An expandable multi-category fine-grained rock classification device and classification method thereof

By designing an expandable rock classification device, using the perceptual hashing algorithm and transformer model combined with a fully connected network and winner tree, the problem of insufficient application of traditional models in the field of rock classification is solved, rapid training and accurate classification are achieved, and dynamic adaptation is achieved. Category changes improve training speed and accuracy.

CN114118204BActive Publication Date: 2024-12-03CHINA UNIV OF GEOSCIENCES (WUHAN) +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202111200806.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-13
Publication Date
2024-12-03
Estimated Expiration
2041-10-13

AI Technical Summary

Technical Problem

The existing transformer model is less used in the field of rock classification, and traditional machine learning algorithms do not perform well under large data volumes.

Method used

A scalable multi-category fine-grained rock classification device is designed, including preprocessors, distributors, decision core groups, evaluators and rules engines. It uses perceptual hashing algorithms and transformer models for feature extraction and classification, and uses all Connect the network and the winner tree for data competition and parameter adjustment, and dynamically increase the decision core to adapt to changes in the number of classification categories.

Benefits of technology

It realizes rapid training and accurate classification under smaller sample data, improves training speed and accuracy, and improves efficiency through distributed deployment and lightweight server operation, allowing dynamically increase categories and refines classification granularity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114118204B_ABST
    Figure CN114118204B_ABST
Patent Text Reader

Abstract

The present invention provides an expandable multi-category fine-grained rock classification device and a classification method thereof, characterized in that the rock classification device comprises: a preprocessor, a distributor, a decision kernel group, an evaluator and a rule engine. The rock classification device and the classification method of the present invention first input the image feature data of the decision kernel, which will be compared with the data in the subsample library of the decision kernel. The transformer model will locate the area where the input data is similar to the data in the subsample library, and then intercept the regional data and scale it to the same size, and then use the improved perceptual hash algorithm to calculate the similarity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of target image classification, and in particular to an expandable multi-category fine-grained rock classification device and a classification method thereof. Background Art

[0002] In the field of target image classification, traditional machine learning algorithms include SVM (support vector machine) and KNN (nearest neighbor algorithm). These traditional machine learning algorithms are more suitable for situations with relatively small amounts of data. Once they encounter massive amounts of data, they often seem to be unable to cope with the situation. In recent years, deep learning has performed better than traditional machine learning methods in the field of target image classification, so deep learning has been deeply studied and developed in the fields of application and target classification. At present, the deep learning networks widely used in image classification are mainly convolutional network models, such as the VGG-19 model built and trained by Karen Simonyan and Andrew Zisserman of Oxford University in 2014, the Inceptionv3 model built and trained by Google, and the ResNet50 model built and trained by Microsoft in 2015. These convolutional network models have good performance in feature extraction before model classification, so they have promoted the development of transfer learning, that is, to obtain the feature data of the image through these excellent models, and then add a small number of network layers for training to obtain a model that meets the requirements. By reducing the parameters to be trained, the training cost is greatly shortened.

[0003] With the development of convolutional networks, many different models have been born, such as MobileNet, Xception and other lightweight networks, which can be applied to mobile devices. The emerging transformer is a framework based on graph neural networks, which is still rarely used in the field of application and image classification, especially in the field of rock classification. Summary of the invention

[0004] The problem solved by the present invention is that the existing transformer model is rarely used in the field of rock classification.

[0005] In order to solve the above problems, on the one hand, the present invention provides an expandable multi-category fine-grained rock classification device, wherein the rock classification device comprises:

[0006] Preprocessors, distributors, decision kernels, evaluators, and rule engines;

[0007] The preprocessor is connected to the trainer inlet and the identifier inlet, and is used to unify the size and format of the input original rock image to obtain the data to be processed, and output it to the trainer inlet or the identifier inlet;

[0008] The distributor is connected to the trainer entrance and the decision core group, and is used to receive the data to be processed output by the trainer entrance, extract feature data and labels therefrom, and distribute the feature data and labels to the decision core group;

[0009] The decision core group is connected to the identifier entrance and the evaluator, and includes a plurality of decision cores and data channels. The decision core, on the one hand, performs a perceptual hash similarity evaluation on the feature data and labels distributed by the distributor and the sample library data in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition. A reward is obtained through the fully connected network output, the expected result, and the competition result, which is used to adjust some parameters of the fully connected network. On the other hand, the data to be processed output by the identifier entrance is input into the fully connected network with relu as the excitation function for binary classification, and the classification result is a value between 0 and 1. The closer to 1, the more the current feature data belongs to the current category, and the identification category is finally output;

[0010] The evaluator is connected to the decision core group and is used to evaluate the results after all decision cores compete and the competition data to determine whether it is necessary to continue to invest in the data channel;

[0011] The rule engine is located in the decision core and is used to add the participating data from each decision core to the winner tree to obtain the category with the highest similarity of current feature data.

[0012] Preferably, the number of the plurality of decision cores depends on the number of desired classifications, and decision cores are dynamically added later to increase the number of classification categories.

[0013] Preferably, the distributor comprises:

[0014] Distributor entry and MobileNet network;

[0015] The distributor inlet is used to extract rock images and sample labels from the input rock samples, and output the rock images to the MobileNet network, and output the sample labels to the decision kernel group;

[0016] The MobileNet network is used to receive the rock image, extract rock feature data therefrom, and output the rock feature data to the decision kernel group.

[0017] Preferably, the decision core includes:

[0018] Training core entry, subsample library, fully connected network and rule engine;

[0019] The training core entrance is used to receive rock feature data and sample labels. The decision core performs perceptual hash similarity evaluation on the rock feature data and sample labels with the data of the sub-sample library in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition to obtain a competition result. The fully connected network output, the expected result given by the sample label, and the competition result are fused through dynamically adjusted network parameters to obtain a reward value, which is fed back to the fully connected network for adjusting some parameters of the fully connected network.

[0020] Preferably, the rule engine comprises:

[0021] The winner tree is used to add the participating data from each decision core to the winner tree for layer-by-layer competition, output the competition results, and obtain the category with the highest similarity of the current feature data.

[0022] On the other hand, the present invention also provides a classification method, which uses the rock classification device as described above, and includes the following steps:

[0023] S1. Input the rock sample data to be trained;

[0024] S2. Unifying the size and format of the input original rock image to be trained by a preprocessor to obtain data to be processed;

[0025] S3. Input the data to be processed into the distributor through the trainer entrance, extract feature data and labels from it through the distributor, and distribute the feature data and labels to the decision core group;

[0026] S4. The decision core in the decision core group performs perceptual hash similarity evaluation on the feature data and labels distributed by the distributor and the sample library data in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition. A reward is obtained through the fully connected network output, expected results, and competition results, which is used to adjust some parameters of the fully connected network;

[0027] S5. Input the rock sample data to be classified;

[0028] S6. Unifying the size and format of the input original rock image to be classified by a preprocessor to obtain data to be processed;

[0029] S7. Input the data to be processed into the decision core group through the identifier entrance;

[0030] S8. The decision core in the decision core group inputs the unprocessed data output by the identifier input port into a fully connected network with relu as the activation function for binary classification. The classification result is a value between 0 and 1. The closer to 1, the more the current feature data belongs to the current category, and the final output is the identification category.

[0031] Preferably, the perceptual hash similarity evaluation in step S4 comprises the following steps:

[0032] S41. Reduce size: reduce the image to a uniform size;

[0033] S42. Color space reduction: (1) Randomly select an unvisited pixel point pi in the image and mark it as Ci; (2) Draw a three-dimensional sphere with radius r as the radius of the pixel space; (3) If the number of pixels in the sphere is greater than the set value L, mark all the pixels in the three-dimensional sphere as Ci, otherwise return to step (1); (4) Take all the marked pixels as the end point and pi as the starting point to obtain a series of vectors, calculate the sum vector of all vectors, move pi along the sum vector, and then repeat steps (2), (3), and (4) to calculate the average pixel value of the pixels with the same mark, which is used to replace the pixel value of the original pixel;

[0034] S43. Grayscale: Convert the image into a grayscale image;

[0035] S44. Calculate DCT: Calculate the DCT transform of the image to obtain a 32*32 DCT coefficient matrix;

[0036] S45. Reduce DCT: retain the 8*8 matrix in the upper left corner of the DCT coefficient matrix;

[0037] S46. Calculate the average value: calculate the mean value of DCT by using the mean hash method;

[0038] S47. Calculate the hash value: According to the 8*8 DCT matrix, set the 64-bit hash value to 0 or 1. The value greater than or equal to the DCT mean is set to "1", and the value less than the DCT mean is set to "0". Together, they form a 64-bit integer, which is the fingerprint of this image.

[0039] Compared with the prior art, the scalable multi-category fine-grained rock classification device and classification method described in the present invention have the following beneficial effects:

[0040] (1) The present invention uses an expandable classifier, and the initial training can be based on smaller sample data, and the data is used by each decision core to make decisions. When there are enough samples in the later stage, secondary training can be performed based on the previous one. The speed is better than that of ordinary methods in terms of training speed and accuracy. Each decision core can be run on a lightweight server, which can improve the training speed;

[0041] (2) The present invention adopts a similarity calculation method, the core of which is an optimized perceptual hash algorithm based on the perceptual hash algorithm + transformer model. The transformer model matches similar areas, inputs two images, extracts feature maps through the same CNN, connects the feature maps in series, and obtains a context feature map by combining position coding; this feature map and the query point after position coding are fed to transformerτ, and the network output is decoded by a multi-layer perceptron D to obtain the matching point x'. In this way, the location of similar areas in the two images is obtained, and then the similar areas are extracted, and the images are scaled to the same size, and then put into the perceptual hash calculator to obtain the similarity result, and finally the rocks are accurately classified. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the structure of the rock classification device of the present invention;

[0043] Figure 2 It is a schematic diagram of the structure of the distributor of the present invention;

[0044] Figure 3 It is a schematic diagram of the decision core structure of the present invention;

[0045] Figure 4 It is a schematic diagram of the rule engine structure of the present invention. DETAILED DESCRIPTION

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] Embodiment 1

[0048] Provide a scalable multi-category fine-grained rock classification device, such as Figure 1 As shown, the rock classification device comprises:

[0049] Preprocessors, distributors, decision kernels, evaluators, and rule engines;

[0050] The preprocessor is connected to the trainer inlet and the identifier inlet, and is used to unify the size and format of the input original rock image to obtain the data to be processed, and output it to the trainer inlet or the identifier inlet;

[0051] The distributor is connected to the trainer entrance and the decision core group, and is used to receive the data to be processed output by the trainer entrance, extract feature data and labels therefrom, and distribute the feature data and labels to the decision core group;

[0052] The decision core group is connected to the identifier entrance and the evaluator, and includes a plurality of decision cores and data channels. The decision core, on the one hand, performs a perceptual hash similarity evaluation on the feature data and labels distributed by the distributor and the sample library data in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition. A reward is obtained through the fully connected network output, the expected result, and the competition result, which is used to adjust some parameters of the fully connected network. On the other hand, the data to be processed output by the identifier entrance is input into the fully connected network with relu as the excitation function for binary classification, and the classification result is a value between 0 and 1. The closer to 1, the more the current feature data belongs to the current category, and the identification category is finally output;

[0053] The evaluator is connected to the decision core group and is used to evaluate the results after all decision cores compete and the competition data to determine whether it is necessary to continue to invest in the data channel;

[0054] The rule engine is located in the decision core and is used to add the participating data from each decision core to the winner tree (ie, a competition tree) to obtain the category with the highest similarity of the current feature data.

[0055] The number of the decision cores depends on the number of desired categories, and decision cores can be added dynamically later to increase the number of classification categories.

[0056] Among them, Figure 2 As shown, the distributor comprises:

[0057] Distributor entry and MobileNet network;

[0058] The distributor inlet is used to extract rock images and sample labels from the input rock samples, and output the rock images to the MobileNet network, and output the sample labels to the decision kernel group;

[0059] The MobileNet network is used to receive the rock image, extract rock feature data therefrom, and output the rock feature data to the decision kernel group.

[0060] Among them, Figure 3 As shown, the decision core includes:

[0061] Training core entry, subsample library, fully connected network and rule engine;

[0062] The training core entrance is used to receive rock feature data and sample labels. The decision core performs perceptual hash similarity evaluation on the rock feature data and sample labels with the data of the sub-sample library in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition to obtain a competition result. The fully connected network output, the expected result given by the sample label, and the competition result are fused through dynamically adjusted network parameters to obtain a reward value, which is fed back to the fully connected network for adjusting some parameters of the fully connected network.

[0063] Among them, Figure 4 As shown, the rule engine includes:

[0064] The winner tree (i.e., a competition tree) is used to add the participating data from each decision core to the winner tree for layer-by-layer competition, output the competition results, and obtain the category with the highest similarity of the current feature data.

[0065] This rock classification device has the following functions:

[0066] 1. Deployment method: This device can be deployed in a distributed manner, using parallel computing to improve training efficiency.

[0067] Each decision core is an independent classifier, which is trained like an independent program. The data flows in a fixed direction from the first decision core to the last decision core.

[0068] 2. New categories can be added dynamically: Categories can be added dynamically, and adjustments after each new category only require parameter updates within a small range.

[0069] The first initial training is to flow data into each decision core, and then let the fully connected network in the decision core train the input feature data and label data. Each subsequent progressive training is that the input feature data generates a bias value X in each decision core. a , the closer to 1, the more likely it is to belong to the category represented by this decision kernel; and generate a similarity, the similarity calculation between the input feature data and the first data previously saved in the decision kernel, and take the average of the highest 50 similarities as the similarity X between the current feature data and the sub-sample library data b The bias value and similarity value are mixed in a certain ratio a:b. The ratio is initialized once. The recommended ratio is 6:4, where the bias value accounts for 6 and the similarity value accounts for 4. Then the ratio changes automatically with training, and then a competitive value is obtained. It will be put into the rule engine to compete with all decision core outputs. The winner will store the feature data in its sub-sample library and classify the feature data into the category represented by this decision core.

[0070] 3. Transformation to fine-grained classification: With the subsequent addition of new samples and the increase in the number of categories, the granularity of the classifier classification can be refined to achieve classification and identification of small categories in large categories of rocks.

[0071] From the first initial training to several subsequent retrainings, the number of classifications of the classification device is gradually increased to achieve the goal of refining the classification granularity.

[0072] The rock classification device in this embodiment uses an expandable classifier. The early training can be based on smaller sample data, and the data is used by each decision core to make decisions. When there are enough samples in the later stage, secondary training can be performed on the basis of the former. The speed is better than that of ordinary methods in terms of training speed and accuracy. Each decision core can be run on a lightweight server, which can improve the training speed.

[0073] Compared with the current image classification network model, this rock classification device takes into account long-term automatic optimization and optimizes the current model through the newly added sample data in the later stage. Therefore, a distributed model mode is adopted. Each decision core is an independent classifier, just like an independent program for training. The first initial training will get a classification model based on fewer samples; each subsequent additional training will generate a bias value and similarity value, and use this to calculate a competitive value for adjusting the model parameters. With the addition of new classification categories and the continuous input of samples in the later stage, the model performance will get better and better.

[0074] Embodiment 2

[0075] A classification method is provided, which uses the rock classification device as described in Example 1, and includes the following steps:

[0076] S1. Input the rock sample data to be trained;

[0077] S2. Unifying the size and format of the input original rock image to be trained by a preprocessor to obtain data to be processed;

[0078] S3. Input the data to be processed into the distributor through the trainer entrance, extract feature data and labels from it through the distributor, and distribute the feature data and labels to the decision core group;

[0079] S4. The decision core in the decision core group performs perceptual hash similarity evaluation on the feature data and labels distributed by the distributor and the sample library data in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition. A reward is obtained through the fully connected network output, expected results, and competition results, which is used to adjust some parameters of the fully connected network;

[0080] S5. Input the rock sample data to be classified;

[0081] S6. Unifying the size and format of the input original rock image to be classified by a preprocessor to obtain data to be processed;

[0082] S7. Input the data to be processed into the decision core group through the identifier entrance;

[0083] S8. The decision core in the decision core group inputs the unprocessed data output by the identifier input port into a fully connected network with relu as the activation function for binary classification. The classification result is a value between 0 and 1. The closer to 1, the more the current feature data belongs to the current category, and the final output is the identification category.

[0084] The perceptual hash similarity evaluation in step S4 includes the following steps:

[0085] S41. Reduce size: reduce the image to a uniform size;

[0086] S42. Color space reduction: (1) Randomly select an unvisited pixel point pi in the image and mark it as Ci; (2) Draw a three-dimensional sphere with radius r as the radius of the pixel space (assuming it is a three-channel image, such as RGB or HSV image); (3) If the number of pixels in the sphere is greater than the set value L, mark all the pixels in the three-dimensional sphere as Ci, otherwise return to step (1); (4) Take all the marked pixels as the end point and pi as the starting point to obtain a series of vectors, calculate the sum vector of all vectors, move pi along the sum vector, and then repeat steps (2), (3), and (4) to calculate the average pixel value of the pixels with the same mark, which is used to replace the pixel value of the original pixel;

[0087] S43. Grayscale: Convert the image into a grayscale image to further simplify the calculation;

[0088] S44. Calculate DCT (DCT, discrete cosine transform): calculate the DCT transform of the image and obtain a 32*32 DCT coefficient matrix;

[0089] S45. Reduce DCT (Discrete Cosine Transform): Although the result of DCT is a 32*32 matrix, we only need to keep the 8*8 matrix in the upper left corner, which presents the lowest frequency in the image;

[0090] S46. Calculate the average value: Calculate the average value of DCT just like the mean hashing;

[0091] S47. Calculate hash value: This is the most important step. According to the 8*8 DCT matrix, set the 64-bit hash value to 0 or 1. The value greater than or equal to the DCT mean is set to "1", and the value less than the DCT mean is set to "0". Together, they form a 64-bit integer, which is the fingerprint of this image.

[0092] Regarding the similarity calculation of the classification method in this embodiment, the image feature data of the decision kernel is first input, which will be compared with the data in the sub-sample library of the decision kernel. The transformer model will locate the area where the input data is similar to the data in the sub-sample library, and then the regional data will be intercepted and scaled to the same size, and then the improved perceptual hashing algorithm will be used to calculate the similarity.

[0093] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A scalable multi-category fine-grained rock classification device, characterized in that: The rock classification device comprises: Preprocessors, distributors, decision kernels, evaluators, and rule engines; The preprocessor is connected to the trainer inlet and the identifier inlet, and is used to unify the size and format of the input original rock image to obtain the data to be processed, and output it to the trainer inlet or the identifier inlet; The distributor is connected to the trainer entrance and the decision core group, and is used to receive the data to be processed output by the trainer entrance, extract feature data and labels therefrom, and distribute the feature data and labels to the decision core group; The decision core group is connected to the identifier entrance and the evaluator, and includes a plurality of decision cores and data channels. The decision core, on the one hand, performs a perceptual hash similarity evaluation on the feature data and labels distributed by the distributor and the sample library data in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition. A reward is obtained through the fully connected network output, the expected result, and the competition result, which is used to adjust some parameters of the fully connected network. On the other hand, the data to be processed output by the identifier entrance is input into the fully connected network with relu as the excitation function for binary classification, and the classification result is a value between 0 and 1. The closer to 1, the more the current feature data belongs to the current category, and the identification category is finally output; The evaluator is connected to the decision core group and is used to evaluate the results after all decision cores compete and the competition data to determine whether it is necessary to continue to invest in the data channel; The rule engine is located in the decision core and is used to add the participating data from each decision core to the winner tree to obtain the category with the highest similarity of current feature data.

2. The rock classification device according to claim 1, characterized in that: The number of the decision cores depends on the number of expected classifications, and decision cores are dynamically added later to increase the number of classification categories.

3. The rock classification device according to claim 1, characterized in that: The distributor comprises: Distributor entry and MobileNet network; The distributor inlet is used to extract rock images and sample labels from the input rock samples, and output the rock images to the MobileNet network, and output the sample labels to the decision kernel group; The MobileNet network is used to receive the rock image, extract rock feature data therefrom, and output the rock feature data to the decision kernel group.

4. The rock classification device according to claim 1, characterized in that: The decision core comprises: Training core entry, subsample library, fully connected network and rule engine; The training core entrance is used to receive rock feature data and sample labels. The decision core performs perceptual hash similarity evaluation on the rock feature data and sample labels with the data of the sub-sample library in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition to obtain a competition result. The fully connected network output, the expected result given by the sample label, and the competition result are fused through dynamically adjusted network parameters to obtain a reward value, which is fed back to the fully connected network for adjusting some parameters of the fully connected network.

5. The rock classification device according to claim 1, characterized in that: The rule engine comprises: The winner tree is used to add the participating data from each decision core to the winner tree for layer-by-layer competition, output the competition results, and obtain the category with the highest similarity of the current feature data.

6. A classification method using the rock classification device according to any one of claims 1 to 5, characterized in that: The steps include: S1. Input the rock sample data to be trained; S2. Unifying the size and format of the input original rock image to be trained by a preprocessor to obtain data to be processed; S3. Input the data to be processed into the distributor through the trainer entrance, extract feature data and labels from it through the distributor, and distribute the feature data and labels to the decision core group; S4. The decision core in the decision core group performs perceptual hash similarity evaluation on the feature data and labels distributed by the distributor and the sample library data in the core to obtain a comprehensive evaluation value, and outputs the comprehensive evaluation value to the rule engine for data competition. A reward is obtained through the fully connected network output, expected results, and competition results, which is used to adjust some parameters of the fully connected network; S5. Input the rock sample data to be classified; S6. Unifying the size and format of the input original rock image to be classified by a preprocessor to obtain data to be processed; S7. Input the data to be processed into the decision core group through the identifier entrance; S8. The decision core in the decision core group inputs the unprocessed data output by the identifier input port into a fully connected network with relu as the activation function for binary classification. The classification result is a value between 0 and 1. The closer to 1, the more the current feature data belongs to the current category, and the final output is the identification category.

7. The classification method according to claim 6, characterized in that: The perceptual hash similarity evaluation in step S4 includes the following steps: S41. Reduce size: reduce the image to a uniform size; S42. Color space reduction: (1) Randomly select an unvisited pixel point pi in the image and mark it as Ci; (2) Draw a three-dimensional sphere with radius r as the radius of the pixel space; (3) If the number of pixels in the sphere is greater than the set value L, mark all the pixels in the three-dimensional sphere as Ci, otherwise return to step (1); (4) Take all the marked pixels as the end point and pi as the starting point to obtain a series of vectors, calculate the sum vector of all vectors, move pi along the sum vector, and then repeat steps (2), (3), and (4) to calculate the average pixel value of the pixels with the same mark, which is used to replace the pixel value of the original pixel; S43. Grayscale: Convert the image into a grayscale image; S44. Calculate DCT: Calculate the DCT transform of the image to obtain a 32*32 DCT coefficient matrix; S45. Reduce DCT: retain the 8*8 matrix in the upper left corner of the DCT coefficient matrix; S46. Calculate the average value: calculate the mean value of DCT by using the mean hash method; S47. Calculate the hash value: According to the 8*8 DCT matrix, set the 64-bit hash value to 0 or 1. The value greater than or equal to the DCT mean is set to "1", and the value less than the DCT mean is set to "0". Together, they form a 64-bit integer, which is the fingerprint of this image.

Citation Information

Patent Citations

  • Method and system for identifying current land utilization situation land type of target land parcel by using field photos

    CN112766417A

  • Rock slice microscopic image deep learning-based rock classification method

    CN113222071A

  • Geo-visual search

    US10248663B1