Multi-scene fire detection method based on deep learning

Abstract

The invention discloses a multi-scene fire detection method based on deep learning. The method comprises the following steps: preparing a multi-scene fire data set, and dividing pictures into a training set, a verification set and a test set in proportion; designing a fire detection network model, wherein the fire detection network model is composed of an input image, a feature extraction module, a residual channel attention module, an Inception feature extraction module, a maximum pooling layer, an average pooling layer, a random inactivation layer and a full connection layer; during model training, a cross entropy loss function is adopted to measure differences, an Adam optimizer is adopted to update parameters, a training set and a verification set are utilized to carry out iterative training, and the optimal weight is selected according to the convergence condition of the loss function and the accuracy is evaluated. The trained model and the weight are deployed to computing power equipment through conversion and quantification, after a camera is connected, the computing power equipment reads an image, preprocesses the image and transmits the image into the model for reasoning, a preset threshold is compared with the reasoning result accuracy, and an alarm is given if the accuracy is greater than the threshold. The network model is suitable for multiple scenes, small in parameter quantity and high in reasoning speed, the small flame detection capability is effectively enhanced, the problems of insufficient small flame detection, large parameter quantity, slow reasoning and the like in the initial stage of a multi-scene fire of an existing model are solved, good practicability and generalization are achieved, and an efficient and reliable solution is provided for fire detection.

Description

Technical Field

[0001] The present invention relates to the field of fire detection technology, and in particular to a multi-scenario fire detection method based on deep learning. Background Art

[0002] Fire detection is an important part of fire safety. Timely detection of fire in the early stage of fire is of great significance to fire control. Fire occurs in a variety of scenes, including forest daytime fires, indoor fires in houses, outdoor fires in houses, forest nighttime fires, car fires, ship fires, etc. With the development of artificial intelligence technology, neural network models based on deep learning have achieved outstanding performance in various applications, which greatly improves the accuracy of detection. In the fire detection model of deep learning neural network based on real-time camera images, the effective flame features of the input image are the key to detection. The existing network model lacks application in multi-scene and small flame fire detection, and the model has a large number of parameters and slow inference speed. Therefore, a multi-scene fire detection method based on deep learning is needed to solve the above problems. Summary of the invention

[0003] The content of the present invention is to provide a multi-scenario fire detection method based on deep learning, which collects detection scene images through a camera, and passes the images to a deep learning neural network model for reasoning, and determines whether a fire occurs based on the reasoning results. The proposed neural network model introduces a feature extraction module, a residual channel attention module, and an Inception module to perform several downsampling and feature extraction on the image to be detected in sequence, and obtains the first, second, third, and fourth eigenvectors respectively. The generalization ability of the method is enhanced by adding a random deactivation layer to the output of the fourth eigenvector, and the fire information in the image to be detected is obtained by feature analysis of the eigenvector after random deactivation. This network model has a small number of parameters and a fast reasoning speed. It can detect fires in a variety of scenarios, and it also has a good effect on the detection of small flames.

[0004] A multi-scenario fire detection method based on deep learning according to the present invention comprises the following steps:

[0005] 1. Preparation of multi-scenario fire dataset;

[0006] 2. Fire detection model design;

[0007] 3. Fire detection model weight parameter training;

[0008] 4. Fire detection model deployment and fire detection.

[0009] Preferably, in step one, specifically:

[0010] Create a multi-scenario fire dataset for network model training. Randomly divide the images into three datasets at a ratio of 53%, 27%, and 20%, named training set, validation set, and test set. The training set and validation set are used to train the network model parameters, and the test set is used to evaluate the detection quality of the network model.

[0011] Preferably, in step 2, specifically:

[0012] A multi-scenario fire detection network model with high accuracy is proposed. The network model consists of an input image, a feature extraction module, a residual channel attention module, an Inception feature extraction module, a maximum pooling layer, an average pooling layer, a random dropout layer, and a fully connected layer. In the network model, the input image is a 3-channel image data tensor with a width and height of 224×224. In the feature extraction module, the input tensor features are expanded and converted to 192×56×56. The tensor is downsampled to 192×28×28 through the maximum pooling module. The residual channel attention block is used to enhance the performance of fire detection. The Inception feature extraction-A module is used to extract the features of the fire, and the one-dimensional data size of the tensor is expanded to 480. The tensor is downsampled to 480×14×14 through the maximum pooling module. The Inception feature extraction-B module is used to extract the features of the fire, and the one-dimensional data size of the tensor is expanded to 896. The tensor is downsampled to 896×7×7 through the maximum pooling module. The Inception feature extraction-C module is used to extract the features of fire and expand the one-dimensional data size of the tensor to 1024. The tensor is downsampled to 1024×1×1 through the average pooling module. The random dropout layer is used to improve the generalization ability of the model. Finally, the fully connected layer is used to classify the network model results to obtain the probability of fire and non-fire.

[0013] The feature extraction module used in the present invention has the function of effectively enhancing the features of the small flame on the boundary of the convolution operation. Specifically, the input feature map is set to be a matrix I, whose size is 224×224, I(I,j) represents the pixel point in the i-th row and j-th column, and each pixel point is represented by three RGB values. t1 is the matrix after column transformation, I t2 is the matrix after row transformation. The rolling feature map is the input feature map matrix I after I t1 Transformation and I t2 The feature map obtained after transformation. t1 The transformation is calculated as:

[0014]

[0015] I t2 The transformation is calculated as:

[0016]

[0017] Through the above transformation operation, this method can pass more features of the input image to the network model. During the training of the network model, more ideal model weight parameters can be obtained, thereby obtaining better test results. The specific implementation method is: a scrolling operation is performed on the feature map of the input image, that is, the pixel data of the bottom two rows are scrolled to the top position, and then the pixel data of the right two columns are scrolled to the left position, thereby completing the creation of the scrolling feature map. The feature extraction module uses a 4×4 convolution kernel with a stride of 4 to convolve the input image feature map and the scrolling feature map, and then splices the convolution results of the input feature map and the scrolling feature map according to the first dimension to finally obtain the output feature map.

[0018] The residual channel attention block used in the present invention can focus on the relevant features of the fire in the network model and enhance the representation of the fire features in the network model. The residual channel attention block is specifically composed of a 3×3 convolution kernel, a GELU activation function, a 3×3 convolution kernel, a channel attention module (CA module) and a residual connection. Among them, the channel attention module is composed of a global average pooling layer GloAvgPool, a 1×1 convolution kernel, a GELU activation function, a 1×1 convolution kernel, and a Sigmoid activation function. The input data X of the channel attention module is a tensor of C×H×W. First, the input data is global pooled in the spatial dimension (F GAP ) operation compresses X and compresses the data dimension to C×1×1 data. GAP The expression is:

[0019]

[0020] Each channel consists of F GAP Global pooling calculates z c , where H and W are the height and width of the input tensor, x c (i,j) is the pixel value at the i-th row and j-th column.

[0021] Use the linear function F D Reduce the channel size of the input data C×1×1 to The dimension is Tensor and use the linear function F U Bundle The tensor is raised to C×1×1. This process can input the channel attention information of the tensor, that is, the channel factor S C , its calculation expression is:

[0022] S C =F U (F D (z c ))

[0023] Finally, by X and S C After the dot product, the output data X' is obtained. The expression of X' is:

[0024] X′=S C ·X

[0025] The Inception feature extraction used in this invention consists of 4 branches, which can obtain feature information in the channel direction. These branches include a 1×1 convolution branch, two 1×1 convolution-3×3 convolution branches, and a 3×3 maximum pooling-1×1 convolution branch. The input data is concatenated through the convolution results of the 4 branches as the output data.

[0026] Preferably, in step three, specifically:

[0027] The model training of the present invention uses the cross entropy loss function as the loss function to measure the difference between the model prediction result and the true label. By using the Adam optimizer to update the parameters of the neural network according to the gradient information of the loss function, the model converges to a better parameter value more quickly. The expression of the cross entropy loss function is:

[0028]

[0029] Among them, y represents the true label, y∈{0,1}, Represents the model prediction probability.

[0030] In actual training, the training set and validation set are fully utilized to perform iterative training operations on the network model. During the training process, the numerical changes of the loss function are continuously monitored until the loss function converges to 0, and the training of the network model is terminated. After the training is completed, the best network weights are selected from the entire training process for subsequent testing.

[0031] Furthermore, the accuracy (Acc) of the selected network model is evaluated to determine whether it can meet the actual application requirements. The accuracy is calculated using a specific expression, which is:

[0032]

[0033] Among them, TP is the number of positive samples correctly predicted as positive, TN is the number of negative samples correctly predicted as negative, FP is the number of negative samples incorrectly predicted as positive, and FN is the number of positive samples incorrectly predicted as negative.

[0034] Preferably, in step 4, specifically:

[0035] The fire detection model designed in step 2 can obtain model weight parameters with high accuracy after training and testing in step 3. The fire detection model designed in step 2 and the weight parameters trained in step 3 are converted and quantized using the conversion tools and quantization tools provided by the computing power device, and the converted and quantized model is deployed to the computing power device. Use a camera to connect to the computing power device where the fire detection model is deployed. The computing power device reads the image from the camera and passes the image to the fire detection model for reasoning after preprocessing. By comparing the preset fire accuracy threshold with the accuracy of the fire model reasoning result, it is determined whether the accuracy of the current reasoning result is greater than the preset fire accuracy threshold. If it is greater than the threshold, an alarm operation is performed.

[0036] Beneficial effects of the present invention: First, by using the feature extraction module, the residual channel attention module and the Inception module to perform multiple downsampling and feature extraction on the image to be detected, and carrying out feature analysis on the feature vector after random inactivation, it is possible to effectively obtain the fire information in the image to be detected, thereby accurately judging whether the fire has occurred. Secondly, the constructed network model has a small amount of parameters and a fast reasoning speed, and can be widely used in fire detection in various different scenes such as forests, indoor and outdoor houses, cars, and ships, greatly improving the practicality and versatility of the detection method, and fully meeting the fire detection needs of diversified scenes. Furthermore, the defects of the large amount of parameters and slow reasoning speed of the existing network model are overcome, and on the basis of ensuring the accuracy of detection, the detection reasoning process is significantly accelerated, the fire situation can be detected in time, and the response efficiency is effectively improved. In addition, it performs well in small flame detection, effectively solves the problem of lack of small flame feature extraction ability in the early stage of multi-scene fire, and is conducive to timely discovering hidden dangers when the small flame stage just appears in the fire, and wins more precious time for fire prevention and control. At the same time, by adding a random inactivation layer to the output of the fourth feature vector, the generalization ability of the method is effectively enhanced, so that it can play a role in fire detection in various different scenes. Finally, a specific cross entropy loss function and Adam optimizer are used to carry out model training, which enables the model to quickly converge to better parameter values. With the help of reasonable data set division and accuracy evaluation methods, the detection quality of the network model can be effectively evaluated, effectively ensuring that it can meet the needs of actual applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the method of the present invention.

[0038] Figure 2 This is a diagram of the multi-scenario fire detection network structure of the method proposed in the present invention.

[0039] Figure 3 Schematic diagram of the feature extraction module of the method proposed in the present invention.

[0040] Figure 4Schematic diagram of the residual channel attention module of the method proposed in the present invention.

[0041] Figure 5 Schematic diagram of the channel attention mechanism of the method proposed in the present invention.

[0042] Figure 6 Schematic diagram of the Inception feature extraction module of the method proposed in the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0044] This embodiment provides a multi-scenario fire detection method based on deep learning, such as Figure 1 As shown, the following steps are included:

[0045] 1. Prepare a multi-scenario fire dataset: Create a multi-scenario fire dataset for training the network model;

[0046] Specifically, the data set includes 15,000 fire images and 15,000 non-fire images. The width and height of the image input to the fire detection neural network model designed by the present invention are 224×224, so the image size in the data set needs to be processed to 224×224. The image processing steps are as follows. First, all images are converted to 3-channel RGB mode. Secondly, the image size is reshaped to N×224 or 224×N, where N is greater than 224 or equal to 224. Then, the size of the data set image is cut to 224×224, and the images containing fire features in the data set are manually checked to ensure that the images input to the neural network model have the characteristics of fire. The 15,000 fire images and 15,000 non-fire images in the data set are randomly divided into three data sets according to the ratios of 8000 / 15000, 4000 / 15000, and 3000 / 15000, respectively named as training set, validation set, and test set. The training set and validation set are used for training the network model parameters, and the test set is used to evaluate the detection quality of the network model.

[0047] 2. Fire detection model design, a high-precision multi-scenario fire detection network model is constructed. The specific composition of the network model is as follows:

[0048] like Figure 2As shown in the figure, it includes an input image, a feature extraction module, a residual channel attention module, an Inception feature extraction module, a maximum pooling layer, an average pooling layer, a random dropout layer, and a fully connected layer. The input image is presented in the form of a 3-channel image data tensor with a width and height of 224×224. In the feature extraction module, the operation process is as follows: the features of the input tensor are extended and converted into a data form of 192×56×56. Subsequently, the tensor is downsampled with the help of the maximum pooling module to be converted into a tensor of 192×28×28. On this basis, the residual channel attention block is introduced to enhance the performance of fire detection.

[0049] Next, the Inception feature extraction-A module is used to extract the features of the fire. During this process, the one-dimensional data size of the tensor is expanded to 480. After that, the tensor is downsampled again through the maximum pooling module to become a 480×14×14 tensor. Then, the Inception feature extraction-B module is used to further extract the fire features, thereby expanding the one-dimensional data size of the tensor to 896. Subsequently, the maximum pooling module is used to downsample the tensor to become a 896×7×7 tensor. The fire features are then extracted through the Inception feature extraction-C module, and the one-dimensional data size of the tensor is expanded to 1024. After that, the tensor is downsampled with the help of the average pooling module to obtain a 1024×1×1 tensor. Finally, the random dropout layer is used to improve the generalization ability of the model. Finally, the fully connected layer is used to classify the results of the network model to obtain the probability results of fire and non-fire.

[0050] Detailed description of feature extraction module: The feature extraction module has the function of effectively enhancing the features of the small flame on the boundary of the convolution operation. Specifically, let the input feature map be a matrix I with a size of 224×224, where I(i,j) represents the pixel point in the i-th row and j-th column, and each pixel point is represented by three RGB values. t1 is the matrix after column transformation, I t2 is the matrix after row transformation. The so-called rolling feature map is the input feature map matrix I undergoes I t1 Transformation and I t2 The feature map obtained after transformation. t1 The transformation is calculated as:

[0051]

[0052] I t2 The transformation is calculated as:

[0053]

[0054] Through the above transformation operation, this method can transfer more features of the input image to the network model, and in the network model training device, it can obtain more ideal model weight parameters, thereby obtaining better test results. Figure 3 As shown in the figure, the specific implementation method is: perform a scrolling operation on the feature map of the input image, that is, scroll the pixel data of the two bottom rows to the top position, and then scroll the pixel data of the two right columns to the left position, so as to complete the creation of the scrolling feature map. The feature extraction module uses a 4×4 convolution kernel with a stride of 4 to convolve the input image feature map and the scrolling feature map, and then concatenates the convolution results of the input feature map and the scrolling feature map according to the first dimension to finally obtain the output feature map.

[0055] Detailed description of the residual channel attention block: The residual channel attention block used in the present invention plays a key role in the entire network model system. It can accurately focus on the relevant features of the fire, thereby significantly enhancing the network model's ability to express and present fire features. Figure 4 As shown in (a), the residual channel attention block is specifically composed of 3×3 convolution kernel, GELU activation function, 3×3 convolution kernel, channel attention module (CA module) and residual connection. Figure 4 As shown in (b), the internal structure includes the global average pooling layer GloAvgPool, 1×1 convolution kernel, GELU activation function, 1×1 convolution kernel and Sigmoid activation function. The schematic diagram of the channel attention module is shown in Figure 5 Specifically, when the channel attention module receives input data X (in the form of a tensor of C×H×W), it first performs global pooling (F GAP ) operation to perform data compression on X, compressing the original C×H×W data dimension into C×1×1 data form. GAP The specific expression of the operation is:

[0056]

[0057] In this expression, H and W represent the height and width of the input tensor, respectively, and x c (i,j) represents the pixel value of the i-th row and j-th column. Through this global pooling operation, each channel can be pooled according to F GAP Calculate the corresponding value.

[0058] Then, using the linear function F D Reduce the channel size of the input data (C×1×1 with C channels) to So the dimension is Next, use the linear function F U The The tensor is again increased to C×1×1. In this process, the channel attention information of the input tensor can be effectively extracted, that is, the channel factor S C , its calculation expression is:

[0059] S C =F U (F D (z c ))

[0060] Finally, the original input data X is combined with the channel factor S C Perform a dot multiplication operation to obtain the output data X', which is expressed as:

[0061] X′=S C ·X

[0062] Through the above series of operations, the residual channel attention block can effectively enhance the network model's ability to capture and characterize fire characteristics, and improve the performance and accuracy of the entire fire detection network model.

[0063] Inception feature extraction module detailed description: Figure 6 As shown, the Inception feature extraction module used in the present invention presents a specific structural composition. It is mainly constructed by 4 different branches, which each play a unique role in the entire feature extraction process and cooperate to achieve effective acquisition of channel direction feature information. One of the branches is a 1×1 convolution branch, which can perform a preliminary 1×1 convolution operation on the input data to extract specific local feature information. The other two branches are combined in the form of 1×1 convolution and 3×3 convolution, that is, the input data is first dimensionalized or preliminarily screened by 1×1 convolution, and then 3×3 convolution is used to further explore the feature details in the data, thereby enriching the dimension and depth of feature extraction from different angles. There is also a branch of 3×3 maximum pooling-1×1 convolution branch, which first uses 3×3 maximum pooling operation to pool the input data to reduce the spatial resolution of the data, reduce the amount of calculation and retain the main feature information, and then further extracts and integrates the features of the pooling results through 1×1 convolution.

[0064] In the actual operation process, the input data will flow into these four branches respectively, and the corresponding convolution or pooling and convolution operations will be completed in each branch. Finally, the convolution results generated by these four branches are connected and combined according to specific series rules to form the final output data. Through this multi-branch structure and series output design, the Inception feature extraction module can make full use of the advantages of different convolution operations, comprehensively and accurately obtain rich feature information in the channel direction, and provide more accurate and comprehensive feature data support for the entire fire detection network model, thereby effectively improving the model's ability to recognize and judge fire characteristics.

[0065] 3. Fire detection model weight parameter training. The model training phase adopts specific strategies and methods. The specific operations are as follows:

[0066] The cross entropy loss function is selected as the key indicator to measure the difference between the model prediction result and the true label. During the training process, the parameters of the neural network are updated with the help of the Adam optimizer based on the gradient information generated by the loss function. In this way, the model can converge to a relatively optimal parameter value state at a more efficient rate. The mathematical expression of the cross entropy loss function is:

[0067]

[0068] Among them, y represents the true label, and its value range is limited to the set {0,1}. It represents the model prediction probability.

[0069] In actual training, the training set and validation set are fully utilized to perform iterative training operations on the network model. During the training process, the numerical changes of the loss function are continuously monitored until the loss function converges and approaches 0, and then the training process of the network model is terminated. After the training is completed, the best network weights are selected from the entire training process for subsequent testing.

[0070] Furthermore, the test set is used to evaluate the accuracy (Acc) of the selected network model to determine whether it can meet the application requirements. The accuracy is calculated using a specific expression, which is:

[0071]

[0072] Among them, TP is the number of positive samples correctly predicted as positive, TN is the number of negative samples correctly predicted as negative, FP is the number of negative samples incorrectly predicted as positive, and FN is the number of positive samples incorrectly predicted as negative. By calculating this indicator, the performance of the network model in the fire detection task can be accurately evaluated, providing a strong basis for the optimization and improvement of the model, while also ensuring the reliability and effectiveness of the model in actual application scenarios.

[0073] 4. Fire detection model deployment and fire detection.

[0074] First, after the fire detection model designed in step 2 has been trained and tested in step 3, it can obtain model weight parameters with high accuracy. Then, the fire detection model designed in step 2 and the weight parameters trained in step 3 are converted and quantized with the help of the conversion tools and quantization tools provided by the computing power equipment. After the conversion and quantization are completed, the processed model is deployed on the computing power equipment to build a complete fire detection operating environment.

[0075] Then, the camera is connected to the computing device where the fire detection model has been deployed. When the connection is successful, the computing device can read real-time image data from the camera. After obtaining the image data, the computing device will perform a series of preprocessing operations on the image to ensure that the format and quality of the image data meet the input requirements of the fire detection model. After the preprocessing is completed, the image data is passed to the fire detection model to start the model's reasoning process.

[0076] After the inference process is completed, the system will automatically compare the preset fire accuracy threshold with the accuracy of the fire model inference result. Through this comparison operation, it can accurately determine whether the accuracy of the current inference result is higher than the preset fire accuracy threshold. Once it is determined that the accuracy of the current inference result is greater than the preset fire accuracy threshold, the system will immediately execute the alarm operation to promptly notify relevant personnel or systems of the fire risk, thereby buying valuable time for fire prevention and control and improving the efficiency and effectiveness of overall fire detection and response.

[0077] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or transformation made by a person skilled in the art based on the invention is within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A multi-scenario fire detection method based on deep learning, characterized in that: The following steps are involved:

1. Preparation of multi-scenario fire dataset; 2. Fire detection model design; 3. Fire detection model weight parameter training; 4. Fire detection model deployment and fire detection.

2. The multi-scenario fire detection method based on deep learning according to claim 1, characterized in that: In step one, specifically: Create a multi-scene fire dataset for network model training. Randomly divide the images into three datasets with a ratio of 53%, 27%, and 20%, and name them training set, validation set, and test set respectively.

3. The multi-scenario fire detection method based on deep learning according to claim 1, characterized in that: In step 2, specifically: The network model consists of an input image, a feature extraction module, a residual channel attention module, an Inception feature extraction module, a maximum pooling layer, an average pooling layer, a random dropout layer, and a fully connected layer, where: The input image is a 3-channel image data tensor with a width and height of 224×224; The feature extraction module is implemented in the following way: suppose the input feature map is a matrix I, whose size is 224×224, I(i, j) represents the pixel point in the i-th row and j-th column, and each pixel point is represented by three RGB values. t1 is the matrix after column transformation, I t2 is the matrix after row transformation. The rolling feature map is the input feature map matrix I after I t1 Transformation and I t2 The feature map obtained after transformation, I t1 The transformation is calculated as: I t2 The transformation is calculated as: The feature map of the input image is scrolled, the pixel data of the bottom two rows are scrolled to the top, and then the pixel data of the right two columns are scrolled to the left to obtain the scrolling feature map. The input image feature map and the scrolling feature map are convolved with a 4×4 convolution kernel with a stride of 4. Then, the convolution results of the input feature map and the scrolling feature map are concatenated along the first dimension to obtain the output feature map, so as to effectively enhance the features of the small flames at the boundary of the convolution operation and pass more features of the input image to the network model. The residual channel attention module consists of a 3×3 convolution kernel, a GELU activation function, a 3×3 convolution kernel, a channel attention module and a residual connection. The channel attention module consists of a global average pooling layer, a 1×1 convolution kernel, a GELU activation function, a 1×1 convolution kernel, and a Sigmoid activation function. The operation is as follows: The input data X of the channel attention module is a tensor of C×H×W. First, the input data is global pooled in the spatial dimension (F GAP ) operation compresses X and compresses the data dimension to C×1×1 data. GAP The expression is: Each channel consists of F GAP Global pooling calculates z c , where H and W are the height and width of the input tensor, x c (i, j) is the pixel value of the i-th row and j-th column, using the linear function F D Reduce the channel size of input data C×1×1 to The dimension is Tensor and use the linear function F U Bundle The tensor is increased to C×1×1. In this process, the channel attention information of the input tensor can be effectively extracted, that is, the channel factor S C , its calculation expression is: S C =F U (F D (Z c )) Finally, the original input data X is combined with the channel factor S C Perform a dot multiplication operation to obtain the output data X', which is expressed as: X′=S C ·X The Inception feature extraction module consists of four branches, including a 1×1 convolution branch, two 1×1 convolution-3×3 convolution branches and a 3×3 maximum pooling-1×1 convolution branch. The input data is connected in series through the convolution results of the four branches as output data to obtain the feature information of the fire scene.

4. The multi-scenario fire detection method based on deep learning according to claim 1, characterized in that: In step three, specifically: The cross entropy loss function is used as the loss function to measure the difference between the model prediction result and the true label. The expression of the cross entropy loss function is: Among them, y represents the true label, y∈{0,1}, Represents the model prediction probability. The parameters of the neural network are updated according to the gradient information of the loss function by using the Adam optimizer, so that the model converges to a better parameter value more quickly. The network model is iteratively trained using the training set and the validation set until the loss function converges to close to 0, the training of the network model is terminated, and the best network weights in the training are selected for testing.

5. The multi-scenario fire detection method based on deep learning according to claim 1, characterized in that: In step 4, specifically: The fire detection model designed in step 2 can obtain model weight parameters with high accuracy after training and testing in step 3. Deploy the fire detection model designed in step 2 and the weight parameters trained in step 3 to the computing device. Use a camera to connect to the computing device where the fire detection model and fire detection weight parameters are deployed. The computing device reads the image from the camera and passes the image to the fire detection model for reasoning after preprocessing. The accuracy of the fire model reasoning result is compared with the preset fire accuracy threshold. When the accuracy of the reasoning result is greater than the preset fire accuracy threshold, an alarm operation is performed.

Images