, characterized by: the computer equipment comprises a memory, a processor and a computer program which is stored in the memory and can run on the processor, wherein the processor realizes the unmanned aerial vehicle inspection defect classification method based on the similarity self-distillation according to any one of claims 1 to 5 when executing the computer program.
8. A computer-readable storage medium, characterized by: the computer readable storage medium stores a computer program for performing the unmanned aerial vehicle inspection defect classification method based on similarity-based self-distillation as claimed in any one of claims 1 to 5.
CN202211164829.6A 2022-09-23 2022-09-23 Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation Active CN115457006B (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202211164829.6A CN115457006B (en) 2022-09-23 2022-09-23 Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation
DE102023125558.4A DE102023125558A1 (en) 2022-09-23 2023-09-20 A method and apparatus for classifying defects in drone inspection based on self-distillation of similarity consistency

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211164829.6A CN115457006B (en) 2022-09-23 2022-09-23 Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation

Publications (2)

Publication Number Publication Date
CN115457006A CN115457006A (en) 2022-12-09
CN115457006B true CN115457006B (en) 2023-08-22

Family

ID=84306526

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211164829.6A Active CN115457006B (en) 2022-09-23 2022-09-23 Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation

Country Status (2)

Country Link
CN (1) CN115457006B (en)
DE (1) DE102023125558A1 (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116416456B (en) * 2023-01-13 2023-10-24 北京数美时代科技有限公司 Self-distillation-based image classification method, system, storage medium and electronic device
CN115937217B (en) * 2023-03-09 2023-05-12 安格利(成都)仪器设备有限公司 Method for automatically detecting corrosion pits by using 3D convolution and correlation operator
CN116385794B (en) * 2023-04-11 2024-04-05 河海大学 Robot inspection defect classification method and device based on attention flow transfer mutual distillation

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019240964A1 (en) * 2018-06-12 2019-12-19 Siemens Aktiengesellschaft Teacher and student based deep neural network training
CN112036231A (en) * 2020-07-10 2020-12-04 武汉大学 Vehicle-mounted video-based lane line and road surface indication mark detection and identification method
WO2021023202A1 (en) * 2019-08-07 2021-02-11 交叉信息核心技术研究院(西安)有限公司 Self-distillation training method and device for convolutional neural network, and scalable dynamic prediction method
CN114022727A (en) * 2021-10-20 2022-02-08 之江实验室 Deep convolution neural network self-distillation method based on image knowledge review
CN114281093A (en) * 2021-12-27 2022-04-05 国网江苏省电力有限公司泰州供电分公司 Defect detection system and method based on unmanned aerial vehicle power inspection
CN114596546A (en) * 2022-01-12 2022-06-07 盛视科技股份有限公司 Vehicle weight recognition method and device, computer and readable storage medium
CN114863248A (en) * 2022-03-02 2022-08-05 武汉大学 Image target detection method based on deep supervision self-distillation
CN114913150A (en) * 2022-05-11 2022-08-16 河海大学 Intelligent identification method for concrete dam defect time sequence image

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180268292A1 (en) * 2017-03-17 2018-09-20 Nec Laboratories America, Inc. Learning efficient object detection models with knowledge distillation
CA3076424A1 (en) * 2019-03-22 2020-09-22 Royal Bank Of Canada System and method for knowledge distillation between neural networks
CN111767711B (en) * 2020-09-02 2020-12-08 之江实验室 Compression method and platform of pre-training language model based on knowledge distillation
JP7283835B2 (en) * 2020-12-17 2023-05-30 之江実験室 Automatic Compression Method and Platform for Pre-trained Language Models Based on Multilevel Knowledge Distillation
US20220261593A1 (en) * 2021-02-16 2022-08-18 Nvidia Corporation Using neural networks to perform object detection, instance segmentation, and semantic correspondence from bounding box supervision
US20220036194A1 (en) * 2021-10-18 2022-02-03 Intel Corporation Deep neural network optimization system for machine learning model scaling

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019240964A1 (en) * 2018-06-12 2019-12-19 Siemens Aktiengesellschaft Teacher and student based deep neural network training
WO2021023202A1 (en) * 2019-08-07 2021-02-11 交叉信息核心技术研究院(西安)有限公司 Self-distillation training method and device for convolutional neural network, and scalable dynamic prediction method
CN112036231A (en) * 2020-07-10 2020-12-04 武汉大学 Vehicle-mounted video-based lane line and road surface indication mark detection and identification method
CN114022727A (en) * 2021-10-20 2022-02-08 之江实验室 Deep convolution neural network self-distillation method based on image knowledge review
CN114281093A (en) * 2021-12-27 2022-04-05 国网江苏省电力有限公司泰州供电分公司 Defect detection system and method based on unmanned aerial vehicle power inspection
CN114596546A (en) * 2022-01-12 2022-06-07 盛视科技股份有限公司 Vehicle weight recognition method and device, computer and readable storage medium
CN114863248A (en) * 2022-03-02 2022-08-05 武汉大学 Image target detection method based on deep supervision self-distillation
CN114913150A (en) * 2022-05-11 2022-08-16 河海大学 Intelligent identification method for concrete dam defect time sequence image

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于深度学习的车道线提取研究;张新宇;《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》(第01期);第C035-260页 *

Also Published As

Publication number Publication date
DE102023125558A1 (en) 2024-03-28
CN115457006A (en) 2022-12-09

Similar Documents

Publication Publication Date Title
CN115457006B (en) Unmanned aerial vehicle inspection defect classification method and device based on similarity consistency self-distillation
CN105825511B (en) A kind of picture background clarity detection method based on deep learning
CN111814661B (en) Human body behavior recognition method based on residual error-circulating neural network
CN111507378A (en) Method and apparatus for training image processing model
CN107392901A (en) A kind of method for transmission line part intelligence automatic identification
CN110163187A (en) Remote road traffic sign detection recognition methods based on F-RCNN
CN110852347A (en) Fire detection method using improved YOLO v3
CN108021947A (en) A kind of layering extreme learning machine target identification method of view-based access control model
CN105550709A (en) Remote sensing image power transmission line corridor forest region extraction method
CN110533086A (en) The semi-automatic mask method of image data
CN110399908A (en) Classification method and device based on event mode camera, storage medium, electronic device
CN115359353A (en) Flower identification and classification method and device
CN113435254A (en) Sentinel second image-based farmland deep learning extraction method
CN114492634B (en) Fine granularity equipment picture classification and identification method and system
CN116486408A (en) Cross-domain semantic segmentation method and device for remote sensing image
CN115222754A (en) Mirror image segmentation method based on knowledge distillation and antagonistic learning
Peng et al. RSBNet: One-shot neural architecture search for a backbone network in remote sensing image recognition
CN115188066A (en) Moving target detection system and method based on cooperative attention and multi-scale fusion
CN114359578A (en) Application method and system of pest and disease damage identification intelligent terminal
Sridharan et al. Deep learning-based ensemble model for classification of photovoltaic module visual faults
CN117876874A (en) Forest fire detection and positioning method and system based on high-point monitoring video
CN104615987B (en) A kind of the wreckage of an plane intelligent identification Method and system based on error-duration model neutral net
Song et al. PDD: Post-Disaster Dataset for Human Detection and Performance Evaluation
CN116385794B (en) Robot inspection defect classification method and device based on attention flow transfer mutual distillation
Jha et al. Classification and Health Prediction in Plants Using Deep Convolutional Neural Networks

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