Flame recognition method and device based on feature fusion, equipment and storage medium

Abstract

The application relates to the technical field of fire monitoring, and discloses a flame recognition method, device and equipment based on feature fusion and a storage medium, the method comprising the following steps: acquiring an infrared video image of a target monitoring area; extracting the video image through a target Gaussian mixture model to obtain a target flame image; determining a corresponding flame feature vector according to the target flame image; and recognizing the flame feature vector through a preset integrated classifier to obtain a corresponding target flame image type; since the target flame image is extracted through the target Gaussian mixture model, the flame feature vector is determined according to the target flame image, and the flame feature vector is recognized through the preset integrated classifier, compared with the prior art, the flame type can be effectively determined with higher accuracy, and the cost and limitations of determining the flame type are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fire monitoring, and in particular to a flame recognition method and device based on feature fusion, equipment and a storage medium. BACKGROUND

[0002] With the continuous improvement of people's safety awareness, fire detection technology is particularly important. Fire detection technology can be used for prediction and early warning at the initial stage of fire, so as to effectively prevent the occurrence of fire and curb its spread and development. At present, the most commonly used equipment for fire detection technology is a smoke and temperature sensing fire detector, and this equipment has been widely used in various places in social life. However, both the smoke and temperature sensing fire detector are contact type fire detectors, and due to the limitation of space in the detection process, there is a certain limitation in the whole detection process. Therefore, video fire detection technology with the advantages of non-contact, rapid response, large detection range and active visibility has emerged. In recent years, it has been widely used in fire monitoring of high and large space buildings such as airports, exhibition halls, workshops and outdoor places such as forests. However, the existing video fire detection technology also has many drawbacks in application, such as the inability to achieve effective monitoring in harsh environments, day and night all-weather monitoring and other functions, resulting in low accuracy of determining the type of flame. Infrared thermal imaging can achieve all-weather imaging and has good application prospects in fire detection, but the existing infrared thermal imaging is mainly used for hot spot detection and cannot accurately identify flames.

[0003] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a flame recognition method, device, equipment and storage medium based on feature fusion, which aims to solve the technical problems of low accuracy, high cost and excessive limitation of determining the type of flame in the prior art.

[0005] To achieve the above purpose, the present application provides a flame recognition method based on feature fusion, which comprises the following steps:

[0006] Obtain an infrared video image of a target monitoring area;

[0007] Extract the video image by a target Gaussian mixture model to obtain a target flame image;

[0008] Determine a corresponding flame feature vector according to the target flame image;

[0009] Identify the flame feature vector by a preset integrated classifier to obtain a corresponding target flame image type.

[0010] Optionally, the extracting from the video image by the target Gaussian mixture model obtains a target flame image, and the method comprises the following steps of:

[0011] obtaining a total number of distribution modes according to the target Gaussian mixture model;

[0012] obtaining a corresponding pixel value according to the video image;

[0013] comparing the total number of distribution modes with the pixel value to obtain a mean deviation;

[0014] when the mean deviation is not located in a target difference range, obtaining a foreground image according to the video image, that is, determining the target flame image.

[0015] Optionally, the determining the corresponding flame feature vector according to the target flame image comprises the following steps of:

[0016] obtaining a corresponding LBP feature and a HOG feature according to the target flame image;

[0017] performing feature fusion on the LBP feature and the HOG feature to obtain the corresponding flame feature vector.

[0018] Optionally, the obtaining the corresponding LBP feature and the HOG feature according to the target flame image comprises the following steps of:

[0019] performing gray scale conversion on the target flame image to obtain a corresponding flame gray scale image;

[0020] determining the corresponding LBP feature according to a target LBP description operator and the flame gray scale image;

[0021] performing gradient calculation on the target flame image according to a target direction by a discrete differential model to obtain corresponding image gradient information;

[0022] determining the corresponding HOG feature according to the image gradient information and the target flame image.

[0023] Optionally, the determining the corresponding LBP feature according to a target LBP description operator and the flame gray scale image comprises the following steps of:

[0024] performing region segmentation on the flame gray scale image according to a target size to obtain a sub-image region of the target size;

[0025] calculating a pixel LBP value in the sub-image region according to a target LBP description operator to obtain the corresponding pixel LBP value;

[0026] counting the pixel LBP value to obtain a target dimension feature vector;

[0027] The target dimension feature vector is normalized, and the normalized target dimension feature vector is spliced to obtain a corresponding LBP feature.

[0028] Optionally, the HOG feature corresponding to the target flame image is determined according to the image gradient information, comprising:

[0029] The target flame image is divided according to a preset pixel size to obtain a target shape unit;

[0030] The target shape unit is equally divided according to the image gradient information to obtain a unit feature vector;

[0031] The target shape units are combined to obtain a target shape unit block;

[0032] The target shape unit block is normalized to obtain a corresponding block feature vector;

[0033] The block feature vector is reduced in dimension by a target component analysis algorithm to obtain a corresponding HOG feature.

[0034] Optionally, the target flame image type is obtained by identifying the flame feature vector through a preset integrated classifier, comprising:

[0035] An initial flame sample data set is obtained in a uniform distribution;

[0036] The initial flame sample data set is trained and learned to obtain a basic classifier;

[0037] The initial flame sample data set is error calculated through the basic classifier to obtain a corresponding current classification error and a classification coefficient;

[0038] The weight distribution of the initial flame sample data set is adjusted according to the current classification error and the classification coefficient to obtain a preset integrated classifier;

[0039] The target flame image type is obtained by identifying the flame feature vector through the preset integrated classifier.

[0040] In addition, in order to achieve the above purpose, the application further provides a flame recognition device based on feature fusion, which comprises:

[0041] An acquisition module is configured to acquire an infrared video image of a target monitoring area;

[0042] An extraction module is configured to extract the video image through a target Gaussian mixture model to obtain a target flame image;

[0043] The determination module is used to determine the corresponding flame feature vector based on the target flame image;

[0044] The recognition module is used to identify the flame feature vector through a preset integrated classifier to obtain the corresponding target flame image type.

[0045] Furthermore, to achieve the above objectives, the present invention also proposes a flame recognition device based on feature fusion, the flame recognition device based on feature fusion comprising: a memory, a processor, and a flame recognition program based on feature fusion stored in the memory and executable on the processor, the flame recognition program based on feature fusion being configured to implement the flame recognition method based on feature fusion as described above.

[0046] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a flame recognition program based on feature fusion, wherein the flame recognition program based on feature fusion, when executed by a processor, implements the flame recognition method based on feature fusion as described above.

[0047] The flame recognition method based on feature fusion proposed in this invention acquires video images of the target monitoring area; extracts target flame images from the video images using a target Gaussian mixture model; determines corresponding flame feature vectors based on the target flame images; and identifies the flame feature vectors using a preset ensemble classifier to obtain the corresponding target flame image type. Because this invention extracts target flame images using a target Gaussian mixture model, determines flame feature vectors based on the target flame images, and then identifies the flame feature vectors using a preset ensemble classifier, compared to existing technologies that detect and determine flame types using contact-based fire detectors, it effectively improves the accuracy of flame type determination and reduces the cost and limitations of flame type determination. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the structure of a flame recognition device based on feature fusion in the hardware operating environment involved in the embodiments of the present invention;

[0049] Figure 2 This is a flowchart illustrating the first embodiment of the flame recognition method based on feature fusion of the present invention;

[0050] Figure 3 This is a flowchart illustrating the second embodiment of the flame recognition method based on feature fusion of the present invention;

[0051] Figure 4 This is a flowchart illustrating the third embodiment of the flame recognition method based on feature fusion of the present invention;

[0052] Figure 5This is a functional module diagram of the first embodiment of the flame recognition device based on feature fusion of the present invention.

[0053] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0055] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a flame recognition device based on feature fusion in the hardware operating environment involved in the embodiments of the present invention.

[0056] like Figure 1 As shown, the feature fusion-based flame recognition device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0057] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on feature fusion-based flame recognition devices and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0058] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a flame recognition program based on feature fusion.

[0059] exist Figure 1The network interface 1004 shown in the feature fusion-based flame identification device is mainly used for data communication with a network integrated platform workstation; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the feature fusion-based flame identification device can be arranged in the feature fusion-based flame identification device, the feature fusion-based flame identification device calls the feature fusion-based flame identification program stored in the memory 1005 through the processor 1001, and executes the feature fusion-based flame identification method provided in the embodiment of the application.

[0060] Based on the above hardware structure, the feature fusion-based flame identification method embodiment of the application is proposed.

[0061] Reference Figure 2 , Figure 2 The flowchart of the first embodiment of the feature fusion-based flame identification method of the application is shown.

[0062] In the first embodiment, the feature fusion-based flame identification method comprises the following steps:

[0063] Step S10, acquiring an infrared video image of a target monitoring area.

[0064] It should be noted that the execution subject of the present embodiment is a feature fusion-based flame identification device, and can also be other devices that can achieve the same or similar functions, such as a monitoring computer, and the present embodiment does not limit this. In the present embodiment, a monitoring computer is taken as an example for description.

[0065] It should be understood that the target monitoring area refers to an area where flames are suspected to appear, and the video image refers to an image of the target monitoring area captured by an imaging device, which can be an infrared thermal imager or other imaging devices, and the present embodiment does not limit this.

[0066] In a specific implementation, the infrared thermal imager monitors the video image of the target area for 24 hours, then encodes the video image through a video encoder, and sends the encoded video image to the monitoring computer to obtain the video image of the target monitoring area.

[0067] Step S20, extracting the video image through a target Gaussian mixture model to obtain a target flame image.

[0068] It can be understood that the target Gaussian mixture model refers to a background model that can effectively overcome periodic changes, each pixel point in the target Gaussian mixture model is composed of a superposition of several Gaussian distributions with different weights, and the target flame image refers to a local area image where flames are suspected to appear, which is a foreground image of the video image of the target monitoring area.

[0069] In a specific implementation, the monitoring computer extracts the foreground image in the video image through the target Gaussian mixture model to obtain the target flame image.

[0070] In step S30, the corresponding flame feature vector is determined according to the target flame image.

[0071] It should be understood that the flame feature vector refers to a feature vector capable of uniquely identifying the type of the target flame image, which is fused from the LBP feature and the HOG feature in the target flame image. Specifically, after obtaining the target flame image, the LBP feature and the HOG feature in the target flame image are extracted, and then the LBP feature and the HOG feature are fused to obtain the corresponding flame feature vector.

[0072] In step S40, the flame feature vector is identified by the preset integrated classifier to obtain the corresponding target flame image type.

[0073] It can be understood that the preset integrated classifier refers to a classifier for identifying fire categories, and the target flame image type refers to the type of the target flame image. The target flame image type includes a flame image type and a non-flame image type. When the target flame image type is the flame image type, the monitoring computer sends an alarm signal and starts the linkage fire extinguishing device. When the target flame image type is the non-flame image type, the next frame of video image is processed.

[0074] Further, in step S40, the following steps are included: obtaining an initial flame sample data set uniformly distributed; training and learning the initial flame sample data set to obtain a basic classifier; calculating the error of the initial flame sample data set by the basic classifier to obtain a corresponding current classification error and a classification coefficient; adjusting the weight distribution of the initial flame sample data set according to the current classification error and the classification coefficient to obtain a preset integrated classifier; identifying the flame feature vector by the preset integrated classifier to obtain the corresponding target flame image type.

[0075] It should be understood that the initial flame sample data set refers to a sample data set composed of foreground image feature vectors and category labels when a fire occurs. The corresponding weight of the initial flame sample data set is uniformly distributed, and the role of the initial flame sample data set in the basic classifier for training and learning is the same. Specifically, the initial flame sample data set is trained and learned to obtain the basic classifier.

[0076] It can be understood that the current classification error refers to the classification error of the basic classifier on the initial flame sample data set, and the classification coefficient refers to the coefficient of the basic classifier in classifying the video image, which is calculated by the following formula:

[0077]

[0078] wherein e m is the current classification error, G m is the gradient value, P is the Gaussian distribution weight, w is the normalized mode weight, I is the number of modes, and x and y refer to the direction angle of each pixel.

[0079]

[0080] wherein a m is the classification coefficient, and e m is the current classification error.

[0081] It should be understood that after obtaining the current classification error and the classification coefficient, the basic classifier is adjusted by the current classification error and the classification coefficient, and the adjustment manner is as follows:

[0082]

[0083] wherein Zm is the normalization factor, w m+1 is the normalized next mode weight, N is the adjustment number, a m is the classification coefficient, and D m+1 is the weight distribution of the adjusted initial flame sample data set.

[0084] The embodiment obtains an infrared video image of a target monitoring area, extracts a target flame image from the video image by using a target Gaussian mixture model, determines a corresponding flame feature vector according to the target flame image, and identifies the flame feature vector by using a preset integrated classifier to obtain a corresponding target flame image type. Compared with the prior art, the embodiment can effectively improve the accuracy of determining the flame type and reduce the cost and limitations of determining the flame type.

[0085] In an embodiment, as Figure 3 described, the second embodiment of the flame recognition method based on feature fusion is proposed based on the first embodiment, and the step S20 comprises:

[0086] Step S201: obtaining the total number of distribution modes according to the target Gaussian mixture model.

[0087] It should be understood that the total number of distribution modes refers to the number of different distribution modes, which is determined by a target Gaussian mixture model, which can be 3-5 or other numbers, and the embodiment does not limit this, and the target Gaussian mixture model is constituted as follows:

[0088]

[0089] Where K is the total number of distribution modes, η(X t ,μ i,t ,∑ i,t ) is the i-th Gaussian distribution at time t, μ i,t is its mean, ∑ i,t is the covariance matrix, is the mean square error of the K-th Gaussian distribution, ω i,t is the weight of the i-th Gaussian distribution at time t.

[0090] Step S202, obtaining corresponding pixel values according to the video image.

[0091] It can be understood that the pixel value refers to the value corresponding to each pixel point in the video image, and the average brightness information of each square in the video image can be determined by the pixel value, specifically, the corresponding pixel information is obtained from the video image, and then the pixel value in the pixel information is extracted.

[0092] Step S203, comparing the total number of distribution modes with the pixel value to obtain a mean deviation.

[0093] It should be understood that the mean deviation refers to the deviation value between the total number of distribution modes and the pixel value, specifically, the total number of distribution modes is compared with the pixel value, and the corresponding mean deviation can be obtained according to the comparison result.

[0094] Step S204, when the mean deviation is not located in the target difference range, obtaining a foreground image according to the video image, that is, determining a target flame image.

[0095] It can be understood that the target difference range refers to the difference range corresponding to the background image, and the target difference range is 0-2.5σ. After obtaining the mean deviation, it is necessary to judge whether the mean deviation is located in the target difference range, if yes, it meets the requirements of the background image, at this time, the image corresponding to the pixel value is the background image, if not, the image corresponding to the pixel value is the foreground image.

[0096] The embodiment obtains the total distribution mode according to the target Gaussian mixture model, obtains the corresponding pixel value according to the video image, compares the total distribution mode with the pixel value to obtain the mean deviation, and obtains the target flame image according to the video image when the mean deviation is not located in the target difference range. Since the embodiment obtains the total distribution mode through the target Gaussian mixture model, obtains the corresponding pixel value according to the video image, compares the total distribution mode with the pixel value, judges whether the mean deviation is located in the target difference range, and obtains the target flame image according to the video image when the mean deviation is not located in the target difference range, the accuracy of obtaining the target flame image can be effectively improved.

[0097] In one embodiment, as Figure 4 described above, the third embodiment of the flame recognition method based on feature fusion is proposed based on the first embodiment, and the step S30 comprises:

[0098] In step S301, the corresponding LBP feature and HOG feature are obtained according to the target flame image.

[0099] It can be understood that the LBP feature refers to a feature for describing the local texture of the target flame image, and the LBP feature has the advantages of rotation invariance and gray invariance. The texture feature of the target flame image can be determined through LBP, and the HOG feature refers to a feature for calculating and counting the gradient direction of the local region of the target flame image.

[0100] Further, in step S301, the target flame image is subjected to gray scale conversion to obtain a corresponding flame gray scale image, the corresponding LBP feature is determined according to the target LBP description operator and the flame gray scale image, the gradient calculation of the target flame image is performed according to the target direction through the discrete differential model to obtain the corresponding image gradient information, and the corresponding HOG feature is determined according to the image gradient information and the target flame image.

[0101] It should be understood that the flame gray scale image refers to an image obtained by converting the target flame image to gray scale, and the size of the target flame image at this time is 128*256. The target LBP description operator refers to an operator for calculating the LBP value of a pixel, and the target LBP operator can be The target LBP description operator is used to calculate the flame gray scale image to obtain the corresponding LBP feature.

[0102] It can be understood that the target direction refers to the direction of gradient calculation of the target flame image, and the target direction includes the horizontal direction and the vertical direction. The discrete differential model refers to a model for gradient calculation of the target flame image, and the discrete differential model can be a one-dimensional discrete differential model (-1, 0, +1) or other dimensional discrete differential models.

[0103] Further, the corresponding LBP feature is determined according to the target LBP description operator and the flame gray image, including: regionally segmenting the flame gray image according to a target size to obtain a sub-image region of the target size; calculating pixels in the sub-image region according to the target LBP description operator to obtain a corresponding pixel LBP value; counting the pixel LBP value to obtain a target dimension feature vector; normalizing the target dimension feature vector, and splicing the normalized target dimension feature vector to obtain the corresponding LBP feature.

[0104] It can be understood that the sub-region image refers to a regionally segmented flame gray image, which can be a 16*16 sub-region image. Then, the target LBP description operator is used to calculate the sub-region image to obtain a pixel LBP value. Then, the pixel LBP value is counted in the form of a histogram to obtain a target dimension feature vector. The target dimension can be 10 dimensions. The normalization of the target dimension feature vector can effectively improve the robustness of the feature vector. At this time, the number of target dimension feature vectors is 8*16. Then, the normalized target dimension feature vector is spliced to obtain an LBP feature. At this time, the dimension of the LBP feature is 1280.

[0105] Further, the corresponding HOG feature is determined according to the image gradient information and the target flame image, including: dividing the target flame image according to a preset pixel size to obtain a target shape unit; equally dividing the target shape unit according to the image gradient information to obtain a unit feature vector; combining the target shape units to obtain a target shape unit block; normalizing the target shape unit block to obtain a corresponding block feature vector; and reducing the dimension of the block feature vector by a target component analysis algorithm to obtain the corresponding HOG feature.

[0106] It should be understood that the preset pixel size refers to the size of the divided target flame image, which can be 8*8. The target shape unit refers to a positive direction unit (cell) obtained by dividing the target flame image. Then, the target shape unit is equally divided according to the gradient direction in the image gradient information to obtain a unit feature vector. The gradient direction can be The unit feature vector refers to a 9-dimensional feature vector. Then, the 2*2 target shape units are combined to obtain a target shape unit block (block). At this time, the dimension of the target shape unit is 36 dimensions. The block feature vector refers to the feature vector of the normalized target shape unit block, which is the final feature vector of the target shape unit block. The target component analysis algorithm refers to an algorithm for reducing the dimension of the vector. The target component analysis algorithm can be a PCA principal component analysis algorithm or other algorithms.

[0107] In step S302, the LBP feature and the HOG feature are fused to obtain a corresponding flame feature vector.

[0108] It should be understood that after the LBP feature and the HOG feature are obtained, the LBP feature is a 1280-dimensional feature vector, and the HOG feature is a 1250-dimensional feature vector. Then, the 1280-dimensional LBP feature and the 1250-dimensional HOG feature are fused to obtain a 2530-dimensional feature vector, which is the corresponding flame feature vector.

[0109] In the embodiment, the corresponding LBP feature and HOG feature are obtained according to the target flame image, the LBP feature and the HOG feature are fused to obtain the corresponding flame feature vector, and the accuracy of obtaining the flame feature vector is effectively improved, and the cost and limitation of determining the flame type are reduced.

[0110] In addition, the embodiment of the present application also provides a storage medium, and the storage medium stores a flame recognition method based on feature fusion. The flame recognition method based on feature fusion is executed by a processor to realize the steps of the flame recognition method based on feature fusion.

[0111] Since the storage medium adopts all the technical solutions of the above-mentioned embodiments, it at least has all the beneficial effects brought by the technical solutions of the above-mentioned embodiments, which will not be repeated here.

[0112] In addition, with reference to Figure 5 , the embodiment of the present application also provides a flame recognition device based on feature fusion, which comprises:

[0113] The acquisition module 10 is configured to acquire an infrared video image of a target monitoring area.

[0114] The extraction module 20 is configured to extract the video image by using a target Gaussian mixture model to obtain a target flame image.

[0115] The determination module 30 is configured to determine a corresponding flame feature vector according to the target flame image.

[0116] The recognition module 40 is configured to recognize the flame feature vector by using a preset integrated classifier to obtain a corresponding target flame image type.

[0117] The embodiment obtains an infrared video image of a target monitoring area; extracts the video image through a target Gaussian mixture model to obtain a target flame image; determines a corresponding flame feature vector according to the target flame image; and identifies the flame feature vector through a preset integrated classifier to obtain a corresponding target flame image type. Compared with the prior art of detecting and determining a flame type through a contact type fire detector, the embodiment can effectively improve the accuracy of determining the flame type and reduce the cost and limitations of determining the flame type.

[0118] It should be noted that the above-described workflow is merely illustrative and does not limit the protection scope of the present application. In actual applications, a person skilled in the art can select part or all of the workflow to achieve the purpose of the embodiment according to actual needs, which is not limited herein.

[0119] In addition, technical details not described in detail in the embodiment can be found in the feature fusion-based flame recognition method provided by any embodiment of the present application, which will not be described herein.

[0120] In an embodiment, the extraction module 20 is further configured to obtain a total number of distribution modes according to the target Gaussian mixture model; obtain a corresponding pixel value according to the video image; compare the total number of distribution modes with the pixel value to obtain a mean deviation; and obtain a foreground image according to the video image when the mean deviation is not located in a target difference range, that is, determine the foreground image as the target flame image.

[0121] In an embodiment, the determination module 30 is further configured to obtain corresponding LBP features and HOG features according to the target flame image; and perform feature fusion on the LBP features and the HOG features to obtain a corresponding flame feature vector.

[0122] In an embodiment, the determination module 30 is further configured to perform grayscale conversion on the target flame image to obtain a corresponding flame grayscale image; determine corresponding LBP features according to a target LBP description operator and the flame grayscale image; perform gradient calculation on the target flame image according to a discrete differential model in a target direction to obtain corresponding image gradient information; and determine corresponding HOG features according to the image gradient information and the target flame image.

[0123] In an embodiment, the determining module 30 is further configured to perform region segmentation on the flame gray image according to a target size to obtain a sub-image region of the target size; calculate a pixel LBP value corresponding to a pixel in the sub-image region according to a target LBP description operator; count the pixel LBP value to obtain a target dimension feature vector; perform normalization processing on the target dimension feature vector, and splice the normalized target dimension feature vector to obtain a corresponding LBP feature.

[0124] In an embodiment, the determining module 30 is further configured to divide the target flame image according to a preset pixel size to obtain a target shape unit; perform average division on the target shape unit according to the image gradient information to obtain a unit feature vector; combine the target shape units to obtain a target shape unit block; perform normalization processing on the target shape unit block to obtain a corresponding block feature vector; and perform dimension reduction on the block feature vector by using a target component analysis algorithm to obtain a corresponding HOG feature.

[0125] In an embodiment, the identifying module 40 is further configured to obtain an initial flame sample data set that is uniformly distributed; perform training learning on the initial flame sample data set to obtain a basic classifier; calculate an error of the initial flame sample data set by using the basic classifier to obtain a corresponding current classification error and a classification coefficient; adjust a weight distribution of the initial flame sample data set according to the current classification error and the classification coefficient to obtain a preset integrated classifier; and identify the flame feature vector by using the preset integrated classifier to obtain a corresponding target flame image type.

[0126] Other embodiments of the flame recognition device based on feature fusion or implementation methods of the present application can refer to the above-mentioned method embodiments, which will not be repeated here.

[0127] In addition, it should be noted that in this document, the terms "comprise", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or system. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of another identical element in the process, method, article or system including the element.

[0128] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0129] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the prior art contribution can be embodied in the form of software products, the computer software product is stored in a storage medium (such as read only memory (Read Only Memory, ROM) / RAM, disk, optical disk), including a number of instructions to make a terminal device (may be a mobile phone, computer, integrated platform workstation, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0130] The above is only the preferred embodiment of the present application, not the patent range of the present application, any equivalent structure or equivalent flow transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A flame recognition method based on feature fusion, characterized in that, The flame recognition method based on feature fusion includes the following steps: Acquire infrared video images of the target monitoring area; The target flame image is obtained by extracting the video image using a target Gaussian mixture model; Determine the corresponding flame feature vector based on the target flame image; The flame feature vector is identified by a preset integrated classifier to obtain the corresponding target flame image type; The step of extracting the target flame image from the video image using a target Gaussian mixture model includes: The total number of distribution patterns is obtained based on the target Gaussian mixture model; The corresponding pixel values ​​are obtained from the video image; The total number of distribution patterns is compared with the pixel value to obtain the mean deviation; When the mean deviation is not within the target difference range, the foreground image is obtained from the video image, which is then determined as the target flame image; Determining the corresponding flame feature vector based on the target flame image includes: The target flame image is converted to grayscale to obtain the corresponding flame grayscale image; The corresponding LBP features are determined based on the target LBP descriptor and the flame grayscale image; The gradient of the target flame image is calculated using a discrete differential model according to the target direction to obtain the corresponding image gradient information; The target flame image is divided into target shape units according to a preset pixel size; The target shape unit is divided into equal parts according to the image gradient information to obtain the unit feature vector; The target shape units are combined to obtain a target shape unit block; The target shape unit block is normalized to obtain the corresponding block feature vector; The dimensionality of the block feature vector is reduced by the target component analysis algorithm to obtain the corresponding HOG features; The LBP and HOG features are fused to obtain the corresponding flame feature vector.

2. The flame recognition method based on feature fusion as described in claim 1, characterized in that, The step of determining the corresponding LBP features based on the target LBP descriptor and the flame grayscale image includes: The flame grayscale image is segmented according to the target size to obtain sub-image regions of the target size; The corresponding pixel LBP value is obtained by calculating the pixels within the sub-image region according to the target LBP descriptor; The LBP values ​​of the pixels are statistically analyzed to obtain the target dimension feature vector; The target dimension feature vector is normalized, and the normalized target dimension feature vectors are concatenated to obtain the corresponding LBP features.

3. The flame recognition method based on feature fusion as described in any one of claims 1 and 2, characterized in that, The step of identifying the flame feature vector through a preset ensemble classifier to obtain the corresponding target flame image type includes: Obtain a uniformly distributed initial flame sample dataset; A basic classifier is obtained by training the initial flame sample dataset. The initial flame sample dataset is used to calculate the error using the basic classifier to obtain the corresponding current classification error and classification coefficient. The weight distribution of the initial flame sample dataset is adjusted based on the current classification error and classification coefficient to obtain a preset ensemble classifier; The flame feature vector is identified by a preset integrated classifier to obtain the corresponding target flame image type.

4. A flame recognition device based on feature fusion, characterized in that, The flame recognition device based on feature fusion includes: The acquisition module is used to acquire infrared video images of the target monitoring area; The extraction module is used to extract the target flame image from the video image using a target Gaussian mixture model; The determination module is used to determine the corresponding flame feature vector based on the target flame image; The identification module is used to identify the flame feature vector through a preset integrated classifier to obtain the corresponding target flame image type; The extraction module is further configured to obtain the total number of distribution patterns based on the target Gaussian mixture model; obtain the corresponding pixel value based on the video image; compare the total number of distribution patterns with the pixel value to obtain the mean deviation; and when the mean deviation is not within the target difference range, obtain the foreground image based on the video image, i.e., determine it as the target flame image. The determining module is further configured to: convert the target flame image to grayscale to obtain a corresponding flame grayscale image; determine the corresponding LBP features based on the target LBP descriptor and the flame grayscale image; perform gradient calculation on the target flame image according to the target direction using a discrete differential model to obtain the corresponding image gradient information; divide the target flame image according to a preset pixel size to obtain target shape units; divide the target shape units equally according to the image gradient information to obtain unit feature vectors; combine the target shape units to obtain target shape unit blocks; normalize the target shape unit blocks to obtain corresponding block feature vectors; reduce the dimensionality of the block feature vectors using a target component analysis algorithm to obtain corresponding HOG features; and fuse the LBP features and HOG features to obtain the corresponding flame feature vectors.

5. A flame recognition device based on feature fusion, characterized in that, The feature fusion-based flame recognition device includes: a memory, a processor, and a feature fusion-based flame recognition program stored in the memory and executable on the processor, wherein the feature fusion-based flame recognition program is configured to implement the feature fusion-based flame recognition method as described in any one of claims 1 to 3.

6. A storage medium, characterized in that, The storage medium stores a flame recognition program based on feature fusion, which, when executed by a processor, implements the flame recognition method based on feature fusion as described in any one of claims 1 to 3.

Images