Building fire early detection method and system based on deep learning image recognition

Abstract

The invention belongs to the field of building fire detection, relates to a deep learning and image recognition technology, is used for solving the problem that fire analysis accuracy and fire alarm timeliness cannot be considered at the same time in the prior art, and particularly relates to a building fire early-stage detection method and system based on deep learning image recognition. The detection platform comprises a target marking module, a verification analysis module, a time sequence analysis module and a database. The target marking module is used for performing target marking analysis on the building fire: recording the internal environment of the building through a high-definition camera, decomposing the recorded video into frame-by-frame analysis images, amplifying the analysis images, segmenting the analysis images into a plurality of analysis areas, and extracting fire feature vectors in the analysis areas; according to the method and the system, closed-loop judgment from feature extraction to risk verification is completed in a single-frame image processing period, and due to the fact that data transmission delay between modules is eliminated, the system can respond to potential fire behaviors in a millisecond level.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of building fire detection, and relates to deep learning and image recognition technology, and particularly relates to a building fire early detection method and system based on deep learning image recognition. BACKGROUND

[0002] The building fire monitoring method based on deep learning image recognition is an advanced method for automatically identifying and alarming in the very early stage of fire (when visible smoke or open fire appears) by analyzing the monitoring camera screen using artificial intelligence technology, which enables the computer to "learn" to identify the visual features of fire like the human eye.

[0003] The invention with the publication number CN117437593A discloses a fire image real-time detection method suitable for large-span space places, which simplifies the ultra-high-definition image deep learning detection into low-resolution image deep learning detection by using far-infrared high-temperature point detection and flame smoke suspected area pre-discrimination method of foreground cumulative image, greatly reducing the complexity of the deep learning detection model and the network parameter size; however, the prior art cannot simultaneously perform fire detection and verification according to a single image, resulting in that the fire analysis accuracy and the timeliness of fire alarm cannot be considered at the same time under the condition of retaining the verification link.

[0004] In view of the above technical problems, the present application provides a solution. SUMMARY

[0005] The application aims to provide a building fire early detection method and system based on deep learning image recognition, which can solve the problem that the prior art cannot consider the fire analysis accuracy and the timeliness of fire alarm at the same time.

[0006] The application needs to solve the technical problem of how to provide a building fire early detection method and system based on deep learning image recognition, which can consider the fire analysis accuracy and the timeliness of fire alarm at the same time.

[0007] The object of the application can be achieved by the following technical solutions.

[0008] The building fire early detection system based on deep learning image recognition comprises a detection platform, which comprises a target marking module, a verification analysis module, a time sequence analysis module and a database.

[0009] The target marking module is used for target marking analysis of building fire: a high-definition camera is used to record the internal environment of the building and the recording is decomposed into analysis images frame by frame, the analysis images are enlarged and segmented into a plurality of analysis regions, fire feature vectors in the analysis regions are extracted, the fire feature vectors include color distribution feature information, texture feature information and shape feature information, a fire feature data set is called from a database, fire feature values of the analysis regions are generated by the fire feature vectors and the fire feature data set, the verification regions are marked by the fire feature values, and the verification regions are sent to the verification analysis module.

[0010] The verification analysis module is used for fire verification analysis of the verification regions: verification coefficients of the verification regions are obtained, and the verification regions are marked as safe regions or warning regions by the verification coefficients.

[0011] The time sequence analysis module is used for time sequence analysis of the building fire risk.

[0012] Further, the process of obtaining the fire feature value of the analysis region includes: comparing the fire feature vector with all data in the fire feature data set one by one and obtaining the cosine similarity of all data in the fire feature data set and the fire feature vector, the feature data set is a set composed of feature vectors extracted from fire scene images in different scenes in historical data and test data, the maximum value of all cosine similarities is marked as the fire feature value of the analysis region, and the feature vector corresponding to the fire feature value in the fire feature data set is marked as the comparison vector.

[0013] Further, the specific process of marking the verification region includes: obtaining a fire feature threshold value from a database, comparing the fire feature value with the fire feature threshold value: if the fire feature value is less than the fire feature threshold value, it is determined that the analysis region does not have fire risk; if the fire feature value is greater than or equal to the fire feature threshold value, it is determined that the analysis region has fire risk, and the corresponding analysis region is marked as the verification region.

[0014] Further, the process of obtaining the verification coefficient of the verification region includes: marking the regions in the analysis region that match the color distribution feature information, the texture feature information and the shape feature information with the comparison vector as color verification regions, texture verification regions and shape verification regions, marking the area value of the overlapping region of the color verification region, the texture verification region and the shape verification region as an overlapping value, marking the ratio of the overlapping value to the area value of the color verification region, the texture verification region and the shape verification region as a color verification value, a texture verification value and a shape verification value, and marking the sum of the color verification value, the texture verification value and the shape verification value as the verification coefficient.

[0015] Further, the specific process of marking the verification area as a safe area or a verification area includes: obtaining a verification threshold value from a database, comparing the verification coefficient with the verification threshold value: if the verification coefficient is less than the verification threshold value, it is determined that the verification area does not have fire characteristics, and the corresponding verification area is marked as a safe area; if the verification coefficient is greater than or equal to the verification threshold value, it is determined that the verification area has fire characteristics, the corresponding verification area is marked as a warning area, a fire warning signal is generated, and the fire warning signal and the warning area are sent to a time sequence analysis module and a mobile terminal of a management personnel.

[0016] Further, the specific process of the time sequence analysis module performing time sequence analysis on the building fire risk includes: magnifying the image of the warning area into a pixel grid image and performing gray scale transformation, obtaining a smoke gray scale range and a flame gray scale range from a database, marking the pixel grid with a gray scale value in the smoke gray scale range and the flame gray scale range as a smoke grid and a flame grid respectively, forming a smoke area from continuous smoke grids and a flame area from continuous flame grids, generating a smoke spread trend and a fire spread trend according to the smoke area and the flame area, and sending the smoke spread trend and the fire spread trend to a mobile terminal of a management personnel.

[0017] Further, the generation process of the smoke spread trend and the fire spread trend includes: extracting the smoke area and the flame area in the analysis image of the warning area in the next L1 frames, sorting all the smoke areas according to the decomposition time sequence of the analysis image to obtain a smoke sequence, sorting all the fire areas according to the decomposition time sequence of the analysis image to obtain a fire sequence, performing time sequence analysis on the pixel brightness or area of the smoke area in the smoke sequence to obtain the smoke spread trend, and performing time sequence analysis on the pixel brightness or area of the fire area in the fire sequence to obtain the fire spread trend.

[0018] The building fire early detection method based on deep learning image recognition includes the following steps:

[0019] Step S1: target marking analysis of building fire: recording the internal environment of the building by a high-definition camera and decomposing the recording into analysis images one by one, magnifying and segmenting the analysis images into a plurality of analysis areas, obtaining fire characteristic values of the analysis areas, and determining whether the analysis areas have fire risks through the fire characteristic values;

[0020] Step S2: fire verification analysis of the verification area: obtaining a verification coefficient of the verification area, and marking the verification area as a safe area or a warning area through the verification coefficient;

[0021] Step S3: Time sequence analysis of building fire risk: Mark the smoke area and fire area in the warning area, extract the smoke area and flame area in the image in the next L1 frame analysis, and generate the smoke spread trend and fire spread trend.

[0022] The present application has the following advantages:

[0023] 1. The present application realizes a closed-loop judgment from feature extraction to risk verification within a single frame image processing period. Since the data transmission delay between modules is eliminated, the system can respond to potential fire within milliseconds. This architecture significantly improves the system's tracking ability of fire spread speed, expands the effective time window of the warning signal, and enables building managers to take timely measures in the early stages of fire development. At the same time, by integrating feature extraction and preliminary verification processes, data redundancy transmission is reduced, and the system's bandwidth occupation and computing resource consumption are decreased. This optimization not only improves the system's real-time performance but also enhances its accuracy and reliability in complex dynamic fire scenarios.

[0024] 2. The present application can consider color, texture, and shape features to improve the accuracy of fire detection. By calculating the area ratio of the overlapping area and each verification area, the consistency between different features can be effectively evaluated, thus more reliably judging potential fire risks. The introduction of the verification coefficient provides a quantitative basis for subsequent fire risk assessment, which helps to reduce false positives and false negatives.

[0025] 3. The present application can realize dynamic tracking and analysis of building fire risk. Through time sequence analysis of the warning area, the fire development situation can be grasped in time, providing more accurate and real-time fire information for firefighters, which helps to develop more targeted fire extinguishing and rescue plans. At the same time, sending the analysis results directly to the manager's mobile terminal can realize rapid response and remote monitoring, improving the efficiency of fire warning and disposal. In addition, this scheme can also reduce the false alarm rate and avoid unnecessary alarms caused by single-frame image judgment errors.

[0026] 4. The present application realizes dynamic time sequence analysis of building fire risk. Through analysis of multiple consecutive frames of images, the changing trend of smoke and flame can be accurately captured, avoiding misjudgment that may be caused by single-frame image analysis. At the same time, smoke and flame are analyzed separately to provide more comprehensive fire information. Using linear regression method to fit the time series can obtain simple and clear trend indicators, which are convenient for managers to quickly judge the fire development situation. This dynamic analysis method can better reflect the real-time development of the fire compared to static analysis, which helps to take effective fire fighting measures in time and improves the efficiency of fire warning and disposal. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0028] Figure 1 The system block diagram of the first embodiment of the present application;

[0029] Figure 2 The method flow chart of the second embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions of the present application will be described in detail below in conjunction with the embodiments. Obviously, the described embodiments only constitute some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0031] In the conventional existing building fire monitoring system, the detection process based on single frame image usually separates feature extraction and verification analysis into independent modules, resulting in time sequence fragmentation during image processing. When the system acquires the video stream of the internal environment of the building through a high-definition camera, it needs to first perform multi-region segmentation and feature extraction on the single frame image, and then transmit the suspicious region to an independent verification module for secondary verification. The multiple transmissions of image data and the communication delay between modules in this process will significantly increase the system response time. Especially when processing high-resolution video streams, the massive analysis regions generated by image segmentation need to be passed through multiple processing units level by level, resulting in that the verification of fire features of key frames lags behind the actual fire development speed.

[0032] For example, in the fire monitoring scene of a large commercial complex, the traditional system deployed uses a 4K resolution camera with 30 frames per second to collect video streams. Each frame of image is segmented into 1024 analysis regions. When a local region with suspected flame appears in a frame, the target detection module needs to transmit all feature vectors to an independent verification server for comparison and analysis. This process produces a communication delay of at least 200 milliseconds. During this period, the flame region in the subsequent video frames may have expanded to adjacent partitions, but the system is still verifying the initial frame, resulting in that the actual fire has entered an irreversible development stage when the warning signal is generated.

[0033] If the above issues are not addressed, the system will be unable to complete the closed-loop judgment of feature extraction and risk verification within a single frame processing cycle, causing early fire features to lose timeliness during cross-module transmission. Redundant data transmission between multiple modules not only consumes network bandwidth resources but also leads to a mismatch between verification results and real-time fire status, causing a simultaneous increase in both false alarm and false negative rates. In dynamic fire scenarios, such architectural flaws will directly weaken the system's ability to track the speed of flame spread, significantly shortening the effective window of early warning signals and failing to meet the millisecond-level response requirements for early fire detection in building safety codes.

[0034] To address the aforementioned issues, this application first considers eliminating data transmission delays between modules. Traditional systems separate target detection and verification analysis into independent modules, resulting in communication delays caused by the need to transmit feature vectors across servers. To resolve this, this application attempts to integrate feature extraction and preliminary verification into the same processing flow, achieving closed-loop judgment within a single frame image by establishing a feature value comparison mechanism. Furthermore, it was found that directly generating quantifiable fire feature values ​​during the target marking stage can avoid data redundancy caused by transmitting the original feature vectors. By constructing a comprehensive evaluation system encompassing multi-dimensional features such as color, texture, and shape, a single module can both perform feature extraction and output risk prediction indicators, thereby reducing cross-module communication links.

[0035] Example 1: As Figure 1 As shown, the building fire early detection system based on deep learning image recognition includes a detection platform, which comprises a target marking module, a verification and analysis module, a time series analysis module, and a database.

[0036] The target marking module is used for target marking analysis of building fire: a high-definition camera is used to record the internal environment of the building and the recording is decomposed into analysis images frame by frame, the analysis images are enlarged and segmented into a plurality of analysis regions, fire feature vectors in the analysis regions are extracted, the fire feature vectors include color distribution feature information, texture feature information and shape feature information, a fire feature data set is called from a database, all data in the fire feature data set are compared with the fire feature vectors one by one, and cosine similarities of all data in the fire feature data set with the fire feature vectors are obtained, the feature data set is a set composed of feature vectors extracted from fire scene images in different scenes in historical data and test data, the maximum value of all cosine similarities is marked as a fire feature value of the analysis region, and a feature vector corresponding to the fire feature value in the fire feature data set is marked as a comparison vector, the fire feature threshold is obtained from the database, and the fire feature value is compared with the fire feature threshold: if the fire feature value is less than the fire feature threshold, it is determined that the analysis region has no fire risk; if the fire feature value is greater than or equal to the fire feature threshold, it is determined that the analysis region has a fire risk, and the corresponding analysis region is marked as a verification region, and the verification region is sent to the verification analysis module.

[0037] In the comparison process, the cosine similarity of the fire feature vector and each feature vector in the data set is calculated one by one, and the maximum value is selected as the fire feature value, and the feature vector corresponding to the maximum value in the data set is marked as the comparison vector.

[0038] Specifically, the fire feature vector is composed of color, texture and shape features of the current analysis region, and the cosine similarity is calculated with the fire feature vectors in the data set in different scenes, so that the similarity between the current feature and the historical fire feature can be accurately measured. For example, when the cosine similarity between the feature vector of the analysis region and the feature vector of a historical fire scene reaches 0.95, the value is selected as the fire feature value, and the corresponding historical feature vector is selected as the comparison vector. This process ensures that the determination of the fire feature value is based on the closest historical data, thereby improving the accuracy of subsequent verification analysis. By selecting the maximum cosine similarity, noise interference can be effectively eliminated, and false judgments caused by partial feature deviation can be avoided, and reliable comparison basis is provided for the verification analysis module.

[0039] As a preferred embodiment, the scheme of the present application is implemented as follows: the detection platform is deployed in the central control room of the building and is connected with high-definition network cameras distributed in various areas of the building. The target marking module uses a deep learning algorithm to process the 4K resolution video stream transmitted by the camera in real time. The video stream is decomposed into single-frame images at a rate of 30 frames per second. Each frame of image is equally divided into 1024 analysis areas. For each analysis area, the system extracts the color distribution histogram of the RGB color space, the texture feature of the gray level co-occurrence matrix, and the shape descriptor of the contour to form a fire feature vector. A large number of feature data sets of fire scenes are pre-stored in the database.

[0040] The process of obtaining the fire feature value of the analysis area includes: comparing the fire feature vector with all data in the fire feature data set one by one and obtaining the cosine similarity of all data in the fire feature data set and the fire feature vector, the feature data set is a set composed of feature vectors extracted from fire scene images in different scenes in historical data and test data, marking the maximum value of all cosine similarities as the fire feature value of the analysis area, and marking the feature vector corresponding to the fire feature value in the fire feature data set as the comparison vector.

[0041] Specifically, first, compare the obtained fire feature vector with each data in the fire feature data set. During the comparison process, calculate the cosine similarity of the fire feature vector and each data in the data set. The formula for calculating the cosine similarity is: ; where A and B represent two vectors, · represents the dot product of the vectors, and ||A|| represents the modulus of vector A.

[0042] Among all the calculated cosine similarity values, the maximum value is selected as the fire feature value of the analysis area. At the same time, the feature vector in the fire feature data set corresponding to this maximum cosine similarity value is marked as the comparison vector for subsequent analysis and use.

[0043] For example, assume that the fire feature vector is [0.8, 0.6, 0.9], and there are 3 data in the fire feature data set: [0.7, 0.5, 0.8], [0.9, 0.7, 0.95], and [0.6, 0.4, 0.7]. The calculation of the cosine similarity gives the results 0.9986, 0.9998, and 0.9962, respectively. The maximum value is 0.9998, so 0.9998 is taken as the fire feature value, and the corresponding [0.9, 0.7, 0.95] is marked as the comparison vector.

[0044] The verification analysis module is configured to perform fire verification analysis on the verification region: color distribution feature information, texture feature information, and shape feature information in the analysis region are matched with the region matching the comparison vector, and the region matching the comparison vector is marked as a color verification region, a texture verification region, and a shape verification region. The area value of the overlapping region of the color verification region, the texture verification region, and the shape verification region is marked as an overlapping value. The ratio of the overlapping value to the area value of the color verification region, the texture verification region, and the shape verification region is marked as a color verification value, a texture verification value, and a shape verification value. The sum of the color verification value, the texture verification value, and the shape verification value is marked as a verification coefficient. The verification threshold is obtained from the database, and the verification coefficient is compared with the verification threshold. If the verification coefficient is less than the verification threshold, it is determined that the verification region does not have fire characteristics, and the corresponding verification region is marked as a safe region. If the verification coefficient is greater than or equal to the verification threshold, it is determined that the verification region has fire characteristics, and the corresponding verification region is marked as a warning region. A fire warning signal is generated and sent to the time sequence analysis module and the mobile terminal of the management personnel.

[0045] The color verification region, the texture verification region, and the shape verification region are generated by matching the corresponding feature information in the comparison vector. The calculation of the overlapping value is based on the spatial overlapping part of the three verification regions. The color verification value, the texture verification value, and the shape verification value reflect the proportion of the overlapping region in each feature dimension. The verification coefficient is a weighted sum of the verification results of the three dimensions. For example, the area of the color verification region is 100 pixels, the area of the texture verification region is 80 pixels, and the area of the shape verification region is 90 pixels. If the area of the overlapping region is 60 pixels, the color verification value is 60 / 100=0.6, the texture verification value is 60 / 80=0.75, the shape verification value is 60 / 90≈0.67, and the verification coefficient is 0.6+0.75+0.67=2.02.

[0046] Specifically, the color distribution feature information, the texture feature information, and the shape feature information are matched with the comparison vector to generate independent verification regions. By calculating the area proportion of the overlapping region of the three, the matching degree of each feature dimension and the fire characteristics can be quantified. The color verification value reflects the consistency of the color feature, the texture verification value reflects the similarity of the texture feature, and the shape verification value reflects the degree of agreement of the shape feature. The sum of the three is the verification coefficient, which can comprehensively evaluate the multi-dimensional matching of the fire characteristics. For example, when the verification coefficient is 2.02 and the verification threshold is 1.8, it is determined that the verification region has fire characteristics and is marked as a warning region. Through the superposition of multi-dimensional verification results, the interference of single feature misjudgment can be effectively eliminated, the accuracy of fire detection is improved, and the real-time performance of the verification link is maintained.

[0047] The fire feature threshold value is calculated by the feature distribution difference between historical fire data and normal environment data, and is used as a reference value for distinguishing risk areas. The comparison between the fire feature value and the threshold value adopts a numerical comparison method, and the verification area screening process is simplified by setting a single decision boundary. When the fire feature value reaches or exceeds the threshold value, the analysis area is marked as a verification area and enters the subsequent verification analysis link; if it does not reach the threshold value, it is directly excluded from the risk and redundant calculation is reduced.

[0048] Specifically, the fire feature threshold value is dynamically updated through the database to ensure its adaptation to different building environments and fire scenarios. In the comparison process, the fire feature value is determined by the maximum cosine similarity of the feature vector and the data set, and the threshold value determination is directly related to the risk probability. When the fire feature value exceeds the threshold value, the system only transmits high-risk areas to the verification analysis module, avoiding the occupation of computing resources by low-risk areas. Through threshold screening, the system prioritizes potential dangerous areas while retaining the verification link, shortens the fire detection response time, and balances the detection accuracy and real-time demand.

[0049] The time sequence analysis module is used for time sequence analysis of building fire risks: the image of the warning area is enlarged into a pixel grid image and subjected to grayscale transformation, the smoke grayscale range and the flame grayscale range are obtained from the database, the pixel grids with grayscale values within the smoke grayscale range and the flame grayscale range are marked as smoke grids and flame grids respectively, the smoke area is composed of consecutive smoke grids, and the flame area is composed of consecutive flame grids, the smoke area and the flame area in the subsequent L1 frames of analysis images in the warning area are extracted, all smoke areas are sorted according to the decomposition time sequence of the analysis images to obtain a smoke sequence, all fire areas are sorted according to the decomposition time sequence of the analysis images to obtain a fire sequence, the smoke spreading trend is obtained by time sequence analysis of the pixel brightness or area of the smoke area in the smoke sequence, the fire spreading trend is obtained by time sequence analysis of the pixel brightness or area of the fire area in the fire sequence, and the smoke spreading trend and the fire spreading trend are sent to the mobile terminal of the management personnel.

[0050] The image of the warning area is converted into a pixel grid image, and each pixel grid is compared with the preset smoke grayscale range and flame grayscale range after grayscale transformation. The marking of smoke grids and flame grids is based on the grayscale range data in the database, and consecutive smoke grids and flame grids form smoke areas and flame areas respectively. The smoke spreading trend and the fire spreading trend are generated by analyzing the area changes in consecutive multiple images, such as time sequence analysis of pixel brightness or area.

[0051] Specifically, the image of the early warning area is enlarged and converted into a high-resolution pixel grid image. After gray scale transformation, the gray scale value of each pixel is mapped to a specific range. Through the pre-stored smoke gray scale range and flame gray scale range in the database, the pixel grid with a gray scale value falling within the corresponding range is classified as a smoke grid or a flame grid. Continuous smoke grids form a smoke area, and continuous flame grids form a flame area. In the subsequent analysis of the image in the continuous L1 frame, the changes in the smoke area and the flame area are extracted and sorted into a time sequence. By analyzing the change trend of the pixel brightness or the area, the spreading direction and speed of the smoke and the flame are predicted. For example, the gradual expansion of the area of the smoke area in the continuous frames or the enhancement of the pixel brightness of the flame area over time are recorded as the trend data. These trend data are sent to the management personnel through the mobile terminal for timely response.

[0052] As a preferred embodiment, the scheme of the application is implemented as follows:

[0053] The specific process of the time sequence analysis module for time sequence analysis of the building fire risk includes: enlarging the image of the early warning area into a pixel grid image and performing gray scale transformation. The smoke gray scale range and the flame gray scale range are obtained through the database. The pixel grids with gray scale values within the smoke gray scale range and the flame gray scale range are marked as smoke grids and flame grids, respectively. The smoke area is composed of continuous smoke grids, and the flame area is composed of continuous flame grids. The smoke spreading trend and the fire spreading trend are generated according to the smoke area and the flame area. The smoke spreading trend and the fire spreading trend are sent to the mobile terminal of the management personnel.

[0054] Specifically, the image of the early warning area can be enlarged into a 1920x1080 pixel high-definition image. The gray scale transformation adopts the RGB-to-gray scale formula: Gray = 0.299xR + 0.587xG + 0.114xB. The smoke gray scale range can be set to 180-220, and the flame gray scale range can be set to 200-255. Continuous smoke grids or flame grids can be defined as at least 8 adjacent pixel points. The smoke area and the flame area can be represented by the minimum circumscribed rectangle. The spreading trend can be obtained by comparing the area changes of the smoke and the flame in the continuous 5 frames of images. The trend information can be sent to the mobile terminal in the form of text description and arrow direction indication.

[0055] As a preferred embodiment, the scheme of the application is implemented as follows:

[0056] The generation process of smoke spread trends and fire spread trends includes: extracting smoke and flame regions from the warning area in the subsequent 10 consecutive frames of analyzed images; sorting all smoke regions according to the decomposition time order of the analyzed images to obtain a smoke sequence; and sorting all fire regions according to the decomposition time order of the analyzed images to obtain a fire sequence. The smoke spread trend is obtained by performing time-series analysis based on the pixel brightness of the smoke regions in the smoke sequence, and the fire spread trend is obtained by performing time-series analysis based on the area of ​​the fire regions in the fire sequence.

[0057] Specifically, to determine the trend of smoke spread, the average pixel brightness value of the smoke region in each frame of the image can be calculated, resulting in a time series containing 10 values. Then, a linear regression method is used to fit this time series, yielding a linear function y=ax+b, where a represents the rate of change of smoke brightness over time. If a is positive, it indicates an increasing trend in smoke brightness; if a is negative, it indicates a decreasing trend in smoke brightness.

[0058] To determine the fire's spread trend, the area of ​​the flame region in each frame of the image can be counted in pixels, resulting in a time series containing 10 values. This time series is then fitted using linear regression, yielding a linear function y = cx + d, where c represents the rate of change of the flame area over time. A positive value for c indicates an expanding flame area, while a negative value indicates a shrinking flame area.

[0059] Finally, the smoke spread trend and fire spread trend are sent to the mobile terminals of the managers in the form of charts or text descriptions, so that the managers can understand the development of the fire in a timely manner and take corresponding measures.

[0060] Example 2: Figure 2 As shown, the early detection method for building fires based on deep learning image recognition includes the following steps:

[0061] Step S1: Target marking analysis of building fire: Record the internal environment of the building using a high-definition camera and decompose the recording into frame-by-frame analysis images. Enlarge the analysis images and divide them into several analysis areas. Obtain the fire feature values ​​of the analysis areas and determine whether the analysis areas have fire risk based on the fire feature values.

[0062] Step S2: Perform fire verification analysis on the verification area: Obtain the verification coefficient of the verification area, and mark the verification area as a safe area or a warning area based on the verification coefficient;

[0063] Step S3: Time sequence analysis on building fire risk: marking the smoke area and fire area in the early warning area, extracting the smoke area and fire area in the early warning area in the subsequent L1 frame analysis images, and generating the smoke spread trend and the fire spread trend.

[0064] The building fire early detection method and system based on deep learning image recognition works as follows: a target marking module records the internal environment of the building through a high-definition camera, and divides the recording into analysis images one by one; the analysis images are enlarged and divided into a plurality of analysis regions; fire feature vectors are extracted from each analysis region, including color distribution feature information, texture feature information and shape feature information; the system retrieves fire feature data sets from a database, compares the fire feature vectors with the fire feature data sets, and generates fire feature values of the analysis regions; the analysis regions are marked according to the fire feature values, and are sent to a verification analysis module; the verification analysis module obtains verification coefficients of the analysis regions, and marks the analysis regions as safe regions or early warning regions according to the verification coefficients; a time sequence analysis module performs time sequence analysis on the building fire risk; this working process realizes closed-loop judgment within a single frame of image, avoids the delay caused by cross-module data transmission, and improves the response speed of the system.

[0065] The above is only an example and description of the structure of the present application, and those skilled in the art can make various modifications or supplements or use similar ways to replace the described specific embodiments, as long as they do not deviate from the structure of the present application or exceed the scope defined by the claims.

[0066] In the description of the present application, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0067] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The preferred embodiments do not describe all the details and limit the present application to the specific embodiments. Obviously, many modifications and changes can be made according to the content of the present application. The present application selects and describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can well understand and utilize the present application. The present application is limited only by the claims and their entire scope and equivalents.

Claims

1. A building fire early detection system based on deep learning image recognition, characterized in that, The system includes a detection platform, which comprises a target marking module, a verification analysis module, a time series analysis module, and a database. The target marking module is used to perform target marking analysis on building fires: it records the interior environment of the building using a high-definition camera and decomposes the recording into frame-by-frame analysis images. The analysis images are magnified and divided into several analysis regions. Fire feature vectors are extracted from the analysis regions. The fire feature vectors include color distribution feature information, texture feature information, and shape feature information. The fire feature dataset is retrieved from the database. Fire feature values ​​for the analysis regions are generated using the fire feature vectors and the fire feature dataset. The verification regions are marked using the fire feature values ​​and then sent to the verification analysis module. The verification analysis module is used to perform fire verification analysis on the verification area: obtain the verification coefficient of the verification area, and mark the verification area as a safe area or a warning area based on the verification coefficient; The time-series analysis module is used to perform time-series analysis on building fire risks.

2. The building fire early detection system based on deep learning image recognition according to claim 1, characterized in that, The process of obtaining fire feature values ​​for the analysis area includes: comparing the fire feature vector with all data in the fire feature dataset one by one and obtaining the cosine similarity between all data in the fire feature dataset and the fire feature vector. The feature dataset is a collection of feature vectors extracted from fire scene images under different scenarios in historical data and test data. The maximum value of all cosine similarities is marked as the fire feature value of the analysis area, and the feature vector in the fire feature dataset corresponding to the fire feature value is marked as the comparison vector.

3. The building fire early detection system based on deep learning image recognition according to claim 2, characterized in that, The specific process of marking the verification area includes: obtaining the fire feature threshold from the database, comparing the fire feature value with the fire feature threshold; if the fire feature value is less than the fire feature threshold, the analysis area is determined to have no fire risk; if the fire feature value is greater than or equal to the fire feature threshold, the analysis area is determined to have fire risk, and the corresponding analysis area is marked as the verification area.

4. The building fire early detection system based on deep learning image recognition according to claim 3, characterized in that, The process of obtaining the verification coefficient of the verification region includes: marking the regions in the analysis region that match the color distribution feature information, texture feature information, and shape feature information with the comparison vector as color verification region, texture verification region, and shape verification region, respectively; marking the area value of the overlapping region of the color verification region, texture verification region, and shape verification region as the overlap value; marking the ratio of the overlap value to the area value of the color verification region, texture verification region, and shape verification region as the color verification value, texture verification value, and shape verification value, respectively; and marking the sum of the color verification value, texture verification value, and shape verification value as the verification coefficient.

5. The building fire early detection system based on deep learning image recognition according to claim 4, characterized in that, The specific process of marking a verification area as a safe area or a verification area includes: obtaining the verification threshold from the database, comparing the verification coefficient with the verification threshold; if the verification coefficient is less than the verification threshold, it is determined that the verification area does not have fire characteristics, and the corresponding verification area is marked as a safe area; if the verification coefficient is greater than or equal to the verification threshold, it is determined that the verification area has fire characteristics, the corresponding verification area is marked as a warning area, a fire warning signal is generated, and the fire warning signal and the warning area are sent to the time series analysis module and the mobile terminal of the management personnel.

6. The building fire early detection system based on deep learning image recognition according to claim 5, characterized in that, The specific process of the time-series analysis module for analyzing building fire risks includes: enlarging the image of the warning area into a pixel grid image and performing grayscale transformation; obtaining the grayscale range of smoke and flame through the database; marking the pixels whose grayscale values ​​are within the grayscale range of smoke and flame as smoke grids and flame grids, respectively; forming a smoke area by continuous smoke grids and a flame area by continuous flame grids; generating smoke spread trends and fire spread trends based on the smoke and flame areas; and sending the smoke spread trends and fire spread trends to the mobile terminals of management personnel.

7. The building fire early detection system based on deep learning image recognition according to claim 6, characterized in that smoke... The generation process of the spread trend and fire spread trend includes: extracting the smoke and flame regions in the subsequent L1 frame analysis images of the warning area; sorting all the smoke regions according to the decomposition time order of the analysis images to obtain the smoke sequence; sorting all the fire regions according to the decomposition time order of the analysis images to obtain the fire sequence; performing time series analysis based on the pixel brightness or area of ​​the smoke regions in the smoke sequence to obtain the smoke spread trend; and performing time series analysis based on the pixel brightness or area of ​​the fire regions in the fire sequence to obtain the fire spread trend.

8. A method for early detection of building fires based on deep learning image recognition, characterized in that, Includes the following steps: Step S1: Target marking analysis of building fire: Record the internal environment of the building using a high-definition camera and decompose the recording into frame-by-frame analysis images. Enlarge the analysis images and divide them into several analysis areas. Obtain the fire feature values ​​of the analysis areas and determine whether the analysis areas have fire risk based on the fire feature values. Step S2: Perform fire verification analysis on the verification area: Obtain the verification coefficient of the verification area, and mark the verification area as a safe area or a warning area based on the verification coefficient; Step S3: Perform time-series analysis on building fire risk: Mark the smoke and fire areas in the warning area, extract the smoke and flame areas in the subsequent L1 frame analysis images of the warning area, and generate smoke spread trend and fire spread trend.

Images