Fire intensity quantification method fusing instance segmentation and optimization convolution self-coding, storage medium, equipment and computer program product

Abstract

The invention discloses a fire intensity quantification method fusing instance segmentation and optimization convolution self-coding, a storage medium, equipment and a computer program product, and the method comprises the steps: collecting flame images in different combustion states, and unifying the resolution of the flame images; constructing a fire intensity quantification model, wherein the fire intensity quantification model comprises a YOLO11seg model, a convolutional self-encoding network and a linear regression model which are connected in sequence; sequentially inputting the flame images with the uniform resolution into the fire intensity quantification model for training until the total loss function of the fire intensity quantification model is converged, and completing the training of the fire intensity quantification model; the flame image is collected again, the resolution of the collected flame image is adjusted, the flame image is input into the trained fire intensity quantification model, and the fire intensity is predicted. According to the method, instance segmentation of the flame image is carried out through the YOLO11seg model, the deep features of the flame region are extracted through the convolutional self-encoding network, and the fire intensity is predicted through the linear regression model.

Description

Technical Field

[0001] The present invention belongs to the technical field of fire emergency response and intelligent monitoring. Specifically, it relates to a method for quantifying fire intensity that combines instance segmentation and optimized convolutional autoencoder, a storage medium, a device, and a computer program product. Background Art

[0002] As one of the most common emergency situations, fire not only poses a direct threat to public safety but also has a profound impact on the stable development of social economy. Therefore, exploring and establishing more efficient and intelligent fire monitoring and early warning methods has become one of the core topics in the field of modern fire safety. Traditional fire monitoring means mainly rely on devices such as temperature sensors and smoke detectors. Although these technologies can achieve preliminary early warning, they have significant limitations in accurately reflecting the scale and development trend of fires, and there is also a false alarm rate. In addition, in complex environments or large spaces, the performance of traditional devices is often limited by the installation location and coverage range, making it difficult to meet the requirements of efficient monitoring.

[0003] With the rapid progress of computer vision technology, image-based fire monitoring technology has emerged and attracted extensive attention and research. However, existing methods mainly rely on object detection technology, and their dependence on a large number of labeled images makes them vulnerable to interference in complex background environments, affecting the accuracy of recognition. In addition, few studies combine flame images to quantitatively analyze the fire intensity, and this lack has, to a certain extent, hindered the formulation of scientific rescue decisions, and there is an urgent need for technological breakthroughs and improvements. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the present invention provides a method for quantifying fire intensity that combines instance segmentation and optimized convolutional autoencoder, a storage medium, a device, and a computer program product, and realizes high-precision prediction of fire intensity through the constructed fire intensity quantification model.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solutions: A method for quantifying fire intensity that combines instance segmentation and optimized convolutional autoencoder, specifically including the following steps:

[0006] Step S1: Collect flame images under different combustion states and unify the resolution of the flame images.

[0007] Step S2: Construct a fire intensity quantification model, including a YOLO11 seg model, a convolutional autoencoder network, and a linear regression model connected in sequence; the YOLO11 seg model is used for instance segmentation of the flame image and retains the flame area in the flame image; the convolutional autoencoder network is used to extract deep features of the flame area; the linear regression model is used to predict the fire intensity according to the deep features of the extracted flame area.

[0008] Step S3: Input the flame images of the unified resolution size into the fire intensity quantization model for training in sequence until the total loss function of the fire intensity quantization model converges, and complete the training of the fire intensity quantization model;

[0009] Step S4: Re - collect flame images, adjust the resolution size of the collected flame images, and input them into the trained fire intensity quantization model to predict the fire intensity.

[0010] Further, the network structure of the YOLO11 seg model in step S2 specifically includes: a backbone network for extracting flame image features, a neck network for integrating flame image features from different levels of the backbone network, and a head network for instance segmentation; the backbone network includes a first convolutional layer, a second convolutional layer, a first C3K2 module, a first receptive field attention convolutional module, a second C3K2 module, a second receptive field attention convolutional module, a third C3K2 module, a third receptive field attention convolutional module, a fourth C3K2 module, an SPPF module, and a C2PSA module connected in sequence; the neck network includes a first up - sampling module, a first Concat module, a fifth C3K2 module, a second up - sampling module, a second Concat module, a sixth C3K2 module, a fourth receptive field attention convolutional module, a third Concat module, a seventh C3K2 module, a fourth Concat module, and an eighth C3K2 module connected in sequence. The output end of the second C3K2 module is also connected to the second Concat module, the output end of the third C3K2 module is also connected to the first Concat module, the output end of the fifth C3K2 module is also connected to the third Concat module, and the output end of the second Concat module is also connected to the sixth C3K2 module; the head network includes a segmentation prototype module, a first image segmentation module, a second image segmentation module, and a third image segmentation module. The output end of the sixth C3K2 module is respectively connected to the input end of the segmentation prototype module and the input end of the first image segmentation module. The input end of the second image segmentation module is connected to the output end of the seventh C3K2 module, and the input end of the third image segmentation module is connected to the output end of the eighth C3K2 module.

[0011] Further, the first receptive field attention convolutional module, the second receptive field attention convolutional module, the third receptive field attention convolutional module, the fourth receptive field attention convolutional module, and the fifth receptive field attention convolutional module all use the average pooling function to aggregate the global information of each receptive field feature, then perform the global information interaction of all receptive field features through 1×1 group convolution operations, and finally use the Softmax function to weight each receptive field feature:

[0012] F = Softmax(g 1×1(AvgPool(X)))×ReLU(Norm(g k×k (X)))

[0013] where g k×k represents grouped convolution with a convolution kernel size of k×k, X represents the feature map input to a certain receptive field attention convolution module, Norm() represents the normalization function, ReLU() represents the activation function, AvgPool() represents the average pooling function, and F represents the receptive field features after weight weighting.

[0014] Furthermore, the convolutional autoencoder network consists of an encoder and a decoder. The encoder is composed of multiple autoencoders, and each sub-encoder consists of a third convolutional layer, a first normalization layer, a first SiLU activation function, and a first pooling layer. The third convolutional layer is used to extract features from the input image to generate a high-dimensional feature map; the first normalization layer is used to ensure the consistency of the distribution of the high-dimensional feature maps input to each layer; the first SiLU activation function is used to perform a non-linear transformation on the feature map output by the first normalization layer; the first pooling layer performs a downsampling operation on the non-linearly transformed feature map. The decoder is composed of multiple sub-decoders, and each sub-decoder consists of an upsampling layer, a fourth convolutional layer, a second normalization layer, and a second SiLU activation function. The upsampling layer is used to expand the spatial dimension of the feature map; the fourth convolutional layer is used to reconstruct the local features of the feature map, enhance the detail information, and adjust the number of channels; the second normalization layer ensures the consistency of the distribution of the reconstructed feature maps input to each layer; the second SiLU activation function is used to perform a non-linear transformation on the feature map output by the second normalization layer.

[0015] Furthermore, step S3 includes the following sub-steps:

[0016] Step S3.1: Annotate the segmentation mask in the flame region of each flame image as the training label of the YOLO11 seg model;

[0017] Step S3.2: Input a flame image into the YOLO11 seg model for instance segmentation to generate a segmentation prototype and a segmentation mask, and use the segmentation mask to set the pixel values of the background region of the flame image to zero to obtain a flame image with the background removed;

[0018] Step S3.3: Input the flame image with the background removed into the convolutional autoencoder network to extract the deep features of the flame region;

[0019] Step S3.4: Input the deep features of the flame region into the linear regression model for regression prediction to predict the fire intensity;

[0020] Step S3.5: Repeat steps S3.2 - S3.4 until the total loss function of the fire intensity quantization model converges, and complete the training of the fire intensity quantization model.

[0021] Further, in step S3.2, the specific process of setting the pixel values of the background area of the flame image to zero using the segmentation mask to obtain the flame image after removing the background is as follows: Construct a full-zero pixel image with the same size as the original flame image, traverse the segmentation mask. If the size of the segmentation mask is inconsistent with the size of the original flame image, use the nearest neighbor interpolation method to adjust the size of the segmentation mask to be the same as the size of the original flame image; otherwise, use the segmentation mask to filter out the pixel values of the flame background area in the original flame image, replace them with full-zero pixel values, and save the flame image after removing the background.

[0022] Further, the total loss function L of the fire intensity quantization model is specifically:

[0023] L = L 1 + L 2 + L 3

[0024] where L 1 represents the loss function of the YOLO11 seg model, L 2 represents the loss function of the convolutional autoencoder network, L 3 represents the loss function of the linear regression model, N represents the number of flame images used for training, i represents the index of N, area i represents the normalized area of the flame region in the i-th flame image, B i represents the flame region in the i-th flame image, BCEWithLogits represents the binary cross-entropy loss function, pred i (x, y) represents the segmentation mask generated by the pixel at the (x, y) position in the i-th flame image, gt i (x, y) represents the true segmentation mask of the pixel at the (x, y) position in the i-th flame image, x i represents the i-th flame image after removing the background, z i represents the deep feature of the flame region reconstructed by the convolutional autoencoder network for the i-th one, y i represents the true fire intensity of the i-th flame image, represents the predicted fire intensity of the i-th flame image.

[0025] Further, the present invention also provides a computer-readable storage medium storing a computer program, and the computer program enables a computer to execute the fire intensity quantization method that fuses instance segmentation and optimized convolutional autoencoder.

[0026] Furthermore, the present invention also provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method for quantifying the fire intensity by fusing instance segmentation and optimizing convolutional auto-encoding is implemented.

[0027] Furthermore, the present invention also provides a computer program product, including a computer program which, when executed by a processor, implements the method for quantifying the fire intensity by fusing instance segmentation and optimizing convolutional auto-encoding.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] (1) By introducing a receptive field attention convolutional module to improve the YOLO11 seg model, the present invention performs instance segmentation on flame images, optimizes the computational efficiency in the segmentation process, thereby being able to accurately remove background interference in flame images, only retaining the key flame regions, and effectively improving the accuracy of feature extraction and the quantization accuracy of the fire intensity.

[0030] (2) By optimizing the convolutional auto-encoding network with the SiLU activation function, the present invention significantly reduces the error in the training and prediction processes of the fire intensity quantization model, and improves its robust performance in complex scenarios.

[0031] In summary, the method for quantifying the fire intensity by fusing instance segmentation and optimizing convolutional auto-encoding in the present invention realizes high-precision quantitative evaluation of the flame combustion intensity, has strong adaptability in complex background and noise environments, and can provide intelligent and reliable technical support for fire emergency response and scientific rescue. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of the method for quantifying the fire intensity by fusing instance segmentation and optimizing convolutional auto-encoding in the present invention;

[0033] Figure 2 is a schematic structural diagram of the YOLO11 seg model in the present invention;

[0034] Figure 3 is a schematic structural diagram of the convolutional auto-encoding network in the present invention;

[0035] Figure 4 is a comparison schematic diagram of the predicted value and the true value of the fire intensity using the method for quantifying the fire intensity by fusing instance segmentation and optimizing convolutional auto-encoding in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] The technical solutions of the present invention will be further explained and illustrated below with reference to the accompanying drawings.

[0037] As Figure 1 is a flowchart of the method for quantifying fire intensity by integrating instance segmentation and optimized convolutional auto-encoding in the present invention. The method for quantifying fire intensity specifically includes the following steps:

[0038] Step S1: Collect flame images under different combustion states and unify the resolution of the flame images to 128×128.

[0039] Step S2: Construct a fire intensity quantification model, including a YOLO11 seg model, a convolutional auto-encoding network, and a linear regression model connected in sequence; among them, the YOLO11 seg model is used for instance segmentation of the flame image to retain the flame area in the flame image; the convolutional auto-encoding network is used to extract deep features of the flame area; the linear regression model is used to predict the fire intensity according to the deep features of the extracted flame area.

[0040] As Figure 2 , the network structure of the YOLO11 seg model in the present invention specifically includes: a backbone network for extracting features of the flame image, a neck network for integrating flame image features from different levels of the backbone network, and a head network for instance segmentation; among them, the backbone network includes a first convolutional layer, a second convolutional layer, a first C3K2 module, a first receptive field attention convolutional module, a second C3K2 module, a second receptive field attention convolutional module, a third C3K2 module, a third receptive field attention convolutional module, a fourth C3K2 module, an SPPF module, and a C2PSA module connected in sequence; the neck network includes a first upsampling module, a first Concat module, a fifth C3K2 module, a second upsampling module, a second Concat module, a sixth C3K2 module, a fourth receptive field attention convolutional module, a third Concat module, a seventh C3K2 module, a fourth Concat module, and an eighth C3K2 module connected in sequence. The output end of the second C3K2 module is also connected to the second Concat module, the output end of the third C3K2 module is also connected to the first Concat module, the output end of the fifth C3K2 module is also connected to the third Concat module, and the output end of the second Concat module is also connected to the sixth C3K2 module; the head network includes a segmentation prototype module, a first image segmentation module, a second image segmentation module, and a third image segmentation module. The output end of the sixth C3K2 module is respectively connected to the input end of the segmentation prototype module and the input end of the first image segmentation module. The input end of the second image segmentation module is connected to the output end of the seventh C3K2 module, and the input end of the third image segmentation module is connected to the output end of the eighth C3K2 module.

[0041] The first receptive field attention convolution module, the second receptive field attention convolution module, the third receptive field attention convolution module, the fourth receptive field attention convolution module, and the fifth receptive field attention convolution module all use the average pooling function to aggregate the global information of each receptive field feature, then perform the interaction of the global information of all receptive field features through 1×1 grouped convolution operations, and finally use the Softmax function to weight each receptive field feature to prevent feature loss:

[0042] F = Softmax(g 1×1 (AvgPool(X))) × ReLU(Norm(g k×k (X)))

[0043] where g k×k represents grouped convolution with a convolution kernel size of k×k, X represents the feature map input to a certain receptive field attention convolution module, Norm() represents the normalization function, ReLU() represents the activation function, AvgPool() represents the average pooling function, and F represents the receptive field feature after weight weighting.

[0044] The present invention improves the YOLO11 seg model by introducing the receptive field attention convolution module, performs instance segmentation on the flame image, optimizes the computational efficiency in the segmentation process, so as to accurately remove the background interference in the flame image, only retain the key flame regions, and effectively improve the accuracy of feature extraction and the quantization accuracy of the fire intensity.

[0045] Such as Figure 3, in the present invention, the convolutional autoencoder network consists of an encoder and a decoder. The encoder is composed of multiple autoencoders, and each sub-encoder consists of a third convolutional layer, a first normalization layer, a first SiLU activation function, and a first pooling layer. The third convolutional layer is used to extract features from the input image to generate a high-dimensional feature map; the first normalization layer is used to stabilize and accelerate the training process to ensure the consistency of the distribution of the high-dimensional feature maps input to each layer; the first SiLU activation function is used to perform a non-linear transformation on the feature map output by the first normalization layer to enhance the model's expressive ability; the first pooling layer performs a downsampling operation on the non-linearly transformed feature map to reduce the spatial dimension. The decoder is composed of multiple sub-decoders, and each sub-decoder consists of an upsampling layer, a fourth convolutional layer, a second normalization layer, and a second SiLU activation function. The upsampling layer is used to expand the spatial dimension of the feature map; the fourth convolutional layer is used to reconstruct the local features of the feature map, enhance the detail information, and adjust the number of channels; the second normalization layer ensures the consistency of the distribution of the reconstructed feature maps input to each layer; the second SiLU activation function is used to perform a non-linear transformation on the feature map output by the second normalization layer to further improve the model's expressive ability and reconstruction quality. The present invention optimizes the convolutional autoencoder network through the SiLU activation function, greatly reducing the error in the training and prediction processes of the fire intensity quantification model and improving its robust performance in complex scenarios.

[0046] Step S3: Input the flame images of the unified resolution size into the fire intensity quantification model in sequence for training until the total loss function of the fire intensity quantification model converges, completing the training of the fire intensity quantification model; specifically, it includes the following sub-steps:

[0047] Step S3.1: Annotate the segmentation mask in the flame area of each flame image through the labelme tool as the training label of the YOLO11 seg model.

[0048] Step S3.2: Input a flame image into the YOLO11 seg model for instance segmentation to generate a segmentation prototype and a segmentation mask. Use the segmentation mask to set the pixel values of the background area of the flame image to zero to obtain the flame image after removing the background. Specifically, construct a full-zero pixel image with the same size as the original flame image, traverse the segmentation mask. If the size of the segmentation mask is inconsistent with the size of the original flame image, use the nearest neighbor interpolation method to adjust the size of the segmentation mask to be consistent with the size of the original flame image; otherwise, use the segmentation mask to filter out the pixel values of the flame background area in the original flame image and replace them with full-zero pixel values, and save the flame image after removing the background for subsequent deep feature extraction and fire intensity quantification analysis.

[0049] Step S3.3: Input the flame image after removing the background into the convolutional autoencoder network to extract the deep features of the flame area.

[0050] Step S3.4: Input the deep features of the flame area into the linear regression model for regression prediction to predict the fire intensity. Specifically, it includes the following sub-steps:

[0051] Step S3.4.1: In the present invention, the linear regression model is a multiple linear regression model And initialize the weight parameters of the multiple linear regression model, where β 0 is the intercept term, and β i are the regression coefficients corresponding to each explanatory variable respectively, ∈ represents the error term, representing the random noise not explained by the multiple linear regression model; X i ={x i,1 ,x i,2 ,…,x i,n} represents the explanatory variables, and n represents the dimension of the explanatory variables;

[0052] Step S3.4.2: Input the deep features of the flame area into the multiple linear regression model for supervised training. Specifically, use the least squares method to determine the weight parameters to minimize the sum of squared residuals between the true value and the predicted value of the fire intensity, so as to effectively predict the predicted value of the fire intensity;

[0053] Step S3.4.3: Use the trained multiple linear regression model to perform regression prediction on the deep features of the flame area to be measured, and obtain a value between 0 and 10, which is used to quantitatively describe the flame combustion intensity.

[0054] Step S3.5: Repeat steps S3.2 - S3.4 until the total loss function of the fire intensity quantization model converges, and complete the training of the fire intensity quantization model.

[0055] The total loss function L of the fire intensity quantization model in the present invention is specifically:

[0056] L = L 1 + L 2 + L 3

[0057] Where L 1 represents the loss function of the YOLO11 seg model, L 2 represents the loss function of the convolutional autoencoder network, L 3 represents the loss function of the linear regression model, N represents the number of flame images used for training, i represents the index of N, area i represents the normalized area of the flame area in the i-th flame image, B idenotes the flame region in the i-th flame image, BCEWithLogits denotes the binary cross-entropy loss function, pred i (x, y) denotes the segmentation mask generated by the pixel at the position (x, y) in the i-th flame image, gt i (x, y) denotes the true segmentation mask of the pixel at the position (x, y) in the i-th flame image, x i denotes the i-th flame image after removing the background, z i denotes the deep feature of the flame region reconstructed by the convolutional autoencoder network for the i-th flame image, y i denotes the true fire intensity of the i-th flame image, denotes the predicted fire intensity of the i-th flame image.

[0058] Step S4: Reacquire the flame image, adjust the resolution of the acquired flame image, and input it into the trained fire intensity quantization model to predict the fire intensity.

[0059] Taking the actual observed value of each flame image as the abscissa and the predicted value of the fire intensity quantization model as the ordinate, a scatter plot is drawn, as Figure 4 shown: Most of the data points are closely distributed around the diagonal, indicating that the model has a high prediction accuracy for different scales of fire sizes; especially in the case of a large fire, the predicted value of the fire intensity quantization model is highly consistent with the actual observed value, showing strong prediction ability and stability; at the same time, since there are fewer labeled images in the training set and most are unlabeled images for feature extraction, the problem of large label demand in traditional image-based fire monitoring technology is solved.

[0060] In a technical solution of the present invention, there is also provided a computer-readable storage medium storing a computer program, and the computer program causes a computer to execute the fire intensity quantization method that fuses instance segmentation and optimizes convolutional autoencoders.

[0061] In a technical solution of the present invention, there is also provided an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the fire intensity quantization method that fuses instance segmentation and optimizes convolutional autoencoders is implemented.

[0062] In a technical solution of the present invention, there is also provided a computer program product including a computer program, and when the computer program is executed by a processor, the fire intensity quantization method that fuses instance segmentation and optimizes convolutional autoencoders is implemented.

[0063] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of the computer storage medium would include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0064] Those of ordinary skill in the art will appreciate that the units and algorithm steps of the examples described in connection with the embodiments disclosed in the present application can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present application.

[0065] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art of this technology, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

Claims

1. A fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder, characterized in that: The specific steps include: Step S1, collecting flame images under different combustion states and unifying the resolution of the flame images; Step S2, constructing a fire intensity quantification model, including a YOLO11 seg model, a convolutional autoencoder network and a linear regression model connected in sequence; the YOLO11 seg model is used for instance segmentation of the flame image, and the flame area in the flame image is retained; the convolutional autoencoder network is used to extract deep features of the flame area; the linear regression model is used to predict the fire intensity according to the extracted deep features of the flame area; Step S3, inputting flame images of uniform resolution into the fire intensity quantization model in sequence for training until the total loss function of the fire intensity quantization model converges, thereby completing the training of the fire intensity quantization model; Step S4: re-collect the flame image, adjust the resolution of the collected flame image, input it into the trained fire intensity quantification model, and predict the fire intensity.

2. The fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder according to claim 1, characterized in that: The network structure of the YOLO11 seg model in step S2 specifically includes: a backbone network for extracting flame image features, a neck network for integrating flame image features from different levels of the backbone network, and a head network for instance segmentation; the backbone network includes a first convolutional layer, a second convolutional layer, a first C3K2 module, a first receptive field attention convolutional module, a second C3K2 module, a second receptive field attention convolutional module, a third C3K2 module, a third receptive field attention convolutional module, a fourth C3K2 module, an SPPF module, and a C2PSA module connected in sequence; the neck network includes a first upsampling module, a first Concat module, a fifth C3K2 module, a second upsampling module, a second Concat module, a sixth C3K2 module, a fourth receptive field attention convolutional module, a third Concat module, a seventh C3K2 module, and a C2PSA module connected in sequence. K2 module, the fourth Concat module and the eighth C3K2 module, the output end of the second C3K2 module is also connected to the second Concat module, the output end of the third C3K2 module is also connected to the first Concat module, the output end of the fifth C3K2 module is also connected to the third Concat module, and the output end of the second Concat module is also connected to the sixth C3K2 module; the head network includes a segmentation prototype module, a first image segmentation module, a second image segmentation module and a third image segmentation module, the output end of the sixth C3K2 module is respectively connected to the input end of the segmentation prototype module and the input end of the first image segmentation module, the input end of the second image segmentation module is connected to the output end of the seventh C3K2 module, and the input end of the third image segmentation module is connected to the output end of the eighth C3K2 module.

3. The fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder according to claim 2, characterized in that: The first receptive field attention convolution module, the second receptive field attention convolution module, the third receptive field attention convolution module, the fourth receptive field attention convolution module and the fifth receptive field attention convolution module all use the average pooling function to aggregate the global information of each receptive field feature, and then interact the global information of all receptive field features through a 1×1 group convolution operation, and finally use the Softmax function to weight each receptive field feature: F=Softmax(g 1×1 (AvgPool(X)))×ReLU(Norm(g k×k (X))) Among them, g k×k represents a grouped convolution with a kernel size of k×k, X represents the feature map of the input receptive field attention convolution module, Norm() represents the normalization function, ReLU() represents the activation function, AvgPool() represents the average pooling function, and F represents the receptive field feature after weighting.

4. The fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder according to claim 2, characterized in that: The convolutional autoencoder network is composed of an encoder and a decoder. The encoder is composed of multiple autoencoders. Each subencoder is composed of a third convolution layer, a first normalization layer, a first SiLU activation function, and a first pooling layer. The third convolution layer is used to extract features from the input image and generate a high-dimensional feature map. The first normalization layer is used to ensure the consistency of the distribution of the high-dimensional feature map input in each layer. The first SiLU activation function is used to perform nonlinear transformation on the feature map output by the first normalization layer; The first pooling layer performs downsampling operations on the feature map with nonlinear changes; the decoder is composed of multiple sub-decoders, each sub-decoder is composed of an upsampling layer, a fourth convolutional layer, a second normalization layer, and a second SiLU activation function, and the upsampling layer is used to expand the spatial dimension of the feature map; the fourth convolutional layer is used to reconstruct the local features of the feature map, enhance the detail information and adjust the number of channels; the second normalization layer ensures the consistency of the distribution of the reconstructed feature map input to each layer; the second SiLU activation function is used to perform nonlinear transformation on the feature map output by the second normalization layer.

5. The fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder according to claim 4, characterized in that: Step S3 includes the following sub-steps: Step S3.1, annotate the flame area in each flame image with a segmentation mask as the training label for the YOLO11 seg model; Step S3.2, input a flame image into the YOLO11 seg model for instance segmentation, generate a segmentation prototype and a segmentation mask, and use the segmentation mask to set the pixel values ​​of the background area of ​​the flame image to zero to obtain the flame image after removing the background; Step S3.3, inputting the flame image after removing the background into the convolutional autoencoder network to extract the deep features of the flame area; Step S3.4, input the deep features of the flame area into the linear regression model for regression prediction to predict the fire intensity; Step S3.5, repeat steps S3.2-S3.4 until the total loss function of the fire intensity quantification model converges, completing the training of the fire intensity quantification model.

6. The fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder according to claim 5, characterized in that: In step S3.2, the pixel values ​​of the background area of ​​the flame image are set to zero using the segmentation mask to obtain the flame image after the background is removed. The specific process is as follows: construct an all-zero pixel image with the same size as the original flame image, traverse the segmentation mask, and if the size of the segmentation mask is inconsistent with the size of the original flame image, use the nearest neighbor interpolation method to adjust the size of the segmentation mask to be consistent with the size of the original flame image; Otherwise, the segmentation mask is used to filter out the pixel values ​​of the flame background area in the original flame image, and replace them with all-zero pixel values, and the flame image after removing the background is saved.

7. The fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoder according to claim 6, characterized in that: The total loss function L of the fire intensity quantification model is specifically: L=L1+L2+L3 Among them, L1 represents the loss function of the YOLO11 seg model, L2 represents the loss function of the convolutional autoencoder network, L3 represents the loss function of the linear regression model, N represents the number of flame images used for training, i represents the index of N, and area i represents the normalized area of ​​the flame region in the i-th flame image, B i represents the flame area in the i-th flame image, BCEWithLogits represents the binary cross entropy loss function, and pred i (x,y) represents the segmentation mask generated by the pixel at position (x,y) in the i-th flame image, gt i (x,y) represents the true segmentation mask of the pixel at position (x,y) in the i-th flame image, x i represents the i-th flame image after background removal, z i represents the deep features of the flame area reconstructed by the convolutional autoencoder network, y i represents the real fire intensity of the i-th flame image, Represents the fire intensity predicted by the i-th flame image.

8. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the fire intensity quantification method integrating instance segmentation and optimized convolutional autoencoding as described in any one of claims 1 to 7.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for quantifying fire intensity by integrating instance segmentation and optimizing convolutional autoencoding as described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the fire intensity quantification method of integrating instance segmentation and optimizing convolutional autoencoding as described in any one of claims 1 to 7 is implemented.

Images