Fire detection method and system based on anchor-free frame structure

Abstract

The invention relates to a fire detection method and system based on an anchor-free frame structure, and the method comprises the steps: obtaining a fire data set, and dividing the fire data set into a training set and a test set according to a preset proportion; a fire detection network is constructed based on an Anchor-Free anchor-free framework and comprises a backbone network used for feature extraction, a neck network used for feature fusion and a detection network used for fire monitoring. Wherein a residual module in the backbone network is reconstructed into a multi-branch structure, and an adaptive attention mechanism is embedded to enhance the feature extraction capability; the neck network is constructed based on an improved feature pyramid network; training the fire detection network by using the training set to obtain a trained fire detection network; and obtaining a to-be-detected fire data set, and inputting the to-be-detected fire data set into the trained fire detection network to obtain a detection result.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and mainly to a fire detection method and system based on an anchorless frame structure. Background Technology

[0002] Fire is a common disaster, making rapid and accurate detection and early warning in its initial stages crucial. Early sensor-based detection methods were significantly limited by environmental and spatial constraints, resulting in slow response times. With the development of computer vision technology, image-based flame detection methods have become the mainstream solution due to their high speed, high accuracy, and precise positioning capabilities.

[0003] Current flame detection methods are typically trained on specific datasets. While they achieve high accuracy in recognizing similar backgrounds and flame patterns within the dataset, they struggle to effectively generalize to complex real-world scenarios. For example, they exhibit poor robustness to interference with similar flame colors (such as sunsets and neon lights), different types of combustibles (liquid, solid, and gaseous flames), and dynamic backgrounds (such as crowds and moving vehicles). This is primarily due to the discrepancy between the distribution of training data and the complex and ever-changing physical environment of the real world, resulting in insufficient environmental adaptability of the model. Secondly, due to the varying distances of surveillance cameras from the fire and the degree of fire spread, the pixel scale of flames in images varies greatly. Existing detection algorithms' backbone networks and feature pyramid structures still fall short in their ability to extract and fuse features at extreme scales (especially small target flames at a distance or large-area flames after spread), leading to low recall and localization accuracy for multi-scale flames, particularly small-scale flames.

[0004] Chinese invention patent application CN117593625A discloses an anchor-frame-free intelligent fire detection method. This technical solution constructs a fire detection model based on an improved YOLOv5 model. The improved YOLOv5 model includes a backbone network for feature extraction, a neck network for multi-scale fusion of features extracted from different levels by the backbone network, and a detection head module for target detection and classification. For the structure in the backbone network, the C3 module is replaced with a C2f module. The neck network includes an FPN-PAN bottleneck structure, where the FPN structure is used to transmit strong semantic features from top to bottom, and the PAN structure is used to transmit strong localization features from bottom to top. The target detection head module adopts a multi-level decoupled head. A dataset is collected and divided into training, validation, and test sets, and the fire detection model is trained multiple times according to preset training parameters. The image to be tested is input into the trained fire detection model, and the detection result is output. However, the above technical solution replaces C2f... The module improvements did not change the basic structural logic of the module, only improving channel utilization and gradient transfer efficiency. They cannot adapt to the characteristics of fire targets, which are characterized by varying scales, irregular shapes, and blurred edges. Furthermore, the lack of an attention mechanism means that it cannot dynamically suppress background interference or specifically enhance key fire features. In addition, the above technical solutions adopt an Anchor-Base architecture, which relies on preset anchor frames. Fire targets, on the other hand, have characteristics such as extremely large scale spans, irregular and dynamic shapes, and strong scene diversity. The anchor frame size / proportion of the Anchor-Base architecture needs to be pre-designed, which cannot cover the target shapes of all fire scenarios.

[0005] Therefore, there is an urgent need for a multi-scale fire detection method that can be applied to real-world scenarios. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a fire detection method and system based on an anchorless frame structure.

[0007] The technical solution of the present invention is as follows: On one hand, this invention proposes a fire detection method based on an anchorless frame structure, the method comprising: Obtain the fire dataset and divide it into training and test sets according to a preset ratio; A fire detection network is constructed based on an anchor-free framework, comprising a backbone network for feature extraction, a neck network for feature fusion, and a detection network for fire monitoring. The residual modules in the backbone network are reconstructed into a multi-branch structure, and an adaptive attention mechanism is embedded to enhance feature extraction capabilities. The neck network is constructed based on an improved feature pyramid network. The fire detection network is trained using the training set to obtain the trained fire detection network; Obtain the fire dataset to be detected and input it into the trained fire detection network to obtain the detection results.

[0008] Preferably, the backbone network adopts ResNet50, and each residual module in ResNet50 is reconstructed into a multi-branch structure, wherein the multi-branch structure is specifically a parallel stacking of sub-modules with the same structure.

[0009] Preferably, the 3×3 convolutional layers in ResNet50 are replaced with an adaptive attention mechanism, wherein: The adaptive attention mechanism includes two parallel branches, where the first branch is a 3×3 convolutional layer and the second branch is a 5×5 convolutional layer; The adaptive attention mechanism performs secondary feature extraction on the extracted input features based on the first branch and the second branch, and then fuses the features after secondary extraction to obtain intermediate features. The intermediate features after global average pooling are processed by a convolution kernel with a size of [missing value]. One-dimensional convolution in adjacent Cross-channel interaction is performed between channels to obtain intermediate features after cross-channel interaction; The Softmax function is used to normalize the intermediate features after cross-channel interaction to generate the weight coefficients of each channel. The weight coefficients are multiplied channel by channel by the intermediate features after cross-channel interaction, and then added to the features after secondary feature extraction to obtain the output features of the adaptive attention mechanism.

[0010] Preferably, the improved feature pyramid network is based on the feature pyramid network and introduces upsampling using subpixel convolution; The subpixel convolution is specifically implemented through pixel recombination, transforming pixels of size [missing information]. The input features, the reconstructed size is The output features, where, Indicates the height of the input feature. Indicates the width of the input feature. This represents the number of channels in the input feature. This represents the inflation rate of the input features.

[0011] Preferably, the specific process of feature fusion in the neck network is as follows: The feature pyramid network is used to perform multi-layer bottom-up feature fusion on the output features of the adaptive attention mechanism to obtain fused features. The output features of the adaptive attention mechanism are upsampled using subpixel convolution to obtain the sampled features of the corresponding layer; The fused features are enhanced using the feature enhancement module to obtain enhanced features.

[0012] Preferably, the feature enhancement module is specifically a multi-layer parallel dilated convolution branch, and each dilated convolution branch is set with a progressively increasing dilation rate from top to bottom; If the current branch is the first dilated convolution branch, the output feature of the current dilated convolution branch is obtained by adding the sampled feature of the first level and the fused feature; otherwise, the output feature of the current dilated convolution branch is obtained by adding the sampled feature of the corresponding level and the output feature of the dilated convolution branch of the previous level. The output features of the dilated convolution branch at the last level are called the augmented features.

[0013] Preferably, the detection network includes a classification branch and a regression branch: The classification branch is based on multiple convolutional layers of preset size and outputs the confidence level that the current location is the center area of ​​the fire. The regression branch outputs the predicted fire location target bounding box and the corresponding fire location centrality, whereby the fire location centrality represents the normalized distance between the current fire location and the preset target center.

[0014] On the other hand, the present invention also provides a fire detection system based on an anchorless frame structure, the system comprising: The data acquisition module is used to acquire fire datasets and divide them into training and test sets according to a preset ratio; The training module incorporates an anchorless framework to construct a fire detection network, including a backbone network for feature extraction, a neck network for feature fusion, and a detection network for fire monitoring. The residual modules in the backbone network are reconstructed into a multi-branch structure, and an adaptive attention mechanism is embedded to enhance feature extraction capabilities. The neck network is constructed based on an improved feature pyramid network. The fire detection network is trained using the training set to obtain the trained fire detection network; The detection module is used to acquire the fire dataset to be detected and input it into the trained fire detection network to obtain the detection results.

[0015] In another aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the present invention.

[0016] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the present invention.

[0017] The present invention has the following beneficial effects: 1. This invention adopts an anchor-free frame architecture design, eliminating the dependence on preset anchor boxes and directly predicting the geometric properties of fire targets such as bounding box coordinates and centrality, without the need to filter candidate boxes through anchor box matching; it solves the problem of insufficient adaptability of anchored architectures caused by the large scale span, irregular shape, and strong scene diversity of fire targets; it improves the adaptability of fire scenes, which requires manual design or clustering of anchor boxes to cover all fire target shapes and avoid missed detections caused by anchor box mismatch; it enhances the model's generalization ability, and there is no need to re-adapt anchor boxes after changing fire scenes; 2. This invention reconstructs the ResNet50 residual module into a multi-branch structure with 32 convolutional groups. By stacking sub-modules in parallel, it simultaneously captures the edge texture and global contour features of fire targets. At the same time, it replaces the original 3×3 convolution with a dual-branch adaptive attention mechanism. This solves the problems of strong generality but weak fire feature specificity and insufficient background interference suppression in traditional modules. It enhances the comprehensiveness of feature extraction, with the multi-branch structure covering fire features at different scales, avoiding the omission of extreme scale features by a single module. It also improves the anti-interference ability in complex scenes. 3. This invention optimizes multi-scale feature fusion through an improved feature pyramid network and introduces sub-pixel convolution for upsampling, achieving the goal of not losing feature details and eliminating the need for additional interpolation calculations. A feature enhancement module with multi-layer parallel dilated convolution branches is designed to expand the receptive field by progressively increasing the dilation rate, thereby capturing multi-scale contextual information. Sub-pixel convolution preserves feature details, while dilated convolution supplements the context, improving the accuracy of multi-scale feature fusion. At the same time, sub-pixel convolution is more efficient than traditional interpolation upsampling, ensuring the requirements of real-time detection. Attached Figure Description

[0018] Figure 1 This is a detailed flowchart of an embodiment of the present invention; Figure 2 This is a fire detection network architecture diagram according to an embodiment of the present invention; Figure 3 This is a diagram illustrating the architecture of the adaptive attention mechanism in an embodiment of the present invention. Figure 4 This is an improved feature pyramid network architecture diagram according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0021] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0023] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0024] Example 1: See Figure 1 This invention provides a fire detection method based on an anchorless frame structure, the method comprising: S1. Obtain the fire dataset and divide it into training and test sets according to a preset ratio; The fire dataset is in MS COCO format and includes multiple images from real-world monitoring, experiments, and online videos, covering diverse environmental backgrounds, such as interference with similar flame colors (e.g., sunset, neon lights), combustion materials of different textures (liquid, solid, gaseous flames), and dynamic backgrounds (e.g., crowds, vehicle movement). S2, please refer to Figure 2 A fire detection network is constructed based on an anchor-free framework, including a backbone network for feature extraction, a neck network for feature fusion, and a detection network for fire monitoring. The Anchor-Free architecture is a detection paradigm that does not rely on predefined anchor boxes. It directly predicts the geometric attributes of the target, such as bounding box coordinates or key points. The detection logic does not rely on the IoU matching between the anchor box and the real target, but is based entirely on the target's own features. Compared with the Anchor-Based architecture, it eliminates the redundant process of "anchor box generation → candidate box screening → IoU matching → anchor box correction" in the Anchor-Based architecture. At the same time, the detection logic of the Anchor-Free architecture is independent of the scene, and there is no need to design or cluster specific anchor boxes for specific fire scenes. After model training, it can be directly transferred to different scenes, and its adaptability is not limited by anchor box design. S21. The residual modules in the backbone network are reconstructed into a multi-branch structure and an adaptive attention mechanism is embedded to enhance the feature extraction capability, so as to adaptively focus on key channels and optimize the receptive field selection. S211. The backbone network adopts ResNet50, and each residual module in ResNet50 is reconstructed into a multi-branch structure. Specifically, the multi-branch structure is a parallel stacking of sub-modules with the same structure. That is, using a parallel stacking of blocks with the same structure can improve the accuracy of the model without increasing the number of parameters, while also reducing the number of hyperparameters and adding an attention module. Furthermore, the reconstructed multi-branch structure adopts a group convolution form, where the number of groups is set to 32. S212. Replace the 3×3 convolutional layer in ResNet50 with an adaptive attention mechanism, which includes two parallel branches, wherein the first branch is a 3×3 convolutional layer and the second branch is a 5×5 convolutional layer. Please see Figure 3 The adaptive attention mechanism performs secondary feature extraction on the extracted input features based on the first branch and the second branch, and then fuses the features after secondary extraction to obtain intermediate features; The intermediate features after global average pooling are processed by a convolution kernel with a size of [missing value]. One-dimensional convolution in adjacent Cross-channel interaction is performed between channels to obtain intermediate features after cross-channel interaction; The Softmax function is used to normalize the intermediate features after cross-channel interaction to generate the weight coefficients of each channel. The weight coefficients are multiplied one channel at a time by the intermediate features after cross-channel interaction, and then added to the features after secondary feature extraction to obtain the output features of the adaptive attention mechanism. S22. The neck network is constructed based on an improved feature pyramid network; S221, please refer to Figure 4 The improved feature pyramid network is specifically based on the feature pyramid network, and introduces upsampling using sub-pixel convolution; The subpixel convolution is specifically implemented through pixel recombination, transforming pixels of size [missing information]. The input features, the reconstructed size is The output features, where, Indicates the height of the input feature. Indicates the width of the input feature. This represents the number of channels in the input feature. Indicates the inflation rate of the input features; S222, The specific process of feature fusion in the neck network is as follows: The feature pyramid network is used to perform multi-layer bottom-up feature fusion on the output features of the adaptive attention mechanism to obtain fused features. The output features of the adaptive attention mechanism are upsampled using subpixel convolution to obtain the sampled features of the corresponding layer; By using a feature enhancement module to enhance the fused features, enhanced features are obtained, which solves the problem of limited contextual information caused by the large receptive field of deep features, thereby significantly improving the detection capability of flames of different scales, especially small-scale targets. The feature enhancement module is specifically a multi-layered parallel dilated convolution branch, with each dilated convolution branch having a progressively increasing dilation rate from top to bottom, for example, the dilation rate is {2, 4, 8, 16, 24}. If the current branch is the first dilated convolution branch, the output feature of the current dilated convolution branch is obtained by adding the sampled feature of the first level and the fused feature; otherwise, the output feature of the current dilated convolution branch is obtained by adding the sampled feature of the corresponding level and the output feature of the dilated convolution branch of the previous level. The output features of the dilated convolution branch at the last level are called the augmented features. S23. The detection network includes a classification branch and a regression branch; the detection network predicts each position on the enhanced feature. The classification branch is based on multiple convolutional layers of preset size and outputs the confidence level that the current position is the center region of the fire. Specifically, it outputs a scalar through 3×3 convolution to represent the confidence level that the current position is the center region of the flame target. The regression branch outputs the predicted fire location target bounding box and the corresponding fire location centrality, where the fire location centrality represents the normalized distance between the current fire location and the preset target center. S24. The loss function of the fire detection network adopts the improved generalized cross-union ratio function, and the calculation method is as follows: ; In the formula, Represents the loss function; Represents the generalized intersection-union ratio function; This represents the preset balance coefficient; This represents the error between the predicted and actual values ​​of fire location centrality. S3. Train the fire detection network using the training set to obtain the trained fire detection network; It also includes performance evaluation of the trained fire detection network based on the test set and preset performance evaluation metrics; S4. Obtain the fire dataset to be detected and input it into the trained fire detection network to obtain the detection results; In one specific embodiment, a fire dataset was built based on the MS COCO format. This dataset contains 13,000 images, with 10,000 images in the training set and 3,000 images in the test set. Taking an image from the test set containing open flames on the exterior of a building as an example, the image is first input into the trained fire detection network. The fire detection network extracts multi-layer semantic features of the image through the backbone feature extraction module and performs target localization and category discrimination of the flame area in the detection head. Finally, it outputs the bounding box coordinates of the fire target and its corresponding confidence score, realizing the automatic identification and annotation of the fire area in the image. Comparative experiments were conducted on a self-built dataset with FCOS, RetinaNet, YOLOF, TOOD, and DETR detection algorithms. The training and testing environment was as follows: Intel Core i5-10500 CPU, NVIDIA GeForce RTX 2080 Ti GPU, and Ubuntu 16.0.7 operating system. The programming languages ​​and deep learning frameworks used were Python 3.7 and PyTorch 1.10.0. The specific detection results are shown in Table 1. Table 1. Detection results of the comparative experiment

[0025] in, Refers to average accuracy; This refers to the calculated value when the IoU between the predicted bounding box and the ground truth bounding box is ≥ 0.5. ; This indicates the detection accuracy for small-sized targets; This indicates the detection accuracy for medium-sized targets; This indicates the detection accuracy for large targets.

[0026] Example 2: This embodiment provides a fire detection system based on an anchorless frame structure, the system comprising: The data acquisition module is used to acquire fire datasets and divide them into training and test sets according to a preset ratio; The training module incorporates an anchorless framework to construct a fire detection network, including a backbone network for feature extraction, a neck network for feature fusion, and a detection network for fire monitoring. The residual modules in the backbone network are reconstructed into a multi-branch structure, and an adaptive attention mechanism is embedded to enhance feature extraction capabilities. The neck network is constructed based on an improved feature pyramid network. The fire detection network is trained using the training set to obtain the trained fire detection network; The detection module is used to acquire the fire dataset to be detected and input it into the trained fire detection network to obtain the detection results.

[0027] Example 3: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a fully automated withstand voltage performance testing system compatible with multiple types of surge arresters as described in any one of Embodiment 1.

[0028] Example 4: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a fully automated withstand voltage performance testing system compatible with multiple types of surge arresters as described in any one of Embodiment 1.

[0029] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0030] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0031] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0032] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0033] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A fire detection method based on an anchorless frame structure, characterized in that, The method includes: Obtain the fire dataset and divide it into training and test sets according to a preset ratio; A fire detection network is constructed based on an anchor-free framework, comprising a backbone network for feature extraction, a neck network for feature fusion, and a detection network for fire monitoring. The residual modules in the backbone network are reconstructed into a multi-branch structure, and an adaptive attention mechanism is embedded to enhance feature extraction capabilities. The neck network is constructed based on an improved feature pyramid network. The fire detection network is trained using the training set to obtain the trained fire detection network; Obtain the fire dataset to be detected and input it into the trained fire detection network to obtain the detection results.

2. The fire detection method based on an anchorless frame structure according to claim 1, characterized in that, The backbone network adopts ResNet50, and each residual module in ResNet50 is reconstructed into a multi-branch structure. Specifically, the multi-branch structure consists of sub-modules with the same structure stacked in parallel.

3. The fire detection method based on an anchorless frame structure according to claim 2, characterized in that, Replace the 3×3 convolutional layers in ResNet50 with an adaptive attention mechanism, where: The adaptive attention mechanism includes two parallel branches, where the first branch is a 3×3 convolutional layer and the second branch is a 5×5 convolutional layer; The adaptive attention mechanism performs secondary feature extraction on the extracted input features based on the first branch and the second branch, and then fuses the features after secondary extraction to obtain intermediate features. The intermediate features after global average pooling are processed by a convolution kernel with a size of [missing value]. One-dimensional convolution in adjacent Cross-channel interaction is performed between channels to obtain intermediate features after cross-channel interaction; The Softmax function is used to normalize the intermediate features after cross-channel interaction to generate the weight coefficients of each channel. The weight coefficients are multiplied channel by channel by the intermediate features after cross-channel interaction, and then added to the features after secondary feature extraction to obtain the output features of the adaptive attention mechanism.

4. The fire detection method based on an anchorless frame structure according to claim 1, characterized in that, The improved feature pyramid network is specifically based on the feature pyramid network, and introduces upsampling using sub-pixel convolution. The subpixel convolution is specifically implemented through pixel recombination, transforming pixels of size [missing information]. The input features, the reconstructed size is The output features, where, Indicates the height of the input feature. Indicates the width of the input feature. This represents the number of channels in the input feature. This represents the inflation rate of the input features.

5. A fire detection method based on an anchorless frame structure according to claim 4, characterized in that, The adaptive feature enhancement module includes multiple convolutional branches with different kernel sizes to extract feature information under different receptive fields. The output features of the convolutional branches are then subjected to global average pooling and generated with one-dimensional convolution and the Softmax function to generate corresponding weight coefficients. These weight coefficients are used to characterize the importance of features at each scale in the current image.

6. A fire detection method based on an anchorless frame structure according to claim 5, characterized in that, The feature enhancement module is specifically a multi-layered parallel dilated convolution branch, with each dilated convolution branch having a progressively increasing dilation rate from top to bottom; If the current branch is the first dilated convolution branch, the output feature of the current dilated convolution branch is obtained by adding the sampled feature of the first level and the fused feature; otherwise, the output feature of the current dilated convolution branch is obtained by adding the sampled feature of the corresponding level and the output feature of the dilated convolution branch of the previous level. The output features of the dilated convolution branch at the last level are called the augmented features.

7. A fire detection method based on an anchorless frame structure according to claim 1, characterized in that, The detection network includes a classification branch and a regression branch: The classification branch is based on multiple convolutional layers of preset size and outputs the confidence level that the current location is the center area of ​​the fire. The regression branch outputs the predicted fire location target bounding box and the corresponding fire location centrality, whereby the fire location centrality represents the normalized distance between the current fire location and the preset target center.

8. A fire detection system based on an anchorless frame structure, characterized in that, The system includes: The data acquisition module is used to acquire fire datasets and divide them into training and test sets according to a preset ratio; The training module incorporates an anchorless framework to construct a fire detection network, including a backbone network for feature extraction, a neck network for feature fusion, and a detection network for fire monitoring. The residual modules in the backbone network are reconstructed into a multi-branch structure, and an adaptive attention mechanism is embedded to enhance feature extraction capabilities. The neck network is constructed based on an improved feature pyramid network. The fire detection network is trained using the training set to obtain the trained fire detection network; The detection module is used to acquire the fire dataset to be detected and input it into the trained fire detection network to obtain the detection results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images