Fire image identification method and system

Abstract

The invention relates to the technical field of general image data processing or generation, in particular to a fire image recognition method and system, and the method comprises the steps: obtaining a fire scene target image, carrying out the preprocessing of the fire scene target image, and generating a fire scene image data set; executing optimization on the semantic segmentation network model U-Net; based on the fire scene image data set and a cross pseudo-supervision method, training the optimized U-Net network model; and based on the optimized U-Net network model, performing fire scene image recognition on the fire scene. According to the invention, an image recognition method for a complex fire hazard condition is provided based on the optimized U-Net network model, and the problems of insufficient training data, poor model recognition effect and the like in the prior art are solved; meanwhile, the method can effectively utilize limited annotation data, improves the working efficiency of fire image detection, and has relatively high universality and expansibility.

Description

Technical Field

[0001] The present application relates to the general field of image data processing or generation technology, and in particular to a fire image recognition method and a system. Background Art

[0002] Fire is one of the disasters that are prone to occur in daily life. In recent years, fires in my country have caused serious losses of life and property. Therefore, by detecting fire images, we can speculate on the cause of the fire, provide important basis for fire investigation, and thus protect people's lives and property.

[0003] Traditional image processing methods mainly rely on color models and edge detection algorithms, which perform detection by analyzing the color features and shapes of flames, smoke, and metals. For example, the color of the flame can be separated using the HSV color space, and the outline of the object can be further determined by combining edge detection algorithms (such as Canny and Sobel). However, traditional image processing methods are too dependent on the color and shape features of the image and are difficult to deal with fire scenes under complex backgrounds. This is because the scenes after the fire usually contain a large number of irregular objects and complex lighting conditions, so traditional methods are prone to false detection and missed detection when dealing with these complex situations. At the same time, there are multimodal methods. In fire scenes, a single visible light image often cannot provide enough information. Multimodal methods that combine infrared imaging and visible light images can significantly improve the accuracy and robustness of detection. Infrared imaging can detect objects hidden in smoke or darkness, especially metal objects and other residues. Although multimodal methods can provide richer information, infrared imaging equipment is expensive, and in practical applications, how to efficiently fuse multiple modal information remains a challenge. The complexity of multimodal information fusion increases the difficulty of algorithm design, and the acquisition of infrared imaging data is greatly affected by environmental conditions, such as smoke density and temperature. Summary of the invention

[0004] The inventors found through research that in recent years, deep learning methods have made significant progress in image content detection, especially in the detection of post-fire scenes. The following models are widely used. The first is YOLO (You Only LookOnce), a real-time target detection algorithm that can quickly detect flames, smoke and metal objects in fire scenes. The advantage of YOLO lies in its balance between speed and detection accuracy. Then there is Faster R-CNN, which generates candidate regions through the region proposal network (RPN) and performs target detection, which is suitable for post-fire image analysis tasks with high accuracy requirements; there are also U-Net and MaskR-CNN, which perform well in image tasks and can finely identify targets in fire images, such as metal residues and other objects. Although the deep learning method has improved accuracy, it has high requirements for computing resources and data volume. Due to the difficulty in obtaining data for fire scenes and the complexity and changeability of the scenes, the generalization ability of the model is insufficient. In addition, training deep learning models usually requires a large amount of labeled data, while labeled data for post-fire scenes is often scarce. Therefore, due to insufficient fire data, some semi-supervised learning methods can also be used to recognize fire images. They combine the characteristics of supervised learning and unsupervised learning, mainly by utilizing a small amount of labeled data and a large amount of unlabeled data to improve the learning effect of the model.

[0005] The purpose of this application is to provide a fire image recognition method and system, which solves the technical problem that the existing technology cannot provide a method for quickly processing fire images in complex scenes by combining a fire scene image dataset and a cross-pseudo-supervision method.

[0006] According to one aspect of the present application, a fire image recognition method is provided, comprising:

[0007] Acquire fire scene target images, perform preprocessing on the fire scene target images, and generate fire scene image data sets;

[0008] Optimize the semantic segmentation network model U-Net;

[0009] Based on the fire scene image dataset and the cross pseudo-supervision method, the optimized U-Net network model is trained;

[0010] Based on the optimized U-Net network model, fire scene image recognition is performed on the fire scene.

[0011] In some embodiments, the fire scene target image includes at least a scene image and a fire image, wherein the scene image includes at least a weather condition image and a lighting condition image, and the fire image includes at least a fire scene image, a video frame, and an image generated by a simulated fire.

[0012] In some embodiments, the preprocessing of the fire scene image specifically includes at least performing the following operations on the fire scene image: normalization processing and data enhancement.

[0013] In some embodiments, the normalization processing includes: unifying the color of the fire scene images, unifying the pixel values ​​of the fire scene images to the same range, and adjusting the fire scene images to the same resolution, and the adjustment of the fire scene images to the same resolution is performed using bilinear difference.

[0014] In some embodiments, the data enhancement includes: performing geometric changes on the fire scene image, and the geometric changes include at least rotation, flipping, translation, and scaling.

[0015] In some embodiments, the optimization of the semantic segmentation network model U-Net specifically includes:

[0016] Based on the U-Net network model structure and combined with residual connections, the ResU-Net model is generated;

[0017] Based on the U-Net network model structure, combined with Transformer, the SwinU-Net model is generated;

[0018] The ResU-Net model and SwinU-Net model are optimized U-Net network models.

[0019] In some embodiments, the training of the optimized U-Net network model based on the fire scene image dataset and the cross pseudo-supervision method at least includes:

[0020] Based on the pseudo labels generated by the ResU-Net model, the SwinU-Net model is trained;

[0021] Based on the pseudo labels generated by the SwinU-Net model, the ResU-Net model is trained.

[0022] In some embodiments, the pseudo-labels generated by ResU-Net are used to train the SwinU-Net model, specifically:

[0023] The ResU-Net model is based on the pseudo-labeled dataset generated by the SwinU-Net model Perform training on unlabeled data and update its parameters; where D b is the pseudo-label dataset of the SwinU-Net model, x i is the i-th labeled image, It is the pseudo label generated by SwinU-Net model training;

[0024] The SwinU-Net model is based on the pseudo-labeled dataset generated by the ResU-Net model Perform training on unlabeled data and update its parameters; where D a is the pseudo-label dataset of the ResU-Net model, x i is the i-th labeled image, are pseudo labels generated by ResU-Net model training.

[0025] In some embodiments, the fire scene image recognition is performed on the fire scene based on the optimized U-Net network model, specifically including:

[0026] The future fire scene image is input into the optimized U-Net network model, and the encoder part of the optimized U-Net network model is used to perform feature extraction on the fire scene image;

[0027] The optimized U-Net network model decoder performs feature fusion and upsampling operations. The ResU-Net model gradually fuses and upsamples multi-level features through the decoder to restore the spatial resolution and generate predictive recognition results. The SwinU-Net model combines multi-scale global feature information through the decoder to generate predictive recognition results.

[0028] According to another aspect of the present application, a fire image recognition system comprises:

[0029] An acquisition module, the acquisition module is used to acquire a fire scene target image, perform preprocessing on the fire scene target image, and generate a fire scene image data set;

[0030] An optimization module, wherein the optimization module is used to optimize the semantic segmentation network model U-Net;

[0031] A training module, wherein the training module is used to perform training on the optimized U-Net network model based on a fire scene image dataset and a cross pseudo-supervision method;

[0032] The recognition module is used to perform fire scene image recognition on the fire scene based on the optimized U-Net network model.

[0033] Compared with the prior art, this application has the following advantages and beneficial effects:

[0034] This application provides an image recognition method for complex fire situations based on the optimized U-Net network model, overcoming the problems of insufficient training data and poor model recognition effect in the prior art; at the same time, the method of this application can effectively utilize limited labeled data, improve the work efficiency of fire image detection, and has strong versatility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 is a flow chart of the fire image recognition method of the present application;

[0037] Figure 2 is a structural diagram of the fire image recognition system of the present application;

[0038] Figure 3 This is a usage scenario diagram of the fire image recognition method of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0040] This embodiment provides a fire image recognition method and system. Figure 1-3 Describe them together.

[0041] Application Overview

[0042] Due to the lack of fire data, some semi-supervised learning methods, such as self-training and consistency regularization, can also be used to identify fire images. The aforementioned learning method combines the characteristics of supervised learning and unsupervised learning, and improves the learning effect of the model by using a small amount of labeled data and a large amount of unlabeled data. The aforementioned semi-supervised method also has some shortcomings in fire image recognition, especially in identifying metals and their properties in images, such as the problem of pseudo-label noise. The aforementioned method relies on the initial prediction results of the model to generate pseudo-labels.

[0043] In a fire scene, due to the interference of smoke, flames and complex background, the prediction results of the initial model may be inaccurate, resulting in a lot of noise in the pseudo-labels. These erroneous pseudo-labels will affect subsequent training and reduce the performance of the model; at the same time, the aforementioned learning methods are often not designed for the particularity of fire scenes and may fail to capture complex attribute information.

[0044] For complex fire images, general semi-supervised methods usually focus on basic classification or segmentation tasks and have difficulty in handling complex attribute prediction.

[0045] Therefore, based on the foregoing content, this application provides a method for rapid processing of fire images in complex scenarios by combining the fire scene image dataset and the cross-pseudo-supervision method, and based on the optimized U-Net network model.

[0046] Exemplary methods:

[0047] A fire image recognition method is executed by a controller, comprising: acquiring a fire scene target image, performing preprocessing on the fire scene target image, and generating a fire scene image data set; optimizing a semantic segmentation network model U-Net; training the optimized U-Net network model based on the fire scene image data set and a cross pseudo-supervision method; and performing fire scene image recognition on the fire scene based on the optimized U-Net network model.

[0048] It should be noted that the semantic segmentation network model U-Net is optimized using a cross-pseudo-supervision method. The prediction results of two different network model structures can be used as pseudo-labels to supervise each other, realizing end-to-end learning between the two network models. During the training process, the labeled data are supervised by the real labels respectively; for the unlabeled data, they are supervised by the pseudo-labels generated by the other network model. The semi-supervised method and the optimized U-Net model are integrated to improve the recognition accuracy and robustness, realize rapid processing of complex fire images, and solve the problem of high cost of manual annotation of fire images.

[0049] Below, each part of the application method will be described in detail:

[0050] First, the target image of the fire scene includes at least a scene image and a fire image, wherein the scene image includes at least a weather condition image and a lighting condition image, wherein the weather condition image includes images of sunny days, rainy days, and windy days; the lighting condition image includes images of daytime, nighttime, strong light, and weak light. The fire image includes at least a fire scene image, a video frame, and an image generated by a simulated fire. In practice, since there will be different objects in the fire scene, especially different types of metals, it is also necessary to pay attention to the various forms, combustion states, different burning times, and degrees of metals in the fire scene to ensure that the model of this application can cope with complex real-world environments. The aforementioned fire scene target images can be obtained from fire research institutions, public security departments, and academic research, or they can be collected from real scenes, or image data can be collected from actual fire rescue scenes, including photos of the fire accident scene, images recorded by firefighters' helmet cameras, etc.

[0051] In some embodiments, performing preprocessing on the fire scene image specifically includes performing at least the following operations on the fire scene image: normalization processing and data enhancement. Ideally, the normalization processing includes: unifying the color of the fire scene image, unifying the pixel values ​​of the fire scene image to the same range, adjusting the fire scene image to the same resolution, and adjusting the fire scene image to the same resolution using bilinear interpolation to perform adjustment; in practice, in order to ensure that the data can be efficiently used by the model and to improve the image quality, all images will be color normalized during the normalization processing so that the pixel values ​​of the image are in the same range, for example, the pixel values ​​can be between 0-1, and such pixels help the model to perform more stably when processing images of different brightness and contrast, and all images will be adjusted to the same resolution, for example, the resolution can be 512 pixels by 512 pixels, to ensure that the model can be trained with a consistent input scale, and the aforementioned resolution adjustment can be completed using bilinear interpolation or other interpolation methods, and bilinear interpolation is preferred in this implementation. Data enhancement includes: performing geometric changes on fire scene images, and the geometric changes include at least rotation, flipping, translation, scaling, etc., the purpose of which is to increase data diversity and prevent model overfitting.

[0052] In some other equivalent implementations, specific transformations such as random cropping, perspective transformation, etc. can also be applied to fire images to simulate scenes from different perspectives; the images can also be processed by color jittering, brightness adjustment, contrast change, etc., so that the model can better adapt to fire scenes under different lighting conditions; Gaussian noise can also be added to the image to simulate sensor noise or environmental noise and enhance the robustness of the model under noisy conditions.

[0053] Some optional implementations are to annotate some of the collected images in detail, including the metal position, outline and related properties of the metal in the fire image, such as burning time, morphological changes, etc. Annotation can be completed through manual annotation or semi-automatic annotation tools to improve efficiency and accuracy. At the same time, according to experimental requirements, a certain amount of images are annotated as training data for the model. These annotated data will be used to supervise the initial training of the model and provide accurate learning signals for the model. At the same time, unlabeled fire images are collected for pseudo-label generation in semi-supervised learning. It should be noted that the aforementioned data should cover various fire scenes as much as possible to increase the diversity and effectiveness of pseudo-labels.

[0054] The above content not only ensures the accuracy and generalization ability of the model during recognition, but also lays a technical foundation for the possible accurate recognition of metals and their properties in fire images.

[0055] In some implementations, the semantic segmentation network model U-Net is optimized, specifically including:

[0056] Based on the U-Net network model structure and combined with residual connections, the ResU-Net model is generated;

[0057] Based on the U-Net network model structure, combined with Transformer, the SwinU-Net model is generated;

[0058] The ResU-Net model and the SwinU-Net model are optimized U-Net network models.

[0059] First of all, it should be noted that

[0060] U-Net is a deep learning model for image processing. The model combines the feature extraction capabilities of convolutional neural networks and the pixel-level predictions of fully convolutional networks, and can effectively perform various operations on images. Its structure is similar to that of an autoencoder, consisting of an encoder and a decoder. The encoder is used to extract the features of the image, while the decoder is used to map these features back to the pixel-level predictions of the original image. In the decoder, U-Net uses jump connections to connect the features in the encoder with the features in the decoder, thereby helping the network better learn the details and contextual information of the image;

[0061] The ResU-Net model is a deep learning model that combines residual connections and U-Net structures. It is improved on the basis of U-Net by adding residual connections. The model structure and improvements are reflected in the introduction of residual networks, which makes gradient transfer smoother and may improve model performance.

[0062] The SwinU-Net model is a model that combines the Transformer and U-Net model structures. Its Transformer architecture based on the local attention mechanism can effectively process large-scale image data. Its model structure and network principle are different from the traditional U-Net model. The network structure uses the Transformer network and uses its self-attention mechanism to improve model performance.

[0063] Furthermore, the difference between the ResU-Net model and the ordinary U-Net model lies mainly in the residual connection of the ResU-Net model. The residual network is a deep neural network architecture composed of jump connections, which can alleviate the gradient vanishing problem in deep networks and promote the learning of deeper features. In other words, in this architecture, each block includes a convolutional layer, followed by a residual connection, adding the output of the convolutional layer to the input, and then passing through a pooling layer, which can effectively alleviate the problem of gradient vanishing and is more conducive to training deep networks. In this embodiment, an attention mechanism, namely CBAM, is added to the first layer of the network and the last layer of the network convolution layer.

[0064] The SwinU-Net model introduces the Swin Transformer into the U-Net model architecture. By pre-training on large-scale datasets, the SwinU-Net model can learn rich image features and patterns, thereby improving the generalization ability and accuracy of the model on specific tasks. The pre-trained model can be used as an initialization starting point to help the model converge to the optimal solution faster, reducing training time and resource consumption, and improving the performance of the model on the target task.

[0065] In some implementations, based on the fire scene image dataset and the cross pseudo-supervision method, training is performed on the optimized U-Net network model, which at least includes:

[0066] Based on the pseudo labels generated by the ResU-Net model, the SwinU-Net model is trained;

[0067] Based on the pseudo labels generated by the SwinU-Net model, the ResU-Net model is trained.

[0068] Furthermore, based on the pseudo labels generated by ResU-Net, the SwinU-Net model is trained as follows:

[0069] The ResU-Net model is based on the pseudo-labeled dataset generated by the SwinU-Net model Perform training on unlabeled data and update its parameters; where D b is the pseudo-label dataset of the SwinU-Net model, x i is the i-th labeled image, It is the pseudo label generated by SwinU-Net model training;

[0070] The SwinU-Net model is based on the pseudo-labeled dataset generated by the ResU-Net model Perform training on unlabeled data and update its parameters; where D a is the pseudo-label dataset of the ResU-Net model, x i is the i-th labeled image, are pseudo labels generated by ResU-Net model training.

[0071] In some other feasible implementations, based on the fire scene image dataset and the cross pseudo-supervision method, the optimized U-Net network model is trained, and the following contents may also be included: simultaneously using labeled data for training, the labeled dataset is in is the i-th labeled image, y iis its corresponding label. The ResU-Net and SwinU-Net models are trained with labeled data. In this implementation, the Dice loss function is used as an auxiliary loss function together with the cross entropy loss function to improve the performance of the model in the task. The loss function formula for labeled data is as follows:

[0072] L N =L BCE (p i ,y i )+L Dice (p i ,y i)

[0073] Among them, L BCE , L Dice They are the cross entropy loss function and the Dice loss function; p i is the model's response to the input image The output result of the prediction; y i is the input image The real label.

[0074] In some other feasible implementations, based on the fire scene image dataset and the cross pseudo-supervision method, the optimized U-Net network model is trained, which may also include the following: using the two models trained with labeled data to add pseudo labels to the unlabeled data at the same time, assuming that the unlabeled dataset is D = {x i In the scenario of semi-supervised learning in this embodiment, the argmax function is used to generate pseudo labels, that is, the classification with the highest probability in the model's prediction of unlabeled data is selected as the pseudo label for training. This embodiment will use the argmax function as the pseudo label generation formula. When an unlabeled image x is input i When , the semi-supervised framework combined with the cross pseudo-supervision technique generates two prediction results:

[0075]

[0076]

[0077] in, Represents the input image x i The prediction results of ResU-Net and SwinU-Net are then and The prediction results of , the pseudo-labels are generated by a semi-supervised framework combining cross-pseudo-supervision techniques:

[0078]

[0079]

[0080] in, are pseudo labels generated by ResU-Net and SwinU-Net training.

[0081] In some other feasible implementations, based on the fire scene image dataset and the cross pseudo-supervision method, the optimized U-Net network model is trained, and the following contents may also be included:

[0082] After training and optimization with labeled and unlabeled data sets, the average of the evaluation parameters of the two models is used as the final evaluation result of the entire semi-supervised model. The calculation method of MeanIntersection over Union, the core evaluation parameter of the entire semi-supervised model, is as follows:

[0083] IoU i =TP i / (TP i +FP i +FN i )

[0084] Among them, TP i Represents the number of true positive examples of category i, the number of pixels correctly predicted by the model as category i; FP i Represents the number of false positives of category i, the number of pixels where the model mistakenly predicts other categories as category i; FN i Represents the number of false negatives for category i, the number of pixels where the model mistakenly predicts category i as other categories;

[0085] Miou=(IoU1+IoU2+...+IoU n ) / n

[0086] Among them, n is the number of categories in the image, and the value range of Miou is between 0 and 1. The closer the value is to 1, the higher the degree of overlap between the model's segmentation result and the true segmentation result, and the better the performance.

[0087] It should also be noted that

[0088] The pseudo-label method is a commonly used technique in semi-supervised learning. It generates prediction labels, namely pseudo-labels, for unlabeled data to use these unlabeled data for model training, thereby improving the generalization ability of the model. It first uses a small amount of labeled data to train an initial model, and then uses this model to predict the unlabeled data. The prediction results are used as pseudo-labels, that is, they are regarded as true labels, and are used together with the original labeled data for further training of the model.

[0089] Cross-pseudo-supervision, or CPS for short, is a technique used in semi-supervised learning, particularly for tasks such as image processing. This method improves the performance of models by leveraging the outputs of two or more models as pseudo-labels for supervised learning. During training, the output of one model is used as the pseudo-label of another model, and vice versa. The core idea of ​​CPS is to improve the performance of models through mutual supervision of models. Since each model generates pseudo-labels during training, these pseudo-labels can be used as additional supervisory information to guide the training of other models. In this way, the model can learn more feature representations, thereby improving its generalization ability.

[0090] In some embodiments, fire scene image recognition is performed on the fire scene based on the optimized U-Net network model, specifically including: inputting the future fire scene image into the optimized U-Net network model, performing feature extraction on the fire scene image through the encoder part of the optimized U-Net network model, performing feature fusion and upsampling operations through the optimized U-Net network model decoder, and then, identifying the processed feature map through the output layer of the optimized U-Net network model to obtain the segmentation boundaries and category information of each target area in the fire image, especially the recognition results of metal objects and their related attributes, specifically including the position of the metal objects, the segmentation contour and its attribute information, such as the burning time, etc.

[0091] Among them, the ResU-Net model gradually fuses and upsamples multi-level features through the decoder to restore the spatial resolution and generate predictive recognition results; the SwinU-Net model combines multi-scale global feature information through the decoder to generate predictive recognition results.

[0092] Among them, combining the recognition results of ResU-Net and SwinU-Net, the model outputs obtained by the two models are: and The average of the output results of the two models is taken as the final output result of the entire semi-supervised model, as shown in the following formula:

[0093]

[0094] Based on the above, this application combines the advantages of ResU-Net and SwinU-Net, has the ability to accurately capture image details, and can effectively understand the global information in the fire scene. At the same time, this application adopts a semi-supervised learning method with cross pseudo-supervision, making full use of a large amount of unlabeled fire image data and reducing the dependence on a large amount of labeled data. In addition, by introducing diverse data collection and image preprocessing strategies in the data preparation stage, it can adapt to the complex situations in various fire scenes and enhance the adaptability to complex scenes. Regardless of different lighting conditions, smoke concentrations, or different fire types and metal burning states, this application can robustly detect and identify target objects and their attributes in the image.

[0095] At the same time, it is worth noting that the cross-pseudo-supervised training method and U-Net model structure used in this application have good versatility and are suitable for other tasks that require target detection and attribute recognition in complex scenes. This means that this method is not only limited to fire image recognition, but can also be extended to other fields, such as natural disaster monitoring, industrial safety detection, etc.

[0096] In order to better illustrate the method of the present application and facilitate understanding, the present application also proposes a system, a fire image recognition system, which includes:

[0097] An acquisition module is used to acquire a fire scene target image, perform preprocessing on the fire scene target image, and generate a fire scene image data set;

[0098] Optimization module, which is used to optimize the semantic segmentation network model U-Net;

[0099] A training module, which is used to train the optimized U-Net network model based on the fire scene image dataset and the cross pseudo-supervision method;

[0100] Recognition module,The recognition module is used to perform fire scene image recognition on the fire scene based on the optimized U-Net network model.

[0101] The system of the present application controls the acquisition module, the optimization module, the training module and the recognition module through a controller to execute the acquisition and processing of fire scene fire images.

[0102] Exemplarily, the computer program of the present application can be divided into one or more modules, one or more modules are stored in the memory and executed by the controller to complete the present application. One or more modules can be a series of computer program instruction segments that can complete specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device. For example, the computer program can be divided into an acquisition module, an optimization module, a training module, and an identification module, and the specific functions of each module are as described above. The aforementioned controller can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0103] The above shows and describes the basic principles, main features and advantages of the present application. For those skilled in the art, it is obvious that the present application is not limited to the details of the above-mentioned exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or basic features of the present application.

[0104] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present application is limited by the appended claims rather than the above description, so it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present application. Any reference sign in the claims should not be regarded as limiting the claims involved.

[0105] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A fire image recognition method, the method being executed by a controller, characterized in that: include: Acquire fire scene target images, perform preprocessing on the fire scene target images, and generate fire scene image data sets; Optimize the semantic segmentation network model U-Net; Based on the fire scene image dataset and the cross pseudo-supervision method, the optimized U-Net network model is trained; Based on the optimized U-Net network model, fire scene image recognition is performed on the fire scene.

2. The method according to claim 1, characterized in that The fire scene target image includes at least a scene image and a fire image, wherein the scene image includes at least a weather condition image and a lighting condition image, and the fire image includes at least a fire scene image, a video frame, and an image generated by a simulated fire.

3. The method according to claim 1, characterized in that The preprocessing of the fire scene image specifically includes at least performing the following operations on the fire scene image: normalization processing and data enhancement.

4. The method according to claim 3, characterized in that The normalization process includes: unifying the colors of the fire scene images, unifying the pixel values ​​of the fire scene images into the same range, and adjusting the fire scene images to the same resolution. The adjustment of the fire scene images to the same resolution is performed using bilinear difference.

5. The method according to claim 3, characterized in that: The data enhancement includes: performing geometric changes on the fire scene image, and the geometric changes include at least rotation, flipping, translation, and scaling.

6. The method according to claim 1, characterized in that The optimization of the semantic segmentation network model U-Net specifically includes: Based on the U-Net network model structure and combined with residual connections, the ResU-Net model is generated; Based on the U-Net network model structure, combined with Transformer, the SwinU-Net model is generated; The ResU-Net model and SwinU-Net model are optimized U-Net network models.

7. The method according to claim 1, characterized in that The training of the optimized U-Net network model based on the fire scene image dataset and the cross pseudo-supervision method at least includes: Based on the pseudo labels generated by the ResU-Net model, the SwinU-Net model is trained; Based on the pseudo labels generated by the SwinU-Net model, the ResU-Net model is trained.

8. The method according to claim 7, characterized in that The pseudo labels generated by ResU-Net are used to train the SwinU-Net model, specifically: The ResU-Net model is based on the pseudo-labeled dataset generated by the SwinU-Net model Perform training on unlabeled data and update its parameters; where D b is the pseudo-label dataset of the SwinU-Net model, x i is the i-th labeled image, It is the pseudo label generated by SwinU-Net model training; The SwinU-Net model is based on the pseudo-labeled dataset generated by the ResU-Net model Perform training on unlabeled data and update its parameters; where D a is the pseudo-label dataset of the ResU-Net model, x i is the i-th labeled image, are pseudo labels generated by ResU-Net model training.

9. The method according to claim 1, characterized in that: The optimized U-Net network model is used to perform fire scene image recognition on the fire scene, specifically including: The future fire scene image is input into the optimized U-Net network model, and the encoder part of the optimized U-Net network model is used to perform feature extraction on the fire scene image; The optimized U-Net network model decoder performs feature fusion and upsampling operations. The ResU-Net model gradually fuses and upsamples multi-level features through the decoder to restore the spatial resolution and generate predictive recognition results. The SwinU-Net model combines multi-scale global feature information through the decoder to generate predictive recognition results.

10. A fire image recognition system, characterized in that: include: An acquisition module, the acquisition module is used to acquire a fire scene target image, perform preprocessing on the fire scene target image, and generate a fire scene image data set; An optimization module, wherein the optimization module is used to optimize the semantic segmentation network model U-Net; A training module, wherein the training module is used to perform training on the optimized U-Net network model based on a fire scene image dataset and a cross pseudo-supervision method; The recognition module is used to perform fire scene image recognition on the fire scene based on the optimized U-Net network model.

Images