CN112100419B - Single weather image recognition method and system based on image retrieval - Google Patents
Single weather image recognition method and system based on image retrieval Download PDFInfo
- Publication number
- CN112100419B CN112100419B CN202010967561.4A CN202010967561A CN112100419B CN 112100419 B CN112100419 B CN 112100419B CN 202010967561 A CN202010967561 A CN 202010967561A CN 112100419 B CN112100419 B CN 112100419B
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- vector
- training
- feedback information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 41
- 239000013598 vector Substances 0.000 claims abstract description 150
- 238000012549 training Methods 0.000 claims abstract description 141
- 238000003860 storage Methods 0.000 claims abstract description 91
- 238000012790 confirmation Methods 0.000 claims description 71
- 238000012937 correction Methods 0.000 claims description 66
- 230000006870 function Effects 0.000 claims description 30
- 238000012360 testing method Methods 0.000 claims description 26
- 238000007781 pre-processing Methods 0.000 claims description 20
- 238000012795 verification Methods 0.000 claims description 16
- 238000000605 extraction Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 10
- 238000004422 calculation algorithm Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 230000015572 biosynthetic process Effects 0.000 claims description 6
- 238000006243 chemical reaction Methods 0.000 claims description 6
- 238000011478 gradient descent method Methods 0.000 claims description 6
- 238000010200 validation analysis Methods 0.000 claims description 6
- 230000000875 corresponding effect Effects 0.000 description 40
- 238000013527 convolutional neural network Methods 0.000 description 9
- 230000009286 beneficial effect Effects 0.000 description 6
- 238000001514 detection method Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 230000002159 abnormal effect Effects 0.000 description 4
- 230000002596 correlated effect Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 239000000463 material Substances 0.000 description 4
- 238000012216 screening Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 238000013499 data model Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 206010039203 Road traffic accident Diseases 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/55—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/56—Information retrieval; Database structures therefor; File system structures therefor of still image data having vectorial format
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Library & Information Science (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a single weather image recognition method and a system based on image retrieval, wherein the method comprises the following steps: step S1, acquiring a training image, and training a preset initial model based on the training image to acquire a retrieval model; storing the first vector features of the training image to a feature store during training; s2, inputting the single search image into a search model, and extracting features to obtain second vector features; and based on the second vector feature, retrieving a corresponding third vector feature from the feature storage library, and outputting weather description of the training image corresponding to the third vector feature as a recognition result. The invention relates to a single weather image recognition method based on image retrieval, which is based on a feature storage library established during model training, realizes the rapid recognition of a single weather image and improves the recognition rate of the single image.
Description
Technical Field
The invention relates to the technical field of image processing and pattern recognition, in particular to a single weather image recognition method and system based on image retrieval.
Background
At present, weather identification in a natural scene plays an important role in the intelligent traffic field, and the improvement of the accuracy of the weather identification can effectively improve the traffic and transportation efficiency. Single image recognition refers to that only one image is input into a model, and recognition is performed based on the content of the image completely. The traditional weather image recognition method utilizes some priori information in the images to classify, combines the characteristics of different information domains into multidimensional characteristics and sends the multidimensional characteristics into a support vector machine to judge classification. Unlike conventional recognition methods, the weather recognition based on deep learning has better recognition effects, such as recognition methods based on image semantic segmentation and image classification. As the progressive progress of technology has promoted the gradual increase in image recognition effect, image recognition has become a popular problem in recent years.
Image recognition methods based on image retrieval have proven to be very effective methods, and are widely used because image retrieval can effectively calculate the similarity between the features of a query graph and the features of images in a repository, and further obtain an output result according to the magnitude of the similarity. However, in the application field of weather identification in a natural scene at present, an image identification method based on image retrieval is not involved, but the identification rate of the existing identification method is not high, so that the problem to be solved is urgent.
Disclosure of Invention
The invention aims to provide a single weather image recognition method based on image retrieval, which is based on a feature storage library established during model training to realize the rapid recognition of a single weather image and improve the recognition rate of the single image.
The embodiment of the invention provides a single weather image identification method based on image retrieval, which comprises the following steps:
step S1, acquiring a training image, and training a preset initial model based on the training image to acquire a retrieval model; storing the first vector features of the training image to a feature store during training;
s2, inputting the single search image into a search model, and extracting features to obtain second vector features; and based on the second vector feature, retrieving a corresponding third vector feature from the feature storage library, and outputting weather description of the training image corresponding to the third vector feature as a recognition result.
Preferably, step S1, a training image is obtained, and a preset initial model is trained based on the training image to obtain a retrieval model; storing the first vector features of the training image to a feature store during training; comprising the following steps:
step S11: preprocessing the training image;
Step S12: extracting features of a training image by using a residual network structure ResNet50, performing array conversion on a training tag list, simultaneously calculating final loss by combining metric loss and cross entropy loss functions, and optimizing network parameters by using a batch random gradient descent method to minimize the loss functions;
step S13, in each training iteration, a first vector feature formation queue extracted from a current batch-size image is introduced into a feature storage library, and when the feature storage library is full, corresponding old vector features in the feature storage library are removed;
step S14: and (3) in the iteration, matching the first vector features of the current batch-size image with all third vector features in a feature memory bank to form a large number of sample pairs, calculating the Similarity by using a metric Loss function Multi-Similarity Loss, and calculating the classification Loss of the current batch image by using a cross entropy Loss function.
Preferably, step S2, inputting a single search image into a search model, and extracting features to obtain second vector features; retrieving a corresponding third vector feature from the feature repository based on the second vector feature, outputting a weather description of the training image corresponding to the third vector feature as a recognition result, including:
Sequentially extracting test images in the test data set, carrying out feature extraction on the images after image normalization preprocessing, and adding fourth vector features of the test images into a feature storage library;
preprocessing a single image to be queried, inputting the single image into a retrieval model, and extracting features to obtain second vector features for retrieving the single image;
and carrying out feature comparison and search on the second vector features in the feature storage library, calculating and inquiring K-class distances between the second vector features and third vector features in the feature storage library by using the K-nearest neighbor algorithm idea, and determining the category labels of the single images based on the K-class distances.
Preferably, storing the first vector feature of the training image in a feature store during training includes:
acquiring first identification information of a training image, wherein the first identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
grouping the feature storage libraries based on grouping information corresponding to the first identification information to obtain a plurality of feature storage groups;
classifying and storing the first vector features into corresponding feature storage groups based on first identification information of the training images;
Retrieving a corresponding third vector feature from a feature store based on the second vector feature, comprising:
acquiring second identification information of the single search image, wherein the second identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
calling a corresponding feature storage group based on the second identification information;
and searching in the feature storage group based on the second vector, and determining a corresponding third vector.
Preferably, the method for identifying the single weather image based on image retrieval further comprises the following steps:
acquiring feedback information for the identification output result;
when the feedback information is the identification result error, acquiring a correct result input by the user;
storing a single search image corresponding to the identification output result and the correct result in a correction database in a correlated way;
when the data quantity in the correction database reaches a preset condition, extracting a single search image in the correction database, adding the single search image into a training image, and retraining the retrieval model;
the preset conditions comprise:
correcting the data quantity in the database to reach a preset first quantity;
and/or the number of the groups of groups,
correcting similar data in the database to reach a preset second quantity;
And/or the number of the groups of groups,
the storage time interval of the similar data in the correction database is smaller than the preset time interval;
and/or the number of the groups of groups,
classifying the data based on the similarity between the data in the correction database, and when the number of the classification items reaches a preset third number;
when the feedback information of the identification output result is acquired, verifying the validity of the feedback information, and when the verification is passed, acquiring the feedback information, otherwise, not acquiring the feedback information; the conditions for passing the verification include:
the number of source users of the same feedback information is larger than the preset number of users;
and/or the number of the groups of groups,
when a single user is used as a source user to feed back feedback information, the confirmation degree of the feedback information is larger than a preset first confirmation value; the validation calculation formula is as follows:
wherein Q is i The confirmation degree of the feedback information when the feedback information is fed back for the ith user; a, a i Feedback confirmation values of i users are allocated in advance; b j The approval value of the user approving the feedback information is the j-th approval value; beta j The user is approved to choose an approval weight value when the feedback information is approved for the j-th approval user; m is the number of endorsement users;
and/or the number of the groups of groups,
when all source users of the same feedback information feed back the feedback information, the sum of the confirmation degrees of the feedback information is larger than a preset second confirmation value;
After retraining the retrieval model, correcting the feedback confirmation value of the source user of the feedback information, wherein the correction formula is as follows:
wherein a' is the feedback confirmation value of the source user after correction; a is the feedback confirmation value of the source user before correction; c is a correction coefficient and is related to the use condition of the feedback information, when the feedback information is verified to be in feedback error in use, c is taken to be-1, when the feedback information is verified to be in feedback correct in use, c is taken to be 1; epsilon is a preset correction amplitude value; epsilon 0 A corrected amplitude value which is a preset corrected amplitude value; a, a max The maximum feedback confirmation value is preset; a, a min Is a preset minimum feedback confirmation value.
The invention also provides a single weather image recognition system based on image retrieval, which comprises:
the model training module is used for acquiring training images, and training a preset initial model based on the training images to acquire a retrieval model;
the feature storage library establishing module is used for storing the first vector features of the training images to the feature storage library in the training process;
the retrieval module is used for inputting the single search image into the retrieval model, extracting the characteristics and obtaining second vector characteristics; retrieving a corresponding third vector feature from the feature store based on the second vector feature,
And the output module is used for outputting weather description of the training image corresponding to the third vector feature as a recognition result.
Preferably, the model training module performs the following operations:
preprocessing the training image;
extracting features of a training image by using a residual network structure ResNet50, performing array conversion on a training tag list, simultaneously calculating final loss by combining metric loss and cross entropy loss functions, and optimizing network parameters by using a batch random gradient descent method to minimize the loss functions;
in each training iteration, introducing a first vector feature formation queue extracted from a current batch-size image into a feature storage library, and eliminating corresponding old vector features in the feature storage library when the feature storage library is full;
and (3) in the iteration, matching the first vector features of the current batch-size image with all third vector features in a feature memory bank to form a large number of sample pairs, calculating the Similarity by using a metric Loss function Multi-Similarity Loss, and calculating the classification Loss of the current batch image by using a cross entropy Loss function.
Preferably, the retrieval module performs the following operations:
sequentially extracting test images in the test data set, carrying out feature extraction on the images after image normalization preprocessing, and adding fourth vector features of the test images into a feature storage library;
Preprocessing a single image to be queried, inputting the single image into a retrieval model, and extracting features to obtain second vector features for retrieving the single image;
and carrying out feature comparison and search on the second vector features in the feature storage library, calculating and inquiring K-class distances between the second vector features and third vector features in the feature storage library by using the K-nearest neighbor algorithm idea, and determining the category labels of the single images based on the K-class distances.
Preferably, storing the first vector feature of the training image in a feature store during training includes:
acquiring first identification information of a training image, wherein the first identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
grouping the feature storage libraries based on grouping information corresponding to the first identification information to obtain a plurality of feature storage groups;
classifying and storing the first vector features into corresponding feature storage groups based on first identification information of the training images;
retrieving a corresponding third vector feature from a feature store based on the second vector feature, comprising:
acquiring second identification information of the single search image, wherein the second identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
Calling a corresponding feature storage group based on the second identification information;
and searching in the feature storage group based on the second vector, and determining a corresponding third vector.
Preferably, the single weather image recognition system based on image retrieval further comprises: the updating module performs the following operations:
acquiring feedback information for the identification output result;
when the feedback information is the identification result error, acquiring a correct result input by the user;
storing a single search image corresponding to the identification output result and the correct result in a correction database in a correlated way;
when the data quantity in the correction database reaches a preset condition, extracting a single search image in the correction database, adding the single search image into a training image, and retraining the retrieval model;
the preset conditions comprise:
correcting the data quantity in the database to reach a preset first quantity;
and/or the number of the groups of groups,
correcting similar data in the database to reach a preset second quantity;
and/or the number of the groups of groups,
the storage time interval of the similar data in the correction database is smaller than the preset time interval;
and/or the number of the groups of groups,
classifying the data based on the similarity between the data in the correction database, and when the number of the classification items reaches a preset third number;
When the feedback information of the identification output result is acquired, verifying the validity of the feedback information, and when the verification is passed, acquiring the feedback information, otherwise, not acquiring the feedback information; the conditions for passing the verification include:
the number of source users of the same feedback information is larger than the preset number of users;
and/or the number of the groups of groups,
when a single user is used as a source user to feed back feedback information, the confirmation degree of the feedback information is larger than a preset first confirmation value; the validation calculation formula is as follows:
wherein Q is i The confirmation degree of the feedback information when the feedback information is fed back for the ith user; a, a i Feedback confirmation values of i users are allocated in advance; b j The approval value of the user approving the feedback information is the j-th approval value; beta j The user is approved to choose an approval weight value when the feedback information is approved for the j-th approval user; m is the number of endorsement users;
and/or the number of the groups of groups,
when all source users of the same feedback information feed back the feedback information, the sum of the confirmation degrees of the feedback information is larger than a preset second confirmation value;
after retraining the retrieval model, correcting the feedback confirmation value of the source user of the feedback information, wherein the correction formula is as follows:
wherein a' is the feedback confirmation value of the source user after correction; a is the feedback confirmation value of the source user before correction; c is a correction coefficient and is related to the use condition of the feedback information, when the feedback information is verified to be in feedback error in use, c is taken to be-1, when the feedback information is verified to be in feedback correct in use, c is taken to be 1; epsilon is a preset correction amplitude value; epsilon 0 A corrected amplitude value which is a preset corrected amplitude value; a, a max The maximum feedback confirmation value is preset; a, a min Is a preset minimum feedback confirmation value.
Compared with the prior art, the invention has the advantages that: firstly, the data images of the data set used by the invention are self-collection and manufacture, and focus on traffic scenes in natural scenes, so that the weather images can be identified, road natural disaster images and traffic accident images can be identified, and the method has good application prospects. Secondly, an image retrieval method is creatively adopted for the weather image recognition task, so that the accuracy of weather image recognition is effectively improved.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Drawings
The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate the invention and together with the embodiments of the invention, serve to explain the invention. In the drawings:
FIG. 1 is a schematic diagram of a single weather image recognition method based on image retrieval in an embodiment of the invention;
FIG. 2 is a schematic diagram of a search model;
fig. 3 is a schematic diagram of a single search image retrieval process.
Detailed Description
The preferred embodiments of the present invention will be described below with reference to the accompanying drawings, it being understood that the preferred embodiments described herein are for illustration and explanation of the present invention only, and are not intended to limit the present invention.
The embodiment of the invention provides a single weather image identification method based on image retrieval, which is shown in fig. 1 and comprises the following steps:
step S1, acquiring a training image, and training a preset initial model based on the training image to acquire a retrieval model; storing the first vector features of the training image to a feature store during training;
s2, inputting the single search image into a search model, and extracting features to obtain second vector features; and based on the second vector feature, retrieving a corresponding third vector feature from the feature storage library, and outputting weather description of the training image corresponding to the third vector feature as a recognition result.
The working principle and the beneficial effects of the technical scheme are as follows:
as shown in fig. 2, feature extraction is performed on the training image by using a CNN network to obtain vector features of the representative image; the CNN utilizes a ResNet network architecture, and realizes that the network goes to a deeper level under the condition of not greatly increasing the calculated amount through a residual block structure; introducing a feature storage mechanism Memory into the extracted image vector features, adding new features extracted from the current mini-batch [ small batch sample ] into a queue for each iteration, and removing old features; in the training process, the MS Loss measurement Loss function is utilized to effectively perform measurement learning, and a retrieval model is obtained through training;
As shown in fig. 3, a single search image is input into a search model to perform feature extraction; extracting features under test image line concentration, and storing the features by using Memory banks; feature vector of single search image is subjected to feature contrast search in Memory Bank, classification is realized by using K neighbor algorithm, category of query image is obtained, and weather image is identified.
Further, the weather training image is subjected to pixel normalization processing, randomly cut and uniformly scaled to an image block with the size of 227 multiplied by 227, randomly horizontally turned over, and then the processed weather image is input into a convolutional neural network;
the Multi-Similarity Loss is adopted as the metric Loss function, the Similarity of the characteristics of the current batch-size and all the characteristics in the Memory Bank of the characteristic Memory is effectively calculated, the metric Loss is calculated, the retrieval network is improved, the Similarity metric and the cross entropy Loss are combined to monitor and train the convolutional neural network, the Similarity metric in the test is consistent with the Similarity metric in the training, the test performance is more stable for images with larger difference in the training and the test, and the method is more suitable for retrieval scenes.
In one embodiment, step S1, obtaining a training image, and training a preset initial model based on the training image to obtain a retrieval model; storing the first vector features of the training image to a feature store during training; comprising the following steps:
Step S11: preprocessing the training image; preprocessing the training set image data before training, firstly establishing indexes for the images and the categories to which the images belong to form a label list; secondly, carrying out normalization processing on weather images, unifying the image sizes to 256×256, randomly selecting image blocks with the size of 227×227 from the 256×256 images during each training, randomly horizontally turning over, and inputting the horizontally turned horizontally into a network for training;
step S12: and carrying out feature extraction on the training image by using a residual network structure ResNet50, carrying out array conversion on a training label list, simultaneously calculating final loss by combining the metric loss and the cross entropy loss function, and carrying out network parameter optimization by a batch random gradient descent method to minimize the loss function. ResNet-50 is smaller in calculation amount and more suitable for real-time scene application. The ResNet-50 convolutional neural network is trained and fine-tuned by using weather images based on the metric loss and the cross entropy loss, so that the ResNet-50 convolutional neural network is suitable for weather retrieval scenes. And a feature embedding layer is added on the ResNet-50 structure, so that low-dimensional features are obtained directly based on metric loss training, and the feature does not need to be subjected to dimension reduction processing.
Step S13, in each training iteration, a first vector feature formation queue extracted from a current batch-size image is introduced into a feature storage library, and when the feature storage library is full, corresponding old vector features in the feature storage library are removed; thereby ensuring that all feature information in Memory Bank is kept up to date.
Step S14: and (3) in the iteration, matching the first vector features of the current batch-size image with all third vector features in a feature memory bank to form a large number of sample pairs, calculating the Similarity by using a metric Loss function Multi-Similarity Loss, and calculating the classification Loss of the current batch image by using a cross entropy Loss function. A more efficient sample training is achieved.
In one embodiment, step S2, inputting a single search image into a search model, and extracting features to obtain second vector features; retrieving a corresponding third vector feature from the feature repository based on the second vector feature, outputting a weather description of the training image corresponding to the third vector feature as a recognition result, including:
sequentially extracting test images in the test data set, carrying out feature extraction on the images after image normalization preprocessing, and adding fourth vector features of the test images into a feature storage library; and (3) sequentially extracting test images in the test data set, carrying out feature extraction on the images after the image normalization processing, and forming a feature Memory Bank of the step S1 to form dynamic storage of feature vectors.
Preprocessing a single image to be queried, inputting the single image into a retrieval model, and extracting features to obtain second vector features for retrieving the single image; preprocessing a single image to be queried, inputting the single image into a retrieval model obtained by training in the step S1, and extracting features of the retrieval image to obtain feature vectors of the retrieval image.
And carrying out feature comparison and search on the second vector features in the feature storage library, calculating and inquiring K-class distances between the second vector features and third vector features in the feature storage library by using the K-nearest neighbor algorithm idea, and determining the category labels of the single images based on the K-class distances. Feature vector of single weather image is subjected to feature contrast searching in a feature Memory Bank, K class distance of query feature vector is calculated by using K neighbor algorithm idea, classification output is realized after sequencing, category labels of query image are obtained, and weather image identification is realized.
Further, the loss function of the training search network adopts cross entropy loss and Multi-Similarity loss; in the Multi-Similarity loss, the weight parameter SCALE_POS is set to 2.0 and SCALE_NEG is set to 40.0;
further, the training time batch-size was set to 55, the test time batch-size was set to 128, and the number of work processes was set to 8.
In one embodiment, storing a first vector feature of a training image to a feature store during a training process includes:
acquiring first identification information of a training image, wherein the first identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
Grouping the feature storage libraries based on grouping information corresponding to the first identification information to obtain a plurality of feature storage groups;
classifying and storing the first vector features into corresponding feature storage groups based on first identification information of the training images;
retrieving a corresponding third vector feature from a feature store based on the second vector feature, comprising:
acquiring second identification information of the single search image, wherein the second identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
calling a corresponding feature storage group based on the second identification information;
and searching in the feature storage group based on the second vector, and determining a corresponding third vector.
The working principle and the beneficial effects of the technical scheme are as follows:
the weather conditions possibly occur in the same place and in the same season are much less than the overall weather conditions, so that the weather conditions can be identified by the representative place [ shooting position information ] and the representative season [ shooting time information ], and the feature storage library is grouped based on the first identification information and the second identification information, so that the data comparison quantity during single image detection is reduced, and the detection speed is improved; the number information of the shooting equipment is also used as identification information, and mainly, after the shooting equipment is installed, the shooting equipment generally does not move, so that the weather condition shot by the equipment is less than the total weather condition.
In one embodiment, the method for identifying the single weather image based on image retrieval further comprises the following steps:
acquiring feedback information for the identification output result;
when the feedback information is the identification result error, acquiring a correct result input by the user;
storing a single search image corresponding to the identification output result and the correct result in a correction database in a correlated way;
when the data quantity in the correction database reaches a preset condition, extracting a single search image in the correction database, adding the single search image into a training image, and retraining the retrieval model;
the preset conditions comprise:
correcting the data quantity in the database to reach a preset first quantity;
and/or the number of the groups of groups,
correcting similar data in the database to reach a preset second quantity;
and/or the number of the groups of groups,
the storage time interval of the similar data in the correction database is smaller than the preset time interval;
and/or the number of the groups of groups,
classifying the data based on the similarity between the data in the correction database, and when the number of the classification items reaches a preset third number;
when the feedback information of the identification output result is acquired, verifying the validity of the feedback information, and when the verification is passed, acquiring the feedback information, otherwise, not acquiring the feedback information; the conditions for passing the verification include:
The number of source users of the same feedback information is larger than the preset number of users;
and/or the number of the groups of groups,
when a single user is used as a source user to feed back feedback information, the confirmation degree of the feedback information is larger than a preset first confirmation value; the validation calculation formula is as follows:
wherein Q is i The confirmation degree of the feedback information when the feedback information is fed back for the ith user; a, a i Feedback confirmation values of i users are allocated in advance; b j The approval value of the user approving the feedback information is the j-th approval value; beta j The user is approved to choose an approval weight value when the feedback information is approved for the j-th approval user; m is the number of endorsement users;
and/or the number of the groups of groups,
when all source users of the same feedback information feed back the feedback information, the sum of the confirmation degrees of the feedback information is larger than a preset second confirmation value;
after retraining the retrieval model, correcting the feedback confirmation value of the source user of the feedback information, wherein the correction formula is as follows:
wherein a' is the feedback confirmation value of the source user after correction; a is the feedback confirmation value of the source user before correction; c is a correction coefficient and is related to the use condition of the feedback information, when the feedback information is verified to be in feedback error in use, c is taken to be-1, when the feedback information is verified to be in feedback correct in use, c is taken to be 1; epsilon is a preset correction amplitude value; epsilon 0 A corrected amplitude value which is a preset corrected amplitude value; a, a max The maximum feedback confirmation value is preset; a, a min Is preset toIs a minimum feedback acknowledgement value for (a).
The working principle and the beneficial effects of the technical scheme are as follows:
acquiring updated materials (single search images with identification errors) through collecting and analyzing feedback information of the identification output results, collecting the materials into a correction database, and when data in the correction database meet preset conditions, invoking the data to retrain a retrieval model so as to update the data model; the update time is determined by adopting preset conditions, so that the influence on the stability of the identification system caused by large fluctuation of the detection model due to frequent update is prevented; in the feedback information collection link, feedback of the feedback information needs to be verified, screening of the feedback information is achieved, and interference feedback information is eliminated. When a user feeds back, introducing the confirmation degree of the feedback information, marking the feedback information, wherein the confirmation degree is mainly determined by the user and the approver approving the feedback information, and the confirmation degree is enhanced for the marking effect of the feedback information by the determination method, so that the feedback information can be reflected to be effective; when the feedback information is verified in use, the verification path is judged to be abnormal data or abnormal feedback by a manager when the verification path is a training model; by correcting the feedback confirmation value of the source user, the accuracy of the next feedback information screening is improved.
The invention also provides a single weather image recognition system based on image retrieval, which comprises:
the model training module is used for acquiring training images, and training a preset initial model based on the training images to acquire a retrieval model;
the feature storage library establishing module is used for storing the first vector features of the training images to the feature storage library in the training process;
the retrieval module is used for inputting the single search image into the retrieval model, extracting the characteristics and obtaining second vector characteristics; retrieving a corresponding third vector feature from the feature store based on the second vector feature,
and the output module is used for outputting weather description of the training image corresponding to the third vector feature as a recognition result.
The working principle and the beneficial effects of the technical scheme are as follows:
as shown in fig. 2, feature extraction is performed on the training image by using a CNN network to obtain vector features of the representative image; the CNN utilizes a ResNet network architecture, and realizes that the network goes to a deeper level under the condition of not greatly increasing the calculated amount through a residual block structure; introducing a feature storage mechanism Memory Bank into the extracted image vector features, adding new features extracted by the current mini-batch into a queue for each iteration, and removing old features; in the training process, the MS Loss measurement Loss function and the cross entropy Loss function are combined to effectively train the model, and a retrieval model is obtained through training;
As shown in fig. 3, a single search image is input into a search model to perform feature extraction; extracting features under test image line concentration, and storing the features by using Memory banks; feature vector of single search image is subjected to feature contrast search in Memory Bank, classification is realized by using K neighbor algorithm, category of query image is obtained, and weather image is identified.
Further, the weather training image is subjected to pixel normalization processing, randomly cut and uniformly scaled to an image block with the size of 227 multiplied by 227, randomly horizontally turned over, and then the processed weather image is input into a convolutional neural network;
the metric Loss function adopts Multi-Similarity Loss to effectively calculate the characteristics of the current batch-size and the Similarity of all the characteristics in a characteristic Memory Bank, calculate the metric Loss and improve the retrieval network.
In one embodiment, the model training module performs the following:
preprocessing the training image;
extracting features of a training image by using a residual network structure ResNet50, performing array conversion on a training tag list, simultaneously calculating final loss by combining metric loss and cross entropy loss functions, and optimizing network parameters by using a batch random gradient descent method to minimize the loss functions;
In each training iteration, introducing a first vector feature formation queue extracted from a current batch-size image into a feature storage library, and eliminating corresponding old vector features in the feature storage library when the feature storage library is full;
and (3) in the iteration, matching the first vector features of the current batch-size image with all third vector features in a feature memory bank to form a large number of sample pairs, calculating the Similarity by using a metric Loss function Multi-Similarity Loss, and calculating the classification Loss of the current batch image by using a cross entropy Loss function.
In one embodiment, the retrieval module performs the following:
sequentially extracting test images in the test data set, carrying out feature extraction on the images after image normalization preprocessing, and adding fourth vector features of the test images into a feature storage library;
preprocessing a single image to be queried, inputting the single image into a retrieval model, and extracting features to obtain second vector features for retrieving the single image;
and carrying out feature comparison and search on the second vector features in the feature storage library, calculating and inquiring K-class distances between the second vector features and third vector features in the feature storage library by using the K-nearest neighbor algorithm idea, and determining the category labels of the single images based on the K-class distances.
In one embodiment, storing a first vector feature of a training image to a feature store during a training process includes:
acquiring first identification information of a training image, wherein the first identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
grouping the feature storage libraries based on grouping information corresponding to the first identification information to obtain a plurality of feature storage groups;
classifying and storing the first vector features into corresponding feature storage groups based on first identification information of the training images;
retrieving a corresponding third vector feature from a feature store based on the second vector feature, comprising:
acquiring second identification information of the single search image, wherein the second identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
calling a corresponding feature storage group based on the second identification information;
and searching in the feature storage group based on the second vector, and determining a corresponding third vector.
The working principle and the beneficial effects of the technical scheme are as follows:
the weather conditions possibly occur in the same place and in the same season are much less than the overall weather conditions, so that the weather conditions can be identified by the representative place [ shooting position information ] and the representative season [ shooting time information ], and the feature storage library is grouped based on the first identification information and the second identification information, so that the data comparison quantity during single image detection is reduced, and the detection speed is improved; the number information of the shooting equipment is also used as identification information, and mainly, after the shooting equipment is installed, the shooting equipment generally does not move, so that the weather condition shot by the equipment is less than the total weather condition.
In one embodiment, the single weather image identification system based on image retrieval further comprises: the updating module performs the following operations:
acquiring feedback information for the identification output result;
when the feedback information is the identification result error, acquiring a correct result input by the user;
storing a single search image corresponding to the identification output result and the correct result in a correction database in a correlated way;
when the data quantity in the correction database reaches a preset condition, extracting a single search image in the correction database, adding the single search image into a training image, and retraining the retrieval model;
the preset conditions comprise:
correcting the data quantity in the database to reach a preset first quantity;
and/or the number of the groups of groups,
correcting similar data in the database to reach a preset second quantity;
and/or the number of the groups of groups,
the storage time interval of the similar data in the correction database is smaller than the preset time interval;
and/or the number of the groups of groups,
classifying the data based on the similarity between the data in the correction database, and when the number of the classification items reaches a preset third number;
when the feedback information of the identification output result is acquired, verifying the validity of the feedback information, and when the verification is passed, acquiring the feedback information, otherwise, not acquiring the feedback information; the conditions for passing the verification include:
The number of source users of the same feedback information is larger than the preset number of users;
and/or the number of the groups of groups,
when a single user is used as a source user to feed back feedback information, the confirmation degree of the feedback information is larger than a preset first confirmation value; the validation calculation formula is as follows:
wherein Q is i The confirmation degree of the feedback information when the feedback information is fed back for the ith user; a, a i Feedback confirmation values of i users are allocated in advance; b j The approval value of the user approving the feedback information is the j-th approval value; beta j The user is approved to choose an approval weight value when the feedback information is approved for the j-th approval user; m is the number of endorsement users;
and/or the number of the groups of groups,
when all source users of the same feedback information feed back the feedback information, the sum of the confirmation degrees of the feedback information is larger than a preset second confirmation value;
after retraining the retrieval model, correcting the feedback confirmation value of the source user of the feedback information, wherein the correction formula is as follows:
wherein a' is the feedback confirmation value of the source user after correction; a is the feedback confirmation value of the source user before correction; c is a correction coefficient and is related to the use condition of the feedback information, when the feedback information is verified to be in feedback error in use, c is taken to be-1, when the feedback information is verified to be in feedback correct in use, c is taken to be 1; epsilon is a preset correction amplitude value; epsilon 0 A corrected amplitude value which is a preset corrected amplitude value; a, a max The maximum feedback confirmation value is preset; a, a min Is a preset minimum feedback confirmation value.
The working principle and the beneficial effects of the technical scheme are as follows:
acquiring updated materials (single search images with identification errors) through collecting and analyzing feedback information of the identification output results, collecting the materials into a correction database, and when data in the correction database meet preset conditions, invoking the data to retrain a retrieval model so as to update the data model; the update time is determined by adopting preset conditions, so that the influence on the stability of the identification system caused by large fluctuation of the detection model due to frequent update is prevented; in the feedback information collection link, feedback of the feedback information needs to be verified, screening of the feedback information is achieved, and interference feedback information is eliminated. When a user feeds back, introducing the confirmation degree of the feedback information, marking the feedback information, wherein the confirmation degree is mainly determined by the user and the approver approving the feedback information, and the confirmation degree is enhanced for the marking effect of the feedback information by the determination method, so that the feedback information can be reflected to be effective; when the feedback information is verified in use, the verification path is judged to be abnormal data or abnormal feedback by a manager when the verification path is a training model; by correcting the feedback confirmation value of the source user, the accuracy of the next feedback information screening is improved.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.
Claims (8)
1. The single weather image recognition method based on image retrieval is characterized by comprising the following steps of:
step S1, acquiring a training image, and training a preset initial model based on the training image to acquire a retrieval model; storing first vector features of the training image to a feature store during training;
s2, inputting a single search image into the search model, and extracting features to obtain second vector features; retrieving a corresponding third vector feature from the feature storage library based on the second vector feature, and outputting weather description of the training image corresponding to the third vector feature as a recognition result;
the single weather image identification method based on image retrieval further comprises the following steps:
acquiring feedback information of the identification result;
when the feedback information is an identification result error, acquiring a correct result input by a user;
The single search image corresponding to the identification result and the correct result are associated and stored in a correction database;
when the data volume in the correction database reaches a preset condition, extracting the single search image in the correction database, adding the single search image into the training image, and retraining the retrieval model;
wherein, the preset conditions include:
the data volume in the correction database reaches a preset first quantity;
and/or the number of the groups of groups,
the similar data in the correction database reaches a preset second quantity;
and/or the number of the groups of groups,
the storage time interval of the similar data in the correction database is smaller than a preset time interval;
and/or the number of the groups of groups,
classifying the data based on the similarity between the data in the correction database, and when the number of classification items reaches a preset third number;
when the feedback information of the identification result is acquired, verifying the validity of the feedback information, and when the verification is passed, acquiring the feedback information, otherwise, not acquiring the feedback information; the conditions for passing the verification include:
the number of the source users of the same feedback information is larger than the preset number of users;
and/or the number of the groups of groups,
When a single user is used as a source user to feed back the feedback information, the confirmation degree of the feedback information is larger than a preset first confirmation value; the validation calculation formula is as follows:
;
wherein,is->The confirmation degree of the feedback information when the individual users feed back the feedback information; />Is->The feedback confirmation values of the individual users are pre-allocated; />Is->An approval value of approving the user of the feedback information; />Is->The approval weight values selected by the approval users when approving the feedback information are selected; />A number of said approving users;
and/or the number of the groups of groups,
when all source users of the same feedback information feed back the feedback information, the sum of the confirmation degrees of the feedback information is larger than a preset second confirmation value;
after retraining the retrieval model, correcting the feedback confirmation value of the source user of the feedback information, wherein the correction formula is as follows:
；
wherein,confirming the value for the feedback of the source user after correction; />Feedback confirmation value for source user before correction; />In order to correct the coefficients, in relation to the use of the feedback information, the feedback information is verified as feedback errors when used, Taking-1, when the feedback information is verified as feedback correct at the time of use, ++>Taking 1; />The correction amplitude value is preset; />A corrected amplitude value which is a preset corrected amplitude value; />The maximum feedback confirmation value is preset; />Is a preset minimum feedback confirmation value.
2. The method for identifying a single weather image based on image retrieval according to claim 1, wherein the step S1 is to acquire a training image, train a preset initial model based on the training image to acquire a retrieval model; storing first vector features of the training image to a feature store during training; comprising the following steps:
step S11: preprocessing the training image;
step S12: extracting features of the training image by using a residual network structure ResNet50, performing array conversion on a training label list, simultaneously calculating final loss by combining metric loss and cross entropy loss functions, and optimizing network parameters by using a batch random gradient descent method to minimize the loss functions;
step S13, in each training iteration, a first vector feature formation queue extracted from a current batch-size image is introduced into a feature storage library, and when the feature storage library is full, the corresponding old vector features in the feature storage library are removed;
Step S14: and in the iteration process, the first vector features of the current batch-size image and all third vector features in the feature storage library are paired to form a large number of sample pairs, similarity is calculated by using a metric Loss function Multi-Similarity Loss, and classification Loss of the current batch image is calculated by using a cross entropy Loss function.
3. The method for identifying single weather image based on image retrieval according to claim 1, wherein said step S2 is to input a single search image into said retrieval model for feature extraction to obtain a second vector feature; retrieving a corresponding third vector feature from the feature repository based on the second vector feature, outputting a weather description of the training image corresponding to the third vector feature as a recognition result, including:
sequentially extracting test images in a test data set, carrying out feature extraction on the images after image normalization preprocessing, and adding fourth vector features of the test images into the feature storage library;
preprocessing a single image to be queried, inputting the single image into the retrieval model, and extracting features to obtain second vector features for retrieving the single image;
And carrying out feature comparison and search on the second vector features in the feature storage library, calculating and inquiring K-type distances between the second vector features and third vector features in the feature storage library by using a K-nearest neighbor algorithm idea, and determining the category label of the single image based on the K-type distances.
4. The method for identifying a single weather image based on image retrieval according to claim 1, wherein said storing a first vector feature of said training image in a feature store during training comprises:
acquiring first identification information of the training image, wherein the first identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
grouping the feature storage libraries based on grouping information corresponding to the first identification information to obtain a plurality of feature storage groups;
storing the first vector feature into the corresponding feature storage group based on the first identification information classification of the training image;
the retrieving, based on the second vector feature, a corresponding third vector feature from the feature store includes:
acquiring second identification information of the single search image, wherein the second identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
Invoking the corresponding feature storage group based on the second identification information;
and searching in the feature storage group based on the second vector, and determining the corresponding third vector.
5. A single weather image recognition system based on image retrieval, comprising:
the model training module is used for acquiring training images, and training a preset initial model based on the training images to acquire a retrieval model;
the feature storage library establishing module is used for storing the first vector features of the training image to a feature storage library in the training process;
the retrieval module is used for inputting a single search image into the retrieval model, extracting the characteristics and obtaining second vector characteristics; retrieving a corresponding third vector feature from the feature store based on the second vector feature,
the output module is used for outputting weather description of the training image corresponding to the third vector feature as a recognition result;
the updating module performs the following operations:
acquiring feedback information of the identification result;
when the feedback information is an identification result error, acquiring a correct result input by a user;
The single search image corresponding to the identification result and the correct result are associated and stored in a correction database;
when the data volume in the correction database reaches a preset condition, extracting the single search image in the correction database, adding the single search image into the training image, and retraining the retrieval model;
wherein, the preset conditions include:
the data volume in the correction database reaches a preset first quantity;
and/or the number of the groups of groups,
the similar data in the correction database reaches a preset second quantity;
and/or the number of the groups of groups,
the storage time interval of the similar data in the correction database is smaller than a preset time interval;
and/or the number of the groups of groups,
classifying the data based on the similarity between the data in the correction database, and when the number of classification items reaches a preset third number;
when the feedback information of the identification result is acquired, verifying the validity of the feedback information, and when the verification is passed, acquiring the feedback information, otherwise, not acquiring the feedback information; the conditions for passing the verification include:
the number of the source users of the same feedback information is larger than the preset number of users;
and/or the number of the groups of groups,
When a single user is used as a source user to feed back the feedback information, the confirmation degree of the feedback information is larger than a preset first confirmation value; the validation calculation formula is as follows:
;
wherein,is->The confirmation degree of the feedback information when the individual users feed back the feedback information; />Is->The feedback confirmation values of the individual users are pre-allocated; />Is->An approval value of approving the user of the feedback information; />Is->The approval weight values selected by the approval users when approving the feedback information are selected; />A number of said approving users;
and/or the number of the groups of groups,
when all source users of the same feedback information feed back the feedback information, the sum of the confirmation degrees of the feedback information is larger than a preset second confirmation value;
after retraining the retrieval model, correcting the feedback confirmation value of the source user of the feedback information, wherein the correction formula is as follows:
；
wherein,confirming the value for the feedback of the source user after correction; />Feedback confirmation value for source user before correction; />In order to correct the coefficients, in relation to the use of the feedback information, the feedback information is verified as feedback errors when used, Taking-1, when the feedback information is verified as feedback correct at the time of use, ++>Taking 1; />The correction amplitude value is preset; />A corrected amplitude value which is a preset corrected amplitude value; />The maximum feedback confirmation value is preset; />Is a preset minimum feedback confirmation value.
6. The image retrieval based single weather image recognition system as claimed in claim 5, wherein said model training module performs the following operations:
preprocessing the training image;
extracting features of the training image by using a residual network structure ResNet50, performing array conversion on a training label list, simultaneously calculating final loss by combining metric loss and cross entropy loss functions, and optimizing network parameters by using a batch random gradient descent method to minimize the loss functions;
in each training iteration, introducing a first vector feature formation queue extracted from a current batch-size image into a feature storage library, and eliminating corresponding old vector features in the feature storage library when the feature storage library is full;
and in the iteration process, the first vector features of the current batch-size image and all third vector features in the feature storage library are paired to form a large number of sample pairs, similarity is calculated by using a metric Loss function Multi-Similarity Loss, and classification Loss of the current batch image is calculated by using a cross entropy Loss function.
7. The single weather image identification system based on image retrieval as claimed in claim 5, wherein said retrieval module performs the following operations:
sequentially extracting test images in a test data set, carrying out feature extraction on the images after image normalization preprocessing, and adding fourth vector features of the test images into the feature storage library;
preprocessing a single image to be queried, inputting the single image into the retrieval model, and extracting features to obtain second vector features for retrieving the single image;
and carrying out feature comparison and search on the second vector features in the feature storage library, calculating and inquiring K-type distances between the second vector features and third vector features in the feature storage library by using a K-nearest neighbor algorithm idea, and determining the category label of the single image based on the K-type distances.
8. The image retrieval based single weather image identification system as claimed in claim 5, wherein said storing the first vector feature of the training image in a feature store during the training process comprises:
acquiring first identification information of the training image, wherein the first identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
Grouping the feature storage libraries based on grouping information corresponding to the first identification information to obtain a plurality of feature storage groups;
storing the first vector feature into the corresponding feature storage group based on the first identification information classification of the training image;
the retrieving, based on the second vector feature, a corresponding third vector feature from the feature store includes:
acquiring second identification information of the single search image, wherein the second identification information comprises: one or more of shooting position information, shooting time information and shooting equipment number information are combined;
invoking the corresponding feature storage group based on the second identification information;
and searching in the feature storage group based on the second vector, and determining the corresponding third vector.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010967561.4A CN112100419B (en) | 2020-09-15 | 2020-09-15 | Single weather image recognition method and system based on image retrieval |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010967561.4A CN112100419B (en) | 2020-09-15 | 2020-09-15 | Single weather image recognition method and system based on image retrieval |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112100419A CN112100419A (en) | 2020-12-18 |
| CN112100419B true CN112100419B (en) | 2024-02-06 |
Family
ID=73759769
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010967561.4A Active CN112100419B (en) | 2020-09-15 | 2020-09-15 | Single weather image recognition method and system based on image retrieval |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112100419B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20240024585A (en) * | 2022-08-17 | 2024-02-26 | 쿠팡 주식회사 | Data providing method and apparatus for the same |
| CN115331062B (en) * | 2022-08-29 | 2023-08-08 | 北京达佳互联信息技术有限公司 | Image recognition method, image recognition device, electronic device and computer-readable storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104598916A (en) * | 2014-09-11 | 2015-05-06 | 单勇 | Establishment method of train recognition system and train recognition method |
| CN106776675A (en) * | 2016-11-01 | 2017-05-31 | 上海斐讯数据通信技术有限公司 | A kind of plants identification method and system based on mobile terminal |
| CN107169148A (en) * | 2017-06-21 | 2017-09-15 | 北京百度网讯科技有限公司 | Image search method, device, equipment and storage medium |
| CN109118779A (en) * | 2018-10-12 | 2019-01-01 | 东软集团股份有限公司 | Break in traffic rules and regulations information identifying method, equipment and readable storage medium storing program for executing |
| CN109740627A (en) * | 2018-11-27 | 2019-05-10 | 南京邮电大学 | A kind of insect image identification identifying system and its method based on parallel-convolution neural network |
| CN110046738A (en) * | 2019-01-28 | 2019-07-23 | 南京林业大学 | A kind of forest fire prediction technique based on artificial intelligence perceptron model |
| CN110197258A (en) * | 2019-05-29 | 2019-09-03 | 北京市商汤科技开发有限公司 | Neural network searching method, image processing method and device, equipment and medium |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11195057B2 (en) * | 2014-03-18 | 2021-12-07 | Z Advanced Computing, Inc. | System and method for extremely efficient image and pattern recognition and artificial intelligence platform |
| US10438302B2 (en) * | 2017-08-28 | 2019-10-08 | The Climate Corporation | Crop disease recognition and yield estimation |
| KR102331675B1 (en) * | 2019-09-09 | 2021-11-30 | 엘지전자 주식회사 | Artificial intelligence apparatus and method for recognizing speech of user |
-
2020
- 2020-09-15 CN CN202010967561.4A patent/CN112100419B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104598916A (en) * | 2014-09-11 | 2015-05-06 | 单勇 | Establishment method of train recognition system and train recognition method |
| CN106776675A (en) * | 2016-11-01 | 2017-05-31 | 上海斐讯数据通信技术有限公司 | A kind of plants identification method and system based on mobile terminal |
| CN107169148A (en) * | 2017-06-21 | 2017-09-15 | 北京百度网讯科技有限公司 | Image search method, device, equipment and storage medium |
| CN109118779A (en) * | 2018-10-12 | 2019-01-01 | 东软集团股份有限公司 | Break in traffic rules and regulations information identifying method, equipment and readable storage medium storing program for executing |
| CN109740627A (en) * | 2018-11-27 | 2019-05-10 | 南京邮电大学 | A kind of insect image identification identifying system and its method based on parallel-convolution neural network |
| CN110046738A (en) * | 2019-01-28 | 2019-07-23 | 南京林业大学 | A kind of forest fire prediction technique based on artificial intelligence perceptron model |
| CN110197258A (en) * | 2019-05-29 | 2019-09-03 | 北京市商汤科技开发有限公司 | Neural network searching method, image processing method and device, equipment and medium |
Non-Patent Citations (4)
| Title |
|---|
| Cross-Batch Memory for Embedding Learning;Xun Wang 等;《Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)》;6388-6397 * |
| Maria Tzelepi 等.Relevance Feedback in Deep Convolutional Neural Networks for Content Based Image Retrieval.《SETN '16: Proceedings of the 9th Hellenic Conference on Artificial Intelligence》.2016,1–7. * |
| 基于交通视频的目标车辆检索方法及应用研究;刘凯雄;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》(第05期);C034-225 * |
| 深度学习在岩石薄片图像检索中的应用研究;岳清清;《中国优秀硕士学位论文全文数据库 信息科技辑》(第08期);I138-1155 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112100419A (en) | 2020-12-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107515895B (en) | Visual target retrieval method and system based on target detection | |
| CN112100419B (en) | Single weather image recognition method and system based on image retrieval | |
| CN112417381B (en) | Method and device for rapidly positioning infringement image applied to image copyright protection | |
| US20050100209A1 (en) | Self-optimizing classifier | |
| CN112149722B (en) | Image automatic labeling method based on unsupervised domain adaptation | |
| CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
| CN108197538A (en) | A kind of bayonet vehicle searching system and method based on local feature and deep learning | |
| CN109993201A (en) | A kind of image processing method, device and readable storage medium storing program for executing | |
| CN104951562B (en) | A kind of image search method based on VLAD dual adaptions | |
| CN107291895B (en) | Quick hierarchical document query method | |
| CN112950445B (en) | Compensation-based detection feature selection method in image steganalysis | |
| CN114723994B (en) | Hyperspectral image classification method based on dual classifier antagonism enhancement network | |
| CN111104398A (en) | Detection method and elimination method for approximate repeated record of intelligent ship | |
| CN115292532A (en) | Remote sensing image domain adaptive retrieval method based on pseudo label consistency learning | |
| CN117557872B (en) | Unsupervised anomaly detection method and device for optimizing storage mode | |
| CN114579794A (en) | Multi-scale fusion landmark image retrieval method and system based on feature consistency suggestion | |
| CN112800232B (en) | Case automatic classification method based on big data | |
| CN114596546A (en) | Vehicle weight recognition method and device, computer and readable storage medium | |
| CN113420291A (en) | Intrusion detection feature selection method based on weight integration | |
| CN112084353A (en) | Bag-of-words model method for rapid landmark-convolution feature matching | |
| CN116721343A (en) | Cross-domain field cotton boll recognition method based on deep convolutional neural network | |
| CN114168782B (en) | Deep hash image retrieval method based on triplet network | |
| CN114332452B (en) | Automatic detection method for image annotation result of target detection or target segmentation | |
| CN116108217A (en) | Fee evasion vehicle similar picture retrieval method based on depth hash coding and multitask prediction | |
| CN110750672B (en) | Image retrieval method based on deep measurement learning and structure distribution learning loss |
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 |