CN113011440A - Coal bed gas well field monitoring heavy identification technology - Google Patents
Coal bed gas well field monitoring heavy identification technology Download PDFInfo
- Publication number
- CN113011440A CN113011440A CN202110296521.6A CN202110296521A CN113011440A CN 113011440 A CN113011440 A CN 113011440A CN 202110296521 A CN202110296521 A CN 202110296521A CN 113011440 A CN113011440 A CN 113011440A
- Authority
- CN
- China
- Prior art keywords
- distance
- sample
- samples
- clustering
- attention
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000012544 monitoring process Methods 0.000 title claims abstract description 19
- 238000005516 engineering process Methods 0.000 title claims abstract description 11
- 239000003245 coal Substances 0.000 title claims abstract description 8
- 238000000034 method Methods 0.000 claims abstract description 30
- 238000010168 coupling process Methods 0.000 claims abstract description 19
- 230000008878 coupling Effects 0.000 claims abstract description 18
- 238000005859 coupling reaction Methods 0.000 claims abstract description 18
- 238000011176 pooling Methods 0.000 claims abstract description 11
- 238000013507 mapping Methods 0.000 claims abstract description 5
- 239000000523 sample Substances 0.000 claims description 52
- 239000011159 matrix material Substances 0.000 claims description 17
- 238000004364 calculation method Methods 0.000 claims description 11
- 239000013074 reference sample Substances 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 8
- 238000010586 diagram Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 6
- 238000003064 k means clustering Methods 0.000 claims description 4
- 230000007246 mechanism Effects 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 238000000354 decomposition reaction Methods 0.000 claims description 2
- 230000002093 peripheral effect Effects 0.000 claims description 2
- 230000000007 visual effect Effects 0.000 claims description 2
- 230000015556 catabolic process Effects 0.000 claims 1
- 238000006731 degradation reaction Methods 0.000 claims 1
- 238000000605 extraction Methods 0.000 abstract description 6
- 238000001514 detection method Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000011840 criminal investigation Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005553 drilling Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000001757 thermogravimetry curve Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Probability & Statistics with Applications (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a coal bed gas well field monitoring re-identification technology. Firstly, constructing a feature extraction network, replacing average pooling with weighted pooling, calculating the inner product between feature attention heat maps according to classification, and extracting the significant features of the target; secondly, performing corresponding feature coupling clustering on the features, randomly selecting high-confidence ID samples to form clustering centers, establishing heterogeneous sample relative distance thresholds, calculating the distance between the features and the clustering centers, and distinguishing target features from background interference and heterogeneous features to obtain feature coupling clustering losses; and finally, mapping reference, similar and heterogeneous samples to a feature space by using a three-configuration network, and comparing the similarity. The method can better cross-camera target re-identification, can well improve the feature identification capability, and improves the accuracy and robustness of re-identification.
Description
Technical Field
The invention relates to a coal bed gas well field monitoring re-identification method technology, and belongs to the technical field of computer vision and image processing.
Background
Along with the rapid development of computer software and hardware, high-performance computers and camera terminals are very common, and are commonly applied to urban intelligent monitoring. With the development and progress of society, the popularization rate of automobiles is higher and higher, and for urban safety, an image analysis technology under monitoring is urgently needed, and a target identification technology becomes the focus of attention of people. The pedestrian re-identification has general application in video monitoring, intelligent transportation and urban calculation, and can quickly find, position and track target pedestrians in large-scale monitoring videos. Unlike pedestrian detection, tracking or classification, pedestrian re-identification can be found as an instance-level object search problem. The pedestrian re-identification problem is a search problem of judging whether a pedestrian image shot in a non-overlapping area belongs to the same person or not in a traffic monitoring scene in a specific range. At present, monitoring cameras exist in areas with developed traffic (crossroads, areas with easily-caused accidents, road sections with large pedestrian flow and the like), and the cameras are used for conducting traffic monitoring, criminal investigation and the like better through researching, so that the method is very valuable. Currently, the related research is mainly divided into three methods: the first method mainly combines a sensor to re-identify pedestrians, and the method is greatly influenced by weather interference and distance factors, and has high corresponding hardware cost, so that the practical requirements cannot be met. The second method mainly utilizes a manual feature extraction mode, and the method mainly utilizes some extraction algorithms to extract some manual features, so that the accuracy cannot be improved. The third method based on deep learning is carried out as soon as possible, the accuracy of feature extraction is higher, and the recognition accuracy is improved. The three-configuration network is an improvement of a twin network, the judgment capability is remarkably improved, and under the condition of complex background interference, the three-configuration network has a good effect on the identification and detection field for the remarkable feature extraction.
Disclosure of Invention
The patent aims to overcome the defects of the prior art introduced above, and provides a coal bed gas well site monitoring re-identification method technology aiming at identifying the same target of a cross-camera under a complex environment by corresponding pedestrian feature discrimination capability of a three-configuration network.
The technical scheme adopted by the invention is as follows: and constructing a three-configuration network based on feature coupling clustering to perform target re-identification. For image data, the extracted semantic features from lower layers are extracted at present by manual feature extraction, the identification is carried out by focusing on feature information such as pedestrian contours, textures and colors, and higher-layer semantic information including spatial characteristics and time and position information can be extracted through a neural network. The high-level semantic information is very important for the feature expression of the target pedestrian, and is well complemented with the extracted bottom-level semantic information, so that the corresponding re-identification is better realized. The method can fully utilize the spatial characteristics to obtain the spatial semantic information of the corresponding target pedestrian, and improves the discrimination capability of the method on the same target of the cross-camera under the complex environment and the reliability and the robustness of re-identification by utilizing the advantage of complementation with the low-level semantic information.
A coal bed gas well site monitoring re-identification method technology comprises the following steps:
firstly, constructing an embedded network structure, adding a weighted attention pooling layer to replace an average pooling layer in the embedded network structure, calculating the inner product between pedestrian feature attention heat maps output in pairs by classification calculation, and emphatically extracting the significant features of target pedestrians;
secondly, by the characteristic coupling clustering method, different ID samples with higher confidence are randomly selected to serve as clustering centers, the characteristic points of the target pedestrians, the background interference points and different pedestrians can be correspondingly clustered and distinguished by establishing the relative distance threshold of the heterogeneous samples and calculating the space Mahalanobis distance between the characteristic points and the clustering centers to obtain corresponding characteristic coupling clustering losses, and then the similarity is calculated through a comparator.
And thirdly, mapping the reference sample, the same-class sample and the different-class sample images into a feature space by integrating three embedded networks into a three-configuration network framework.
The method has the advantages that the advantages of similar feature coupling clustering are fully utilized, meanwhile, a weighted attention pool is used for replacing an average pooling layer to extract corresponding significant feature information of a target pedestrian and construct a feature heat map, positive correlation features of a specific target are activated through a weighted attention pool module, and better weight deviation is given to corresponding channels by using the spatial features of the significant features, so that better pedestrian weight recognition is realized.
The invention can better identify the same pedestrian crossing the camera, can well improve the characteristic identification capability and improve the accuracy and the robustness of pedestrian re-identification.
Drawings
FIG. 1 is a schematic diagram of a three-configuration network pedestrian re-identification structure based on feature coupling clustering.
FIG. 2 is a graph comparing the accuracy of reordering and re-ordering on a MarKet1501 data set.
FIG. 3 is a graph comparing the exact curves of the present invention on a MarKet1501 data set with other classical algorithms.
FIG. 4 is a comparison of confidence accuracy of the present invention reordering and reordering on a MarKet1501 data set.
FIG. 5 is a comparison of confidence accuracy of the present invention and a classical algorithm on a MarKet1501 data set.
FIG. 6 is a graph of characteristic coupling clustering effect in the present invention.
FIG. 7 is a diagram of a weighted attention pool structure in the present invention.
FIG. 8 is a diagram 1 illustrating the effect of the invention in identifying pedestrians in coal-bed gas wells.
FIG. 9 is a diagram 2 illustrating the effect of the invention in identifying pedestrians in coal-bed gas wells.
Detailed Description
The following further describes a specific embodiment of the present invention with reference to the drawings and technical solutions.
A coal bed gas well site monitoring re-identification method technology comprises the following steps:
the method comprises the steps of firstly, constructing an embedded network, extracting sample characteristics, extracting characteristics of related images under the action of each subnet, extracting the salient characteristics of reference pedestrians, similar samples and heterogeneous samples by using a convolution network, and projecting the salient characteristics to a characteristic space.
For the first step of the embedded network structure, we replace the average pooling layer with a weighted attention pool, the structure is shown in FIG. 1. The invention is composed of two network structures of an embedded network and a three-configuration network. The original average pooling layer pooling is represented by a matrix using a weighted attention pool and by a matrix using a second order pool, then a decomposition is performed on the matrix of weights, and then a bottom-up and top-down attention interpretation is performed using the degraded results. The specific implementation mode is as follows:
wherein X ∈ Rn*f,W∈Rf*fX represents a feature matrix, n represents the number of samples, and f represents the number of channels
Then proceed to the lowest rank approximation WkWeight, Wk∈Rf*f:
Wherein a, b ∈ Rf*1Where a, b denote the rows and columns of the weight matrix and k denotes the number of categories.
We chose to simulate the visual saliency of class-specific "top-down" attention mechanism from bottom-up, independent of class. It is usually mandatory that one of the attention parameter vectors is class independent, so we turn bkSo our last low rank attention pool mechanism module is denoted as:
Gradeattention(X,a,b)＝(Xak)T(Xbk) (formula 3)
Finally, the method continues to be simplified as follows:
wherein T isk＝Xak,h＝Xb。
This is equivalent to the top-down (special class) TkAnd a bottom-up significance-based h-attention map. We have previously averaged or maximized the pool by reducing the feature matrix to Rf*1Then generates the classification number by FC (full connection layer) weight vector W processing. We can see the attention thermogram as Xa ∈ Rn*1B is considered as a classifier of attention pool features, so finally our GradeattentionCan be seen as the inner product between two attention heat maps defined over all n spatial locations.
In our experiments, we predicted a single-channel bottom-up saliency map that is the same as the last feature map, and used a linear classifier on top, and we also generated n1×n2The top-down attention map Xa for the x Num _ class dimension, where Num _ class is the number of classes. We multiply the two attention maps by the spatial average, generating an output prediction for the Num _ class dimension (Xa)k)T(Xbk) These operations are equivalent to first characterizing and characterizing featuresCharacteristic XT(Xb)) and then passed through a classifier (a (X)T(Xb))。
And secondly, mapping the reference sample, the same-class sample and the different-class sample images into a feature space by integrating three embedded networks into a three-configuration network framework. According to the characteristic coupling clustering method, different ID samples with high confidence coefficient are randomly selected to serve as clustering centers, the characteristic points of the target pedestrians, the background interference points and different pedestrians can be correspondingly clustered and distinguished by establishing the relative distance threshold of heterogeneous samples and calculating the space Mahalanobis distance between the characteristic points and the clustering centers to obtain corresponding characteristic coupling clustering losses, and then the similarity is calculated through a comparator.
In a spatial feature set, feature coupling clustering is used for dividing same classes and different classes, so that later-stage re-identification is facilitated, and the specific implementation process is as follows:
randomly selecting different ID samples with higher confidence as Cluster centerscenterHowever, each time of selecting the cluster center can only select one of all the IDs of the samples as the cluster center, for each cluster center, the space Euclidean distance between the peripheral samples and the cluster center is calculated, and a negative sample relative distance threshold value d is definednNegative sample dnDefining the shortest distance as the distance between the centers of different clusters, when we calculate the distance between the centers of clusters and other samples SiDistance between, if dist (cluster)center,Fi)>dnIn which F isiRepresenting the bottom i feature samples, centered at the cluster center, and dnDividing a cluster with the distance as the radius, clustering the candidate characteristic points, and selecting min (cluster)center,Fi))>dnThe farthest point of (a) is re-used as a cluster center, and the continued clustering division is performed in the same manner, so that all samples can be labeled with corresponding similar and dissimilar pseudo labels, wherein we use a feature coupled clustering loss to perform the corresponding distance calculation, unlike K-means clustering, our distance calculationData-coupled clustering automatically labels part of the data as similar high confidence, while K-means clustering divides all data into several clusters, and our cluster center labels similar samples as identical IDs. In the clustering process, a plurality of candidate feature points probably meet the constraint, and the point with the largest difference is selected as a clustering center sample point.
min(DistM(fi,fj))>dn
Wherein f isi,fjAll belong to the feature map in the image library to be retrieved, ClustercenterIs the center of the cluster, dnIs a threshold value.
Spatial feature points are calculated using mahalanobis distance:
the mahalanobis distance between feature points is calculated as follows:
sigma is a covariance matrix of a multi-dimensional random variable, if the covariance matrix is a unit vector, dimensions are independently and identically distributed, the mahalanobis distance is the euclidean distance, and since the euclidean distance has a large error in the distance in the flow pattern space under different dimensions, the computed euclidean distances are not necessarily similar, and later, people want to adopt a normalized euclidean distance to eliminate the problem of inconsistent dimension scales, but with the increase of sample amount, the difference of sample distribution is larger and larger, the normalized euclidean distance cannot well eliminate outlier, so we adopt the mahalanobis distance to calculate the spatial distance; the mahalanobis distance is specifically derived as follows:
the feature points are rotated to Principal Components (PCA) first, so that all dimensions become linearly independent and mutually independent, and new coordinates are generated at the moment.
The mahalanobis distance is the euclidean distance after the rotation change, and the mahalanobis distance equation 6 is calculated as:
when DistCluster<dnThen, taking the cluster center as the center, clustering the feature points, dnDistance, conversely, we push the distance away by dnThe distance of (c). For TripletNet (three-state network), since three sample pairs, i.e. sample x, are input, positive sample x+Negative sample x-. Selecting positive sample and negative sample x in positive sample set and negative sample set+,x-Thus, by feature coupled clustering as described above, we estimate the average of the cluster centers for each positive sample:
wherein N is+Is the number of positive samples
The resulting relative distance calculation for positive and negative samples can be reflected by the following equation:
where i belongs to the positive sample set and j belongs to the negative sample set, they satisfy the relationship:
Distpostive+θ≤Distnegtive(formula 10)
Where θ is a distance deviation, θ < dn
One coupled cluster loss can be found:
wherein xiIs a set of positive samples, xjIs a set of negative samples, N+Is the number of positive samples.
If it is not
If the partial derivative is less than or equal to 0, the partial derivative of the positive and negative samples is 0, otherwise, the partial derivative of the positive and negative samples is
The constraint of this Loss function is the same as the Triplet Loss, so that the intra-class distance should be much smaller than the inter-class distance. But this loss can be done for multiple samples for sample distance variation and the distance between cluster center and sample is calculated and not a trivial calculation.
Thirdly, after the three-configuration network structure is processed, after the characteristic information is obtained through calculation, a reference sample is calculated through the following calculation formula, and the distance comparison function is solved through pairwise l2 norms of the positive sample and the negative sample:
TripletNet () represents a characteristic distance function obtained after passing through a three-component network structure to obtain a cmp comparator, if the result of the comparator is greater than 0, we can know that the positive and negative samples have high similarity or the negative sample is similar to a reference sample, otherwise, if the difference between the positive and negative samples is greater than or equal to 0, we can know that the positive and negative samples have good identification and classification effects, and the positive sample is better than the reference sample xThis x+Zoom in, negative sample x-Is pushed away.
The invention utilizes a MarKet1501 data set to carry out training and experimental data comparison on the network structure, and the average accuracy and the confidence accuracy can be verified in a complex environment through drawing an average accuracy curve comparison graph of the invention and a classical algorithm on the MarKet1501 data set, a characteristic coupling process graph and a network structure graph of a confidence accuracy data set and a detection effect graph in a related coal-bed gas well.
In the MarKet1501 dataset, we used mean accuracy and confidence accuracy for re-recognition evaluation. The average accuracy is a commonly used evaluation index in target detection and label image classification, and because more than one label is used in a multi-classification task, a single classification standard of a common single-label image, namely mAP, can not be used, namely the average accuracy AP in the multi-classification task is summed and then averaged. Average accuracy mean versus curve plots for the various identification methods are shown in fig. 2 and 3. As can be seen from the graph, compared with other classical algorithms, the coal-bed gas well field monitoring re-identification method technology has better accuracy. The confidence accuracy rate refers to the probability that a correct result exists in a search result with a highest confidence rank, namely the accuracy rate of finding image tags similar to target pedestrians under different cameras. Compared with other algorithms, the coal-bed gas well field monitoring re-identification method has good confidence accuracy, as shown in fig. 4 and 5, and therefore whether the same pedestrian is identified under the condition of crossing cameras can be accurately identified.
The characteristic clustering coupling algorithm and the weighted attention pool used in the invention are shown in the process diagrams of fig. 6 and 7, which can show that the identification accuracy rate can be improved well by the characteristic clustering coupling algorithm and the weighted attention pool in the invention, and the heavy identification can be improved well. The pedestrian identification method is applied to practical projects, pedestrian identification in coal-bed gas wells and oil drilling wells, can effectively help yield and capacity monitoring, intelligent security and protection, safety accident monitoring and the like, and the effect graphs are shown in fig. 8 and 9, so that the pedestrian identification method has good performance and accuracy in heavy identification tasks.
Claims (1)
1. A coal bed gas well site monitoring and re-identification technology is characterized by comprising the following steps:
firstly, constructing an embedded network structure, extracting sample characteristics, adding a weighted attention pooling layer to replace an average pooling layer in the embedded network structure, calculating the inner product between pedestrian characteristic attention heat maps output in pairs by classification calculation, and emphatically extracting the salient characteristics of a target pedestrian;
the network structure consists of an embedded network and a three-configuration network; performing matrix representation on the original average pooling layer by using a weighted attention pool, performing representation by using a matrix of a second-order pool, performing degradation decomposition on the matrix of the weight, and performing bottom-up and top-down attention interpretation by using a degraded result;
the specific implementation mode is as follows:
wherein X ∈ Rn*f,W∈Rf*fX represents a feature matrix, n represents the number of samples, and f represents the number of channels;
for the above WkThe weight matrix is used for carrying out minimum rank approximation, Wk∈Rf*f:
Wherein a, b ∈ Rf*1Wherein a, b represent the row and column of the weight matrix, and k represents the number of categories;
selecting to simulate a class-specific "top-down" attention mechanism from bottom-up visual saliency, independent of class; force one of the attention parameter vectors to be class independent, therefore let bkB, so lastThe low rank attention pool mechanism module is represented as:
Gradeattention(X,a,b)＝(Xak)T(Xbk) (formula 3)
Continuing to simplify as follows:
wherein T isk＝Xak,h＝Xb；
Equation 4 is equivalent to the equation for top-down TkSolving inner products between the attention mapping based on the significance h from bottom to top; average or maximum pool is to reduce the feature matrix to Rf*1Then, generating a classification number through the weight vector W processing of the FC; the method considers the attention thermal diagram as Xa epsilon Rn*1B is considered as a classifier of the attention pool features, so the final GradeattentionSeen as the inner product between two attention heat maps defined over all n spatial locations;
by predicting the first and last feature map identical single-channel bottom-up saliency maps and using a linear classifier at the top, n is also generated1×n2A top-down attention map Xa for the x Num _ class dimension, where Num _ class is the number of classes; multiplying the two attention maps by the spatial average generates an output prediction (Xa) in the Num _ class dimensionk)T(Xbk) These operations are equivalent to first associating the feature with a salient feature XT(Xb)) and then passed through a classifier (a (X)T(Xb))；
Secondly, mapping reference samples, similar samples and heterogeneous sample images into a feature space by integrating three embedded networks into a three-configuration network framework; through a characteristic coupling clustering method, different ID samples with higher confidence are randomly selected to serve as clustering centers, the characteristic points of target pedestrians, background interference points and different pedestrians are correspondingly clustered and distinguished by establishing a heterogeneous sample relative distance threshold value and calculating the space Mahalanobis distance between the characteristic points and the clustering centers to obtain corresponding characteristic coupling clustering losses, and then the similarity is calculated through a comparator; the concrete implementation is as follows:
randomly selecting different ID samples with higher confidence as Cluster centerscenterHowever, only one of all the IDs of the samples can be selected as the cluster center for each selection of the cluster centers, and for each cluster center, the spatial euclidean distance between the peripheral samples and the cluster center is calculated, and a negative sample relative distance threshold d is definednNegative sample dnDefining the shortest distance between the cluster centers as the distance between the cluster centers and other samples SiDistance between, if dist (Cluster)center,Fi)>dnWill be centered on the cluster center, where FiRepresents the ith feature set as dnDividing a Cluster with the distance as the radius, clustering the candidate characteristic points, and selecting min (Cluster)center,Fi))>dnThe farthest point of the data is used as a clustering center again, and the clustering is continued to be divided in the same way, so that all samples are attached with corresponding similar and dissimilar pseudo labels, wherein, a characteristic coupling clustering loss is used for calculating corresponding distances, unlike K-means clustering, data coupling clustering can automatically mark part of data as similar high confidence, while K-means clustering is to divide all data into several clusters, and the clustering center marks similar samples as the same ID; in the clustering process, a plurality of candidate characteristic points probably meet the constraint, and the point with the maximum difference is selected as a clustering center sample point;
min(DistM(fi,fj))>dn
wherein f isi,fjAll belong to the feature map in the image library to be retrieved, ClustercenterIs a cluster center，dnIs a threshold value;
spatial feature points are calculated using mahalanobis distance:
the mahalanobis distance between feature points is calculated as follows:
sigma is a covariance matrix of a multi-dimensional random variable, if the covariance matrix is a unit vector, all dimensions are independently and identically distributed, the mahalanobis distance is the euclidean distance, and the computed euclidean distances are not necessarily similar due to large errors of the euclidean distances in flow pattern spaces under different dimensions, and then a normalized euclidean distance is adopted to eliminate the problem of inconsistent dimension scales, but the normalized euclidean distance cannot well eliminate outler after the difference of sample distribution is larger and larger along with the increase of sample amount, so the mahalanobis distance is adopted to calculate the spatial distance; the mahalanobis distance is specifically as follows:
firstly, rotating the feature points to the principal component to enable all dimensions to become linearly independent and mutually independent, and generating new coordinates at the moment;
the mahalanobis distance is the euclidean distance after the rotation change, and the mahalanobis distance formula 8 becomes:
when DistCluster<dnThen, taking the cluster center as the center, clustering the feature points, dnDistance, otherwise, the distance is pushed far by dnThe distance of (d); for TripletNet (three-state network), since three sample pairs, i.e. reference sample x, positive sample x, are input+Negative sample x-(ii) a Selecting positive sample and negative sample x in positive sample set and negative sample set+,x-Thereby coupling through the above-mentioned featuresClustering, estimating the average value of the cluster centers of each positive sample:
wherein N is+Is the number of positive samples;
the resulting relative distance calculation for positive and negative samples is reflected by the following equation:
where i belongs to the positive sample set and j belongs to the negative sample set, they satisfy the relationship:
Distpostive+θ≤Distnegtive(formula 10)
Where θ is a distance deviation, θ < dn；
Obtaining a coupled clustering loss:
wherein xiIs a set of positive samples, xjIs a set of negative samples, N+Is the number of positive samples;
if it is not
If the partial derivative is less than or equal to 0, the partial derivative of the positive and negative samples is 0, otherwise, the partial derivative of the positive and negative samples is
The constraint of this Loss function is the same as the Triplet Loss, so that the intra-class distance should be much smaller than the inter-class distance;
thirdly, after the three-configuration network structure is processed, after the characteristic information is obtained through calculation, a reference sample is calculated through the following calculation formula, and the distance comparison function is solved through pairwise l2 norms of the positive sample and the negative sample:
TripletNet () represents a characteristic distance function obtained after passing through a three-configuration network structure, and a cmp comparator is obtained, if the result of the comparator is greater than 0, the positive and negative samples are known to have higher similarity or the negative sample is also similar to the reference sample, otherwise, if the difference between the positive and negative samples is greater than or equal to 0, the positive sample x is known to have good identification and classification effects, and compared with the reference sample, the positive sample x+Zoom in, negative sample x-Is pushed away.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110296521.6A CN113011440B (en) | 2021-03-19 | 2021-03-19 | Coal-bed gas well site monitoring and re-identification technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110296521.6A CN113011440B (en) | 2021-03-19 | 2021-03-19 | Coal-bed gas well site monitoring and re-identification technology |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113011440A true CN113011440A (en) | 2021-06-22 |
| CN113011440B CN113011440B (en) | 2023-11-28 |
Family
ID=76403360
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110296521.6A Active CN113011440B (en) | 2021-03-19 | 2021-03-19 | Coal-bed gas well site monitoring and re-identification technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113011440B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114333039A (en) * | 2022-03-03 | 2022-04-12 | 济南博观智能科技有限公司 | Portrait clustering method, device and medium |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109800794A (en) * | 2018-12-27 | 2019-05-24 | 上海交通大学 | A kind of appearance similar purpose identifies fusion method and system across camera again |
| CN110110642A (en) * | 2019-04-29 | 2019-08-09 | 华南理工大学 | A kind of pedestrian's recognition methods again based on multichannel attention feature |
| CN111476168A (en) * | 2020-04-08 | 2020-07-31 | 山东师范大学 | Cross-domain pedestrian re-identification method and system based on three stages |
| CN112101150A (en) * | 2020-09-01 | 2020-12-18 | 北京航空航天大学 | Multi-feature fusion pedestrian re-identification method based on orientation constraint |
| CN112270355A (en) * | 2020-10-28 | 2021-01-26 | 长沙理工大学 | Active safety prediction method based on big data technology and SAE-GRU |
| WO2021029881A1 (en) * | 2019-08-14 | 2021-02-18 | Google Llc | Systems and methods using person recognizability across a network of devices |
| CN112507941A (en) * | 2020-12-17 | 2021-03-16 | 中国矿业大学 | Cross-vision field pedestrian re-identification method and device for mine AI video analysis |
| CN113158901A (en) * | 2021-04-22 | 2021-07-23 | 天津大学 | Domain-adaptive pedestrian re-identification method |
| US20220147735A1 (en) * | 2020-11-10 | 2022-05-12 | Nec Laboratories America, Inc. | Face-aware person re-identification system |
-
2021
- 2021-03-19 CN CN202110296521.6A patent/CN113011440B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109800794A (en) * | 2018-12-27 | 2019-05-24 | 上海交通大学 | A kind of appearance similar purpose identifies fusion method and system across camera again |
| CN110110642A (en) * | 2019-04-29 | 2019-08-09 | 华南理工大学 | A kind of pedestrian's recognition methods again based on multichannel attention feature |
| WO2021029881A1 (en) * | 2019-08-14 | 2021-02-18 | Google Llc | Systems and methods using person recognizability across a network of devices |
| CN111476168A (en) * | 2020-04-08 | 2020-07-31 | 山东师范大学 | Cross-domain pedestrian re-identification method and system based on three stages |
| CN112101150A (en) * | 2020-09-01 | 2020-12-18 | 北京航空航天大学 | Multi-feature fusion pedestrian re-identification method based on orientation constraint |
| CN112270355A (en) * | 2020-10-28 | 2021-01-26 | 长沙理工大学 | Active safety prediction method based on big data technology and SAE-GRU |
| US20220147735A1 (en) * | 2020-11-10 | 2022-05-12 | Nec Laboratories America, Inc. | Face-aware person re-identification system |
| CN112507941A (en) * | 2020-12-17 | 2021-03-16 | 中国矿业大学 | Cross-vision field pedestrian re-identification method and device for mine AI video analysis |
| CN113158901A (en) * | 2021-04-22 | 2021-07-23 | 天津大学 | Domain-adaptive pedestrian re-identification method |
Non-Patent Citations (7)
| Title |
|---|
| ASLAM, MUHAMMAD，等: "Adaptive energy-efficient clustering path planning routing protocols for heterogeneous wireless sensor networks", 《SUSTAINABLE COMPUTING-INFORMATICS & SYSTEMS》, vol. 12, pages 57 - 71 * |
| GUOYIN REN，等: "A Cross-Camera Multi-Face Tracking System Based on Double Triplet Networks", 《IEEE ACCESS》, vol. 9, pages 43759 - 43774, XP011845661, DOI: 10.1109/ACCESS.2021.3061572 * |
| WEIYU ZENG，等: "Clustering-Guided Pairwise Metric Triplet Loss for Person Reidentification", 《IEEE INTERNET OF THINGS JOURNAL》, vol. 9, no. 16, pages 15150 - 15160, XP011916811, DOI: 10.1109/JIOT.2022.3147950 * |
| ZHIGANG CHANG，等: "Distribution Context Aware Loss for Person Re-identification", 《2019 IEEE VISUAL COMMUNICATIONS AND IMAGE PROCESSING (VCIP)》, pages 1 - 4 * |
| 丁嘉婕: "基于质量增强和注意力机制的井下行人重识别算法", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 2, pages 138 - 1902 * |
| 厍向阳，等: "融合随机擦除和残差注意力网络的行人重识别", 《计算机工程与应用》, vol. 58, no. 3, pages 215 - 221 * |
| 唐志鸿，等: "密集场景下基于轮廓运动的人群行为分析", 《江苏通信》, vol. 37, no. 4, pages 102 - 107 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114333039A (en) * | 2022-03-03 | 2022-04-12 | 济南博观智能科技有限公司 | Portrait clustering method, device and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113011440B (en) | 2023-11-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tsintotas et al. | The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection | |
| Lu et al. | Safetynet: Detecting and rejecting adversarial examples robustly | |
| CN107679078B (en) | Bayonet image vehicle rapid retrieval method and system based on deep learning | |
| Murillo et al. | Localization in urban environments using a panoramic gist descriptor | |
| Lynen et al. | Placeless place-recognition | |
| US9275269B1 (en) | System, method and apparatus for facial recognition | |
| Kim et al. | Predicting good features for image geo-localization using per-bundle vlad | |
| CN111709311A (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN106778501B (en) | Video face online identification method based on compression tracking and IHDR incremental learning | |
| Chen et al. | Vehicle re-identification using distance-based global and partial multi-regional feature learning | |
| CN104794219A (en) | Scene retrieval method based on geographical position information | |
| CN113361464B (en) | Vehicle weight recognition method based on multi-granularity feature segmentation | |
| CN111242064A (en) | Pedestrian re-identification method and system based on camera style migration and single marking | |
| CN109933619B (en) | Semi-supervised classification prediction method | |
| Thoma et al. | Soft contrastive learning for visual localization | |
| Liu et al. | Urban area vehicle re-identification with self-attention stair feature fusion and temporal Bayesian re-ranking | |
| Wang et al. | Online visual place recognition via saliency re-identification | |
| Han et al. | A novel loop closure detection method with the combination of points and lines based on information entropy | |
| Cai et al. | Patch-NetVLAD+: Learned patch descriptor and weighted matching strategy for place recognition | |
| CN113011440A (en) | Coal bed gas well field monitoring heavy identification technology | |
| Špaňhel et al. | Vehicle fine-grained recognition based on convolutional neural networks for real-world applications | |
| Jin Kim et al. | Learned contextual feature reweighting for image geo-localization | |
| Liu et al. | Multi-cue fusion: Discriminative enhancing for person re-identification | |
| CN115050044B (en) | Cross-modal pedestrian re-identification method based on MLP-Mixer | |
| Wozniak et al. | Deep embeddings-based place recognition robust to motion blur |
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 |