A flame target detection method based on improved YOLOv11 and related equipment

Abstract

This application discloses a flame target detection method and related equipment based on an improved YOLOv11 model. The method includes: acquiring the backbone network, neck network, and detection head of the YOLOv11 model; determining the replacement positions and number of convolutional modules to be replaced in the backbone network and neck network, and then replacing the corresponding convolutional modules with DSConv modules according to the replacement positions and number of replacements, thereby obtaining a first improved backbone network and a first improved neck network; embedding a first MSCA module into the first improved backbone network and a second MSCA module into the first improved neck network, finally obtaining an improved YOLOv11 model; and inputting a flame image into the improved YOLOv11 model for detection. This application can reduce the false detection rate and, without sacrificing the accuracy of flame feature extraction, reduce the computational load of the model and improve detection efficiency, and can be widely applied in the field of target detection technology.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and in particular to a flame target detection method and related equipment based on an improved YOLOv11. Background Technology

[0002] Overhead transmission lines often traverse rugged, densely vegetated mountainous and forested areas. Wildfires pose a serious threat to the safe operation of these lines, and power outages caused by wildfires can directly result in economic losses. The risk of wildfires increases significantly, especially during peak periods of outdoor fire use such as Qingming Festival and Spring Festival. The high concentrations of smoke produced by wildfires can drastically reduce air insulation strength, endangering the safety of power grid equipment.

[0003] To address this challenge, power grid companies are accelerating the development of digital monitoring, installing online wildfire monitoring devices in hazardous areas to collect images and use target detection models to detect wildfires. However, early wildfires often manifest as small targets with a very low pixel percentage (<1%), and their shape changes rapidly due to wind direction. Existing detection models lack feature extraction capabilities and have weak ability to perceive details and model context for small targets. They are prone to misidentifying non-fire targets such as cloud reflections and streetlight halos as flames or smoke, resulting in a high false detection rate. Secondly, a single transmission line generates more than 100,000 image data per day, and existing detection models struggle to balance detection accuracy with computational efficiency, failing to meet the engineering requirements for real-time monitoring of wildfires along overhead transmission lines. Summary of the Invention

[0004] The main objective of this application is to propose a flame target detection method and related equipment based on an improved YOLOv11, which can reduce the missed detection of small targets, lower the false detection rate, and reduce the computational load of the model and improve the detection efficiency without sacrificing the accuracy of flame feature extraction.

[0005] To achieve the above objectives, one aspect of this application proposes a flame target detection method based on an improved YOLOv11, comprising: Acquire flame images; Obtain the framework of the YOLOv11 model, which includes a backbone network, a neck network, and a detection head; Calculate the computational cost of each convolutional module in the backbone network and the neck network; Based on the computational load, determine the replacement location and number of convolutional modules to be replaced in the backbone network and the neck network; Based on the replacement position and replacement quantity, the convolutional modules at corresponding positions in the backbone network and the neck network are replaced with DSConv modules to obtain the first improved backbone network and the first improved neck network. A first MSCA module is embedded at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion position and / or downsampling transition position of several C3K2 modules in the first improved neck network to obtain a second improved neck network. Based on the second improved backbone network, the second improved neck network, and the detection head, an improved YOLOv11 model is obtained; The flame image is input into the improved YOLOv11 model for detection, and the flame target detection result is obtained.

[0006] In some embodiments, calculating the computational cost of each convolutional module in the backbone network and the neck network includes: Obtain the feature map size and number of channels corresponding to each convolutional module in the backbone network and the neck network; The computational cost of each convolutional module in the backbone network and the neck network is calculated based on the feature map size and number of channels corresponding to each convolutional module.

[0007] In some embodiments, determining the replacement location and number of convolutional modules to be replaced in the backbone network and the neck network based on the computational load includes: Each convolutional module is sorted in descending order of computational complexity to obtain a candidate replacement sequence; From the candidate replacement sequence, the first N convolutional modules are selected sequentially and replaced with DSConv modules to construct multiple intermediate improved models with different replacement scales, where N is a positive integer greater than zero; Ablation experiments were conducted on the multiple intermediate improved models, and the replacement positions and number of convolutional modules to be replaced in the backbone network and the neck network were determined based on the results of the ablation experiments.

[0008] In some embodiments, inputting the flame image into the improved YOLOv11 model for detection to obtain flame target detection results includes: The second improved backbone network performs multi-scale feature extraction on the flame image to obtain a multi-scale feature map; The multi-scale feature map is input into the second improved neck network, and the second improved neck network fuses the multi-scale feature map to obtain a multi-scale feature pyramid; The multi-scale feature pyramid is input into the detection head, and the classification and regression branches in the detection head are decoupled for prediction to obtain the candidate box for each category and the confidence level of the candidate box. Each candidate box is filtered based on the confidence level corresponding to each category to obtain the flame target detection result.

[0009] In some embodiments, the second improved backbone network includes a plurality of first DSConv modules, a plurality of first C3K2 modules, an SPPF module, a C2PSA module, and a first MSCA module; The second improved backbone network performs multi-scale feature extraction on the flame image to obtain a multi-scale feature map, including: The second improved backbone network performs primary feature extraction and downsampling on the flame image through multiple first DSConv modules in the first layer to obtain a first-scale feature map; The first-scale feature map is enhanced and downsampled by multiple first DSConv modules and multiple first C3K2 modules at the second level to obtain a second-scale feature map; the second-scale feature map is then fused by multi-scale pooling through the SPPF module to obtain the multi-scale pooled fused feature map. The features obtained after multi-scale pooling fusion are sequentially passed through the C2PSA module and the first MSCA module for feature attention enhancement to obtain a third-scale feature map.

[0010] In some embodiments, the second improved neck network includes a plurality of upsampling modules, a plurality of second C3K2 modules, a plurality of second MSCA modules, and a plurality of second DSConv modules; The step of inputting the multi-scale feature map into the second improved neck network, and the second improved neck network fusing the multi-scale feature map to obtain a multi-scale feature pyramid, includes: The third-scale feature map is upsampled through the first upsampling module, and the upsampled third-scale feature map is laterally connected and feature fused with the second-scale feature map to obtain the second-scale fused feature map. The second scale fusion feature map is upsampled through the second upsampling module, and the upsampled second scale fusion feature map is horizontally connected and feature fused with the first scale feature map to obtain the first scale fusion feature map. The third-scale feature map, the second-scale fused feature map, and the first-scale fused feature map are convolved and integrated through the corresponding second C3K2 module, multi-scale context enhancement is performed through the corresponding second MSCA module, and convolution operation and scale alignment are performed through the corresponding second DSConv module to obtain a multi-scale feature pyramid.

[0011] In some embodiments, the step of filtering each candidate box according to the confidence level corresponding to each category to obtain the flame target detection result includes: Sort the candidate boxes for each category in descending order of confidence to obtain a candidate box list; Remove candidate boxes from the candidate box list whose confidence level is lower than the first preset threshold to obtain a preliminary candidate box list; Non-maximum suppression is performed on each category in the initial candidate box list, and the box with the highest confidence is selected as the retained box in turn. Calculate the intersection-union ratio (IUU) of the retained box with the remaining candidate boxes of the same category, and delete candidate boxes whose IUU is greater than a second preset threshold to obtain a deduplicated set of flame detection boxes; Based on the deduplicated set of flame detection frames, the target detection result is obtained.

[0012] To achieve the above objectives, another aspect of this application proposes a flame target detection device based on an improved YOLOv11, the device comprising: The image acquisition module is used to acquire flame images; The framework acquisition module is used to acquire the framework of the YOLOv11 model, which includes a backbone network, a neck network, and a detection head. The computational complexity statistics module is used to calculate the computational complexity of each convolutional module in the backbone network and the neck network. The replacement position determination module is used to determine the replacement position and number of convolutional modules to be replaced in the backbone network and the neck network based on the computational load. The first improvement module is used to replace the convolutional modules at corresponding positions in the backbone network and the neck network with DSConv modules according to the replacement position and the replacement quantity, so as to obtain the first improved backbone network and the first improved neck network. The second improved module is used to embed a first MSCA module at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion position and / or downsampling transition position corresponding to several C3K2 modules in the first improved neck network to obtain a second improved neck network. An improved model generation module is used to obtain an improved YOLOv11 model based on the second improved backbone network, the second improved neck network, and the detection head; The target detection module is used to input the flame image into the improved YOLOv11 model for detection and obtain the flame target detection result.

[0013] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0014] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0015] This application also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0016] The embodiments of this application include at least the following beneficial effects: This application provides a flame target detection method based on an improved YOLOv11. This method first acquires a flame image; then acquires the framework of a YOLOv11 model, which includes a backbone network, a neck network, and a detection head; calculates the computational cost of each convolutional module in the backbone and neck networks; based on the computational cost, determines the replacement positions and number of convolutional modules to be replaced in the backbone and neck networks; and based on the replacement positions and number of replacements, replaces the corresponding convolutional modules in the backbone and neck networks with DSConv modules. A first improved backbone network and a first improved neck network are obtained. A first MSCA module is embedded at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network to obtain a second improved neck network. Based on the second improved backbone network, the second improved neck network, and the detection head, an improved YOLOv11 model is obtained. The flame image is input into the improved YOLOv11 model for detection to obtain the flame target detection result. By embedding MSCA modules at the end of the residual connection branches of the C2PSA module in the first improved backbone network, and embedding multiple MSCA modules after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network, the ability to extract multi-scale dynamic features of flames / smoke is improved, the ability to capture early small targets is enhanced, the missed detection of small targets is reduced, and the false detection rate is significantly reduced. At the same time, the replacement positions and number of convolutional modules with DSConv modules are determined according to the computational load, which can reduce the computational load of the model and improve the detection efficiency without losing the accuracy of flame feature extraction, thus meeting the engineering requirements of real-time monitoring of wildfires in overhead power line corridors. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of this application; Figure 2 This is a flowchart of a flame target detection method based on an improved YOLOv11 provided in an embodiment of this application; Figure 3 This is a schematic diagram of the framework of the improved YOLOv11 model provided in the embodiments of this application; Figure 4 This is a schematic diagram comparing the detection results provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of a flame target detection device based on an improved YOLOv11 provided in an embodiment of this application; Figure 6This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0019] It is understood that the terms "first," "second," "third," "fourth," etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where 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 (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or 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.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Because early wildfires often manifest as small targets of smoke with a very low pixel percentage (<1%), and their shape changes rapidly due to wind direction, existing detection models lack sufficient feature extraction capabilities and have weak ability to perceive details and model contexts of small targets. They are prone to misjudging non-fire targets such as cloud reflections and streetlight halos as flames or smoke, resulting in a high false detection rate. Secondly, a single transmission line generates more than 100,000 image data per day, and existing detection models struggle to balance detection accuracy with computational efficiency, failing to meet the engineering requirements for real-time monitoring of wildfires along overhead power line corridors.

[0023] In view of this, this application provides a flame target detection method and related equipment based on an improved YOLOv11. The method first acquires a flame image; then acquires the framework of a YOLOv11 model, which includes a backbone network, a neck network, and a detection head; calculates the computational cost of each convolutional module in the backbone and neck networks; based on the computational cost, determines the replacement positions and number of convolutional modules to be replaced in the backbone and neck networks; and based on the replacement positions and number of replacements, replaces the corresponding convolutional modules in the backbone and neck networks with DSConv modules, thus obtaining... The first improved backbone network and the first improved neck network are constructed. A first MSCA module is embedded at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network to obtain a second improved neck network. Based on the second improved backbone network, the second improved neck network and the detection head, an improved YOLOv11 model is obtained. The flame image is input into the improved YOLOv11 model for detection to obtain the flame target detection result. By embedding MSCA modules at the end of the residual connection branches of the C2PSA module in the first improved backbone network and embedding multiple MSCA modules after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network, the ability to extract multi-scale dynamic features of flames / smoke is improved, the ability to capture early small targets is enhanced, the missed detection of small targets is reduced, and the false detection rate is significantly reduced. Simultaneously, by determining the replacement positions and number of convolutional modules with DSConv modules based on computational complexity, the computational complexity of the model can be reduced and the detection efficiency improved without sacrificing the accuracy of flame feature extraction, meeting the engineering requirements for real-time monitoring of wildfires along overhead power lines. The flame target detection method based on improved YOLOv11 provided in this application relates to the field of target detection technology. The flame target detection method based on improved YOLOv11 provided in this application can be applied to terminals, servers, or software running on terminals or servers.In some embodiments, the terminal may be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited thereto; the server may be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or as a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server may also be a node server in a blockchain network; the software may be an application implementing a flame target detection method based on the improved YOLOv11, but is not limited to the above forms.

[0024] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0025] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided in an embodiment of this application. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0026] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0027] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0028] Terminal 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc. It can also be a vehicle-mounted terminal of the various device types described above, but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment does not impose any limitations.

[0029] For example, based on Figure 1 The implementation environment shown in this application embodiment provides a flame target detection method based on improved YOLOv11. The following description uses the application of this flame target detection method based on improved YOLOv11 in server 101 as an example. It can be understood that this method can also be applied in terminal 102.

[0030] Reference Figure 2 , Figure 2 This flowchart illustrates a flame target detection method based on an improved YOLOv11 applied to a server, as provided in this application embodiment. The execution subject of this method can be any of the aforementioned computer devices (including servers or terminals). (Refer to...) Figure 2 The method may include the following steps: S100, Obtain the flame image.

[0031] For example, embodiments of this application acquire flame images from multiple sources, perform standardized annotation and normalization processing to enhance the flame images, thereby constructing a dedicated dataset of high-definition wildfire flame images suitable for overhead power transmission line corridor scenarios, providing a high-quality image data foundation for the subsequent training, verification and detection of the YOLOv11 model.

[0032] Specifically, firstly, image data was collected from multiple sources to ensure sample diversity: for example, high-definition visible light / thermal imaging images taken by inspection drones and 24-hour monitoring data from intelligent observation cameras on iron towers / mountain tops were extracted from historical alarm data of a provincial power grid transmission line intelligent monitoring system; simultaneously, images of wildfires were collected online, and samples similar to the background of the transmission line corridor were selected to supplement the dataset. The dataset includes two categories: smoke and fire images and background images, with the background images covering typical interference objects such as sunlight, mountain fog, and sky clouds.

[0033] Secondly, the acquired images were labeled: the LabelImg labeling tool was used to label smoke targets (category name: smoke) and flame targets (category name: fire) with rectangular bounding boxes, and the target category and spatial coordinate parameters (xmin, xmax, ymin, ymax) were recorded, generating an XML labeling file in PASCAL VOC format. Background images were not labeled. To ensure the reliability of the labeling, randomly selected labeled images were used to evaluate category consistency using the Fleiss' Kappa coefficient, and the average bounding box intersection-union ratio was used to evaluate location consistency.

[0034] Finally, the dataset was partitioned and augmented: the dataset was divided into training, validation and test sets in an 8:1:1 ratio to construct three types of flame image datasets: real samples, collected samples and mixed samples. The collected flame image datasets were preprocessed, including scaling (to adapt to 640×640 input size), normalization, random flipping, brightness or contrast adjustment, etc., to expand sample diversity, alleviate class imbalance problem and improve the model's generalization ability.

[0035] S200. Obtain the framework of the YOLOv11 model, wherein the framework of the YOLOv11 model includes a backbone network, a neck network, and a detection head.

[0036] S300. Calculate the computational cost of each convolutional module in the backbone network and the neck network.

[0037] For example, the steps of calculating the computational cost of each convolutional module in the backbone network and the neck network include S310-S320: S310. Obtain the feature map size and number of channels corresponding to each convolutional module in the backbone network and the neck network; S320. Calculate the computational cost of each convolutional module in the backbone network and the neck network based on the feature map size and number of channels corresponding to each convolutional module. Specifically, based on the feature map size (height and width) and the number of channels, the number of floating-point operations (FLOPs) is used as a measure of computational cost to calculate the computational cost of each standard convolutional module, thus obtaining the computational cost of each standard convolutional module.

[0038] S400. Based on the computational load, determine the replacement position and number of convolutional modules to be replaced in the backbone network and the neck network.

[0039] The replacement position and number of convolutional modules to be replaced refer to the replacement position and number of DSConv modules in the backbone and neck networks.

[0040] The DSConv module refers to the distributed offset convolution module. It introduces a variable quantization kernel (VQK) and a distributed offset mechanism on the basis of depthwise separable convolution. By using a small number of full-precision parameters to correct the spatial and channel distribution of the quantization kernel, it achieves adaptive sampling of input features. This distributed offset convolution module can significantly compress parameter storage and reduce computation while maintaining the expressive power of the convolution operator.

[0041] For example, the steps of determining the replacement location and number of convolutional modules to be replaced in the backbone network and the neck network include S410-S430: S410. Sort each convolutional module in descending order of computational complexity to obtain a candidate replacement sequence.

[0042] S420. From the candidate replacement sequence, select the first N convolutional modules in sequence to replace them with DSConv modules to construct multiple intermediate improved models with different replacement scales, where N is a positive integer greater than zero. In this embodiment of the application, when selecting the first N convolutional modules in the candidate replacement sequence to replace the DSConv modules in descending order of computational cost, priority is given to selecting convolutional modules in the middle positions of the backbone network and neck network that have a high proportion of computational cost and low sensitivity to localization / classification.

[0043] S430. Perform ablation experiments on the multiple intermediate improved models respectively, and determine the replacement positions and number of convolutional modules to be replaced in the backbone network and the neck network based on the results of the ablation experiments (that is, the replacement positions and number of convolutional modules in the backbone network and the neck network to be replaced with DSConv modules).

[0044] Specifically, ablation experiments were conducted on each intermediate improved model to evaluate its detection accuracy and inference speed (FPS). During the ablation experiments, when the detection accuracy met the preset threshold requirements and the inference speed reached the target, the replacement was stopped, thereby determining the final replacement position and number of DSConv modules. This approach can reduce the computational load of the model and improve detection efficiency without sacrificing the accuracy of flame feature extraction.

[0045] S500: According to the replacement position and replacement quantity, replace the convolutional modules at corresponding positions in the backbone network and the neck network with DSConv modules to obtain the first improved backbone network and the first improved neck network.

[0046] The DSConv module in this embodiment can be replaced with other efficient convolutional structures, such as depthwise separable convolution, Ghost convolution, deformable convolution, etc. in the MobileNet series, or different quantization strategies can be adopted, as long as the computational complexity and model size can be reduced while ensuring detection accuracy.

[0047] S600. A first MSCA module is embedded at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network to obtain a second improved neck network.

[0048] In this embodiment, a multi-scale convolutional attention (MSCA) module is introduced. This MSCA module uses multi-branch deep strip convolution combined with channel attention mechanism to achieve dynamic aggregation of multi-scale contextual information, which can take into account both local details and global receptive field, and is used to improve the expressive ability of small targets in complex backgrounds. Specifically, the processing steps of the MSCA module include: (1) extracting multi-scale contextual features from the input features by multi-branch deep strip convolution; wherein, multi-branch deep strip convolution uses convolution kernels of different scales and directions to capture diverse contextual information; (2) concatenating or fusing the features output by each branch to obtain aggregated features; (3) performing global average pooling on the aggregated features to obtain channel description vectors, and generating channel weights through 1×1 convolution or multilayer perceptron; (4) using the channel weights to recalibrate the aggregated features to obtain enhanced features; (5) outputting the enhanced features, or adding the residuals of the features of the main branch to maintain training stability.

[0049] For example, in this embodiment of the application, a first MSCA module is inserted at the end of the residual connection branch of the C2PSA module in the first improved backbone network. After the C2PSA module performs global semantic recalibration on the deep feature map based on its internal polarization self-attention mechanism, the first MSCA module further enhances the high-level semantic features through parallel multi-scale convolution and channel attention, thereby improving the ability to focus on flame / smoke targets and suppressing background interference such as clouds, fog, and lights.

[0050] By embedding multiple second MSCA modules after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network, multi-scale context modeling can effectively compensate for the loss of small target details caused by downsampling and enhance the perception ability of mid-to-low-level features for small flames and thin smoke.

[0051] In this embodiment, by embedding the MSCA module at the aforementioned positions in the first improved neck network and the first improved backbone network respectively, a stronger multi-scale attention capability is injected into the network without destroying the stability of the original backbone network and neck network, thereby achieving multi-level and multi-scale feature enhancement of flame / smoke targets.

[0052] The MSCA module in this embodiment can also be replaced with other attention modules with multi-scale modeling capabilities, such as CBAM, ECA-Net variants, pyramid attention modules, etc., but the overall idea of ​​embedding attention structures at specific positions in the first improved network and the first improved backbone network should still be maintained.

[0053] S700. Based on the second improved backbone network, the second improved neck network, and the detection head, an improved YOLOv11 model is obtained.

[0054] S800: Input the flame image into the improved YOLOv11 model for detection to obtain the flame target detection result. Exemplarily, the step of inputting the flame image into the improved YOLOv11 model for detection to obtain the flame target detection result includes S810-S830: S810. The second improved backbone network performs multi-scale feature extraction on the flame image to obtain a multi-scale feature map. The second improved backbone network includes multiple first DSConv modules, multiple first C3K2 modules, SPPF modules, C2PSA modules, and first MSCA modules; Furthermore, the second improved backbone network performs multi-scale feature extraction on the flame image to obtain a multi-scale feature map, including steps S811-S814: S811. The second improved backbone network performs primary feature extraction and downsampling on the flame image through multiple first DSConv modules at the first level to obtain a first-scale feature map. For example, the first layer includes two first DSConv modules, namely the first first DSConv module and the second first DSConv module. Specifically, the preprocessed flame image is first input into the second improved backbone network, and then the two first DSConv modules of the first layer are used to perform primary feature extraction and downsampling operations to capture the basic texture features in the image and obtain a shallow feature map (i.e., a first-scale feature map).

[0055] S812. The first-scale feature map is enhanced and downsampled by multiple first DSConv modules and multiple first C3K2 modules at the second level to obtain the second-scale feature map. For example, the second layer may include three first DSConv modules; specifically, the shallow feature map (i.e., the first scale feature map) is processed by the three first DSConv modules and multiple first C3K2 modules of the second layer to enhance the details of the features and downsample them to obtain the middle layer feature map (the second scale feature map).

[0056] S813. The second scale feature map is fused by multi-scale pooling through the SPPF module to obtain the multi-scale pooled feature.

[0057] S814. The features after multi-scale pooling fusion are sequentially passed through the C2PSA module and the first MSCA module for feature attention enhancement to obtain a third-scale feature map. Specifically, the features after multi-scale pooling fusion are input into the C2PSA module for deep extraction and enhancement of mid-to-high-level semantic features. Then, at the end of the residual connection branch of the C2PSA module, attention enhancement is applied to the extracted mid-to-high-level semantic features through the embedded first MSCA module to obtain more representative deep features, and finally a deep feature map (third-scale feature map) is obtained.

[0058] S820. The multi-scale feature map is input into the second improved neck network, and the second improved neck network fuses the multi-scale feature map to obtain a multi-scale feature pyramid. The second improved neck network includes multiple upsampling modules, multiple second C3K2 modules, multiple second MSCA modules, and multiple second DSConv modules; Furthermore, the second improved neck network fuses multi-scale feature maps to obtain a multi-scale feature pyramid, including steps S821-S823: S821. The third-scale feature map is upsampled through the first upsampling module, and the upsampled third-scale feature map is laterally connected and feature fused with the second-scale feature map to obtain the second-scale fused feature map. Among them, the horizontal connection can be achieved by feature splicing or element-by-element addition.

[0059] S822. Upsample the second scale fusion feature map through the second upsampling module, and then connect the upsampled second scale fusion feature map with the first scale feature map laterally and fuse the features to obtain the first scale fusion feature map. S823. The third-scale feature map, the second-scale fused feature map, and the first-scale fused feature map are respectively convolved and integrated through the corresponding second C3K2 module, multi-scale context enhancement is performed through the corresponding second MSCA module, and convolution operation and scale alignment are performed through the corresponding second DSConv module to obtain a multi-scale feature pyramid.

[0060] Specifically, the third-scale feature map, the second-scale fused feature map, and the first-scale fused feature map are convolved and integrated using the corresponding second C3K2 modules, then multi-scale context enhancement is performed using the corresponding second MSCA modules, and finally convolutional operations and scale alignment are performed using the corresponding second DSConv modules to obtain a multi-scale feature pyramid. In this implementation, multiple second MSCA modules are embedded after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network to enhance the response of the flame / smoke region and suppress cloud and fog background interference.

[0061] In this embodiment, the second improved neck network, relying on the multi-scale fusion advantage of the FPN structure and the second MSCA module, can efficiently fuse multi-scale feature maps into a multi-scale feature pyramid.

[0062] S830. Input the multi-scale feature pyramid into the detection head, and perform decoupled prediction through the classification branch and regression branch in the detection head to obtain the candidate box for each category and the confidence level corresponding to the candidate box. For example, the detection head adopts a decoupled structure design and introduces the DSConv module in the core convolutional layer to replace the traditional standard convolutional module, which further optimizes the computational efficiency while ensuring detection accuracy. At the same time, according to the Anchor-Free decoding mechanism, it realizes accurate localization and category determination of flame and smoke targets. The steps include: for the feature map of each scale in the multi-scale feature pyramid, the feature map is first passed through a shared convolutional layer containing the DSConv module for feature integration; the integrated features are input into the classification branch and the regression branch respectively: the classification branch outputs the flame / smoke category score, and the regression branch outputs the bounding box regression parameters; and the Anchor-Free decoding is used to map the regression parameters into candidate boxes in the image coordinate system, and the category score and the preset target score are calculated to obtain the confidence level corresponding to each candidate box.

[0063] S840. Filter each candidate box according to the confidence level corresponding to each category to obtain the flame target detection result.

[0064] For example, the steps of filtering each candidate box according to the confidence level corresponding to each category to obtain the flame target detection result include S841-S845: S841. Sort the candidate boxes for each category in descending order of confidence to obtain a candidate box list; Specifically, first, all candidate boxes are aggregated, including all candidate boxes for the two categories of flame and smoke, and the confidence level of each candidate box is associated with it. Then, the candidate boxes are sorted by category, arranged in descending order of confidence level, to obtain a list of candidate boxes for each category.

[0065] S842. Remove candidate boxes from the candidate box list whose confidence level is lower than the first preset threshold to obtain a preliminary candidate box list.

[0066] S843. Perform non-maximum suppression processing on each category in the preliminary candidate box list, and select the current highest confidence box as the retained box in turn. For example, in the embodiments of this application, the NMS strategy or the Soft-NMS strategy can be used for non-maximum suppression processing.

[0067] S844. Calculate the intersection-union ratio (IUU) of the retained box with the remaining candidate boxes of the same category, and delete the candidate boxes whose IUU is greater than the second preset threshold to obtain a set of deduplicated flame detection boxes. Specifically, based on the selected retained bounding box, the intersection-union ratio (IUR) between the retained bounding box and the remaining candidate boxes of the same category is calculated. The IUR is used to quantify the degree of overlap between two candidate boxes. The calculated IUR is compared with a second preset threshold, and all candidate boxes of the same category with an IUR greater than the second preset threshold are deleted to avoid the same target being detected multiple times and duplicate candidate boxes appearing. Subsequently, among the remaining candidate boxes, the candidate box with the highest current confidence is selected as the new retained box, and the above IUR calculation and elimination steps are repeated until there are no candidate boxes of the same category with excessive overlap, and finally the deduplicated flame detection box set is obtained.

[0068] S845. Based on the deduplicated set of flame detection frames, the target detection result is obtained.

[0069] It should be noted that the improved YOLOv11 model in step S800 has been pre-trained using the high-resolution wildfire image dataset constructed above.

[0070] To explain in detail the principles of the technical solution of this application, the overall process of this application will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principles of this application and should not be regarded as a limitation of this application.

[0071] In one specific embodiment, the flame target detection process based on the improved YOLOv11 in this application includes the following steps: S1. Collect a dataset of flame images; By acquiring flame images from multiple sources, standardizing and enhancing the flame images through normalized annotation and processing, a dedicated dataset of high-definition wildfire flame images suitable for overhead power transmission line corridor scenarios is constructed.

[0072] (1) Data Acquisition: Image data was collected through multiple sources, such as high-definition visible light / thermal imaging images taken by inspection drones and 24-hour monitoring data from intelligent observation cameras on iron towers / mountain tops, extracted from historical alarm data of a provincial power grid transmission line intelligent monitoring system; at the same time, wild environmental fire images were collected through the network, and samples similar to the background of the transmission line channel were selected to supplement the dataset. The dataset contains two categories: smoke and fire images and background images, among which the background images cover typical interference objects such as sunlight, mountain fog, and sky clouds.

[0073] (2) Labeling: The LabelImg labeling tool was used to label smoke targets (category name: smoke) and flame targets (category name: fire) with rectangular boxes, and the target category and spatial coordinate parameters (xmin, xmax, ymin, ymax) were recorded to generate an XML labeling file in PASCAL VOC format. The background image was not labeled. To ensure the reliability of the labeling, the Fleiss' Kappa coefficient was used to evaluate the category consistency of randomly selected labeled images, and the average bounding box intersection-union ratio was used to evaluate the position consistency.

[0074] (3) Dataset partitioning and augmentation: The dataset is divided into training set, validation set and test set in a ratio of 8:1:1. Three types of flame image datasets are constructed: real samples, collected samples and mixed samples. The collected flame image datasets are preprocessed, including scaling (adapting to 640×640 input size), normalization, random flipping, brightness / contrast adjustment, etc., to expand the diversity of samples, alleviate the class imbalance problem and improve the generalization ability of the model.

[0075] S2. Improved network structure design of YOLOv11 model The basic framework of the YOLOv11 model includes an input layer, a backbone network, a neck network, and a head. The C3K2 block, SPPF module, and C2PSA block in the basic framework design of the YOLOv11 model enhance multi-scale feature extraction and spatial attention capabilities. In this application example, improvements are made to the backbone network and neck network within the basic framework of the YOLOv11 model. The improvement process is as follows: S2.1 Determine the replacement locations and number of DSConv modules in the backbone and neck networks, specifically: The feature map size and number of channels for each convolutional module in the backbone and neck networks are obtained. Based on the feature map size (height and width) and number of channels, floating-point operations (FLOPs) are used as a measure of computational cost to calculate the computational cost of each standard convolutional module. Each convolutional module is then sorted in descending order of computational cost to obtain a candidate replacement sequence. From this sequence, the middle convolutional modules in the backbone and neck networks with high computational cost and low sensitivity to localization / classification are prioritized. Finally, the first N convolutional modules from the candidate replacement sequence are selected sequentially. The modules were replaced with DSConv modules to construct multiple intermediate improved models with different replacement scales. Ablation experiments were conducted on each of the intermediate improved models to evaluate their detection accuracy and inference speed. During the ablation experiments, the replacement was stopped when the detection accuracy met the preset threshold and the inference speed reached the target, thereby determining the final replacement position and number of DSConv modules. Finally, based on the replacement position and number of DSConv modules, the convolutional modules at corresponding positions in the backbone network and the neck network were replaced with DSConv modules to obtain the preliminary improved backbone network and the preliminary improved neck network.

[0076] S2.1 Embedded MSCA module, specifically: In the initially improved backbone network, an MSCA module is embedded at the end of the residual connection branch of the C2PSA module to obtain the final improved backbone network; and multiple MSCA modules are embedded after the fusion position and / or downsampling transition position of several C3K2 modules in the initially improved neck network to obtain the final improved neck network.

[0077] The MSCA module uses multi-branch deep strip convolution combined with channel attention mechanism to achieve dynamic aggregation of multi-scale contextual information, which can take into account both local details and global receptive field, and is used to improve the expressive ability of small targets in complex backgrounds.

[0078] S3. Based on the final improved neck network, the final improved backbone network, and the detection head, an improved YOLOv11 model is obtained, such as... Figure 3 As shown, Figure 3 A schematic diagram of the framework of the improved YOLOv11 model provided for embodiments of this application.

[0079] The final improved backbone network may include multiple DSConv modules, multiple C3K2 modules, SPPF modules, C2PSA modules, and MSCA modules; the final improved neck network may include multiple upsample modules, multiple C3K2 modules, multiple MSCA modules, and multiple DSConv modules.

[0080] After obtaining the improved YOLOv11 model, the improved YOLOv11 model is trained using the high-resolution wildfire image dataset constructed in step S1.

[0081] S4. Target detection is performed on the flame image to be detected using the improved YOLOv11 model, specifically including: S4.1. Multi-scale feature extraction is performed on the flame image to obtain multi-scale feature maps (shallow feature map, mid-level feature map, and deep feature map): The preprocessed flame image is input into the final improved backbone network. First, it is passed through the first two DSConv modules for primary feature extraction and downsampling to obtain the shallow feature map. Then, the shallow feature map is input into multiple subsequent DSConv modules and multiple C3K2 modules for detail enhancement and downsampling to obtain the mid-level feature map step by step. The mid-level feature map is then passed through the SPPF module for multi-scale pooling fusion, the C2PSA module, and the embedded MSCA module for attention enhancement of mid-to-high-level semantic features, highlighting key information in the flame / smoke region and suppressing background interference such as clouds, fog, and lights. Finally, the backbone network outputs feature maps containing at least three scales: shallow, mid-level, and deep.

[0082] S4.2. The multi-scale feature maps are fused to obtain a multi-scale feature pyramid, including: The deep feature map is upsampled, and then the upsampled deep feature map is horizontally concatenated with the mid-level feature map (using feature concatenation or element-wise addition). This concatenation is then performed to obtain the mid-level fused feature. The mid-level fused feature is further upsampled, and then horizontally concatenated with the shallow feature map. This concatenation is then performed to obtain the shallow fused feature. The deep feature map, mid-level fused feature, and shallow fused feature are then concatenated and integrated using the corresponding C3K2 module, multi-scale context enhancement using the corresponding MSCA module, and convolutional operations and scale alignment using the corresponding DSConv module to obtain the multi-scale feature pyramid.

[0083] S4.3 Input the multi-scale feature pyramid into the detection head to obtain the candidate boxes for each category and the confidence scores corresponding to the candidate boxes, including: The multi-scale feature pyramid is input into the decoupled detection head. The shared convolutional layer, classification branch, and regression branch in the decoupled detection head all use the DSConv module to replace the original standard convolutional module for convolution operations. First, the feature map of each scale in the multi-scale feature pyramid is integrated through a shared convolutional layer containing the DSConv module to obtain integrated features. Then, the integrated features are input in parallel into the classification and regression branches. The classification branch outputs the category score for each location on the feature map at that scale, indicating whether it belongs to flame or smoke. The regression branch outputs the bounding box regression parameters corresponding to each location on the feature map at that scale. Finally, an anchor-free decoding method is used to map the bounding box regression parameters to candidate box coordinates in the image coordinate system. The category score and the preset target score are calculated to obtain the confidence score corresponding to each candidate box, outputting the flame and smoke candidate boxes and their confidence scores at each scale.

[0084] S4.4. Filter each candidate box to obtain the final flame target detection result, including: First, the candidate boxes for flames and smoke at various scales output by the detection head, along with their corresponding confidence levels, are aggregated and sorted in descending order of confidence level. Then, based on a pre-set confidence threshold for the wildfire monitoring scenario, candidate boxes with confidence levels below the pre-set threshold are removed, resulting in a preliminary candidate box set. Next, non-maximum suppression is performed on each category in the preliminary candidate box set, selecting the candidate box with the highest current confidence level as the retained box. The intersection-union ratio (IU) of this box with the remaining candidate boxes of the same category is calculated, and overlapping boxes with IU exceeding a preset threshold are deleted to remove redundant detection boxes. Finally, the final set of flame and smoke detection boxes is obtained. Furthermore, this application can also generate alarm information containing target type, location coordinates, confidence level, and timestamp based on the detection box set, and upload the alarm information to the power transmission line wildfire monitoring platform to complete the output of flame target detection results.

[0085] Alternatively, YOLOv11 in this application can be replaced with YOLOv8, YOLOv10 or other lightweight object detection frameworks. In the new framework, the MSCA module can be embedded in a similar manner as described above, and the corresponding convolutional module can be replaced with the DSConv module to build an improved model suitable for wildfire detection on power transmission lines.

[0086] At the deployment level, this application can deploy the improved YOLOv11 model on edge computing terminals (such as industrial control computers or embedded GPU devices), receive multi-channel camera image streaming through streaming media services, set batch processing strategies according to device computing power, and ensure that the inference speed of nearly 42 frames per second can still be achieved in multi-channel concurrent scenarios, so as to realize near real-time automatic monitoring of wildfires in power transmission line channels.

[0087] To verify the effectiveness of the proposed flame target detection method based on the improved YOLOv11, flame images in the same scene were detected using different target detection models. The detection results are as follows: Figure 4 As shown, where, Figure 4 (a) shows the detection results of the flame target using the Faster R-CNN model; Figure 4 (b) shows the detection results of the flame target using the YOLOv8 model; Figure 4 (c) shows the detection results of the flame target using the YOLOv10 model; Figure 4 (d) shows the detection results of the flame target using the unmodified YOLOv11 model; Figure 4 (e) shows the detection results of the flame target using the improved YOLOv11 model of this application. Figure 4 As can be seen from the comparison of the detection results in (a) to (e), the improved YOLOv11 model proposed in this application has a better detection effect on flame and smoke targets in complex backgrounds, effectively improving the detection accuracy of small targets and reducing the false detection rate.

[0088] In summary, this application provides a flame target detection method and related equipment based on an improved YOLOv11. The method first acquires a flame image; then obtains the framework of a YOLOv11 model, which includes a backbone network, a neck network, and a detection head; calculates the computational cost of each convolutional module in the backbone and neck networks; based on the computational cost, determines the replacement positions and number of convolutional modules to be replaced in the backbone and neck networks; and finally, based on the replacement positions and number of replacements, replaces the corresponding convolutional modules in the backbone and neck networks with DSConv modules, resulting in... The first improved backbone network and the first improved neck network are constructed. A first MSCA module is embedded at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network to obtain a second improved neck network. Based on the second improved backbone network, the second improved neck network and the detection head, an improved YOLOv11 model is obtained. The flame image is input into the improved YOLOv11 model for detection to obtain the flame target detection result. By embedding MSCA modules at the end of the residual connection branches of the C2PSA module in the first improved backbone network, and embedding multiple MSCA modules after the fusion positions and / or downsampling transition positions of several C3K2 modules in the first improved neck network, the ability to extract multi-scale dynamic features of flames / smoke is improved, the ability to capture early small targets is enhanced, the missed detection of small targets is reduced, and the false detection rate is significantly reduced. At the same time, the replacement positions and number of convolutional modules with DSConv modules are determined according to the computational load, which can reduce the computational load of the model and improve the detection efficiency without losing the accuracy of flame feature extraction, thus meeting the engineering requirements of real-time monitoring of wildfires in overhead power line corridors.

[0089] like Figure 5 As shown in the figure, this application also provides a schematic diagram of a flame target detection device based on an improved YOLOv11. This device can implement the above-described method and may include: Image acquisition module 21 is used to acquire flame images; The framework acquisition module 22 is used to acquire the framework of the YOLOv11 model, wherein the framework of the YOLOv11 model includes a backbone network, a neck network, and a detection head. The computational complexity statistics module 23 is used to calculate the computational complexity of each convolutional module in the backbone network and the neck network. The replacement position determination module 24 is used to determine the replacement position and number of convolutional modules to be replaced in the backbone network and the neck network based on the computational load. The first improvement module 25 is used to replace the convolutional modules at corresponding positions in the backbone network and the neck network with DSConv modules according to the replacement position and the replacement quantity, so as to obtain the first improved backbone network and the first improved neck network. The second improvement module 26 is used to embed a first MSCA module at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network, and to embed multiple second MSCA modules after the fusion position and / or downsampling transition position corresponding to several C3K2 modules in the first improved neck network to obtain a second improved neck network. The improved model generation module 27 is used to obtain an improved YOLOv11 model based on the second improved backbone network, the second improved neck network and the detection head; The target detection module 28 is used to input the flame image into the improved YOLOv11 model for detection and obtain the flame target detection result.

[0090] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0091] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned flame target detection method based on the improved YOLOv11. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0092] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0093] Please see Figure 6 , Figure 6 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 to implement the flame target detection method based on the improved YOLOv11 of the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0094] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described flame target detection method based on the improved YOLOv11.

[0095] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0096] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0097] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0098] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0099] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0100] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0102] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0104] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it 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 all or 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 multiple 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 of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A flame target detection method based on an improved YOLOv11, characterized in that, include: Acquire flame images; Obtain the framework of the YOLOv11 model, which includes a backbone network, a neck network, and a detection head; Calculate the computational cost of each convolutional module in the backbone network and the neck network; Based on the computational load, determine the replacement location and number of convolutional modules to be replaced in the backbone network and the neck network; Based on the replacement position and replacement quantity, the convolutional modules at corresponding positions in the backbone network and the neck network are replaced with DSConv modules to obtain the first improved backbone network and the first improved neck network. A first MSCA module is embedded at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion position and / or downsampling transition position of several C3K2 modules in the first improved neck network to obtain a second improved neck network. Based on the second improved backbone network, the second improved neck network, and the detection head, an improved YOLOv11 model is obtained; The flame image is input into the improved YOLOv11 model for detection, and the flame target detection result is obtained.

2. The flame target detection method based on the improved YOLOv11 according to claim 1, characterized in that, The computational cost of calculating each convolutional module in the backbone network and the neck network includes: Obtain the feature map size and number of channels corresponding to each convolutional module in the backbone network and the neck network; The computational cost of each convolutional module in the backbone network and the neck network is calculated based on the feature map size and number of channels corresponding to each convolutional module.

3. The flame target detection method based on the improved YOLOv11 according to claim 1, characterized in that, The step of determining the replacement location and number of convolutional modules to be replaced in the backbone network and the neck network based on the computational load includes: Each convolutional module is sorted in descending order of computational complexity to obtain a candidate replacement sequence; From the candidate replacement sequence, the first N convolutional modules are selected sequentially and replaced with DSConv modules to construct multiple intermediate improved models with different replacement scales, where N is a positive integer greater than zero; Ablation experiments were conducted on the multiple intermediate improved models, and the replacement positions and number of convolutional modules to be replaced in the backbone network and the neck network were determined based on the results of the ablation experiments.

4. The flame target detection method based on the improved YOLOv11 according to claim 1, characterized in that, The step of inputting the flame image into the improved YOLOv11 model for detection to obtain the flame target detection result includes: The second improved backbone network performs multi-scale feature extraction on the flame image to obtain a multi-scale feature map; The multi-scale feature map is input into the second improved neck network, and the second improved neck network fuses the multi-scale feature map to obtain a multi-scale feature pyramid; The multi-scale feature pyramid is input into the detection head, and the classification and regression branches in the detection head are decoupled for prediction to obtain the candidate box for each category and the confidence level of the candidate box. Each candidate box is filtered based on the confidence level corresponding to each category to obtain the flame target detection result.

5. The flame target detection method based on the improved YOLOv11 according to claim 4, characterized in that, The second improved backbone network includes multiple first DSConv modules, multiple first C3K2 modules, SPPF modules, C2PSA modules, and first MSCA modules; The second improved backbone network performs multi-scale feature extraction on the flame image to obtain a multi-scale feature map, including: The second improved backbone network performs primary feature extraction and downsampling on the flame image through multiple first DSConv modules in the first layer to obtain a first-scale feature map; The first-scale feature map is enhanced and downsampled by multiple first DSConv modules and multiple first C3K2 modules at the second level to obtain the second-scale feature map. The second-scale feature map is fused using multi-scale pooling through the SPPF module to obtain the multi-scale pooled fused features. The features obtained after multi-scale pooling fusion are sequentially passed through the C2PSA module and the first MSCA module for feature attention enhancement to obtain a third-scale feature map.

6. The flame target detection method based on the improved YOLOv11 according to claim 5, characterized in that, The second improved neck network includes multiple upsampling modules, multiple second C3K2 modules, multiple second MSCA modules, and multiple second DSConv modules; The step of inputting the multi-scale feature map into the second improved neck network, and the second improved neck network fusing the multi-scale feature map to obtain a multi-scale feature pyramid, includes: The third-scale feature map is upsampled through the first upsampling module, and the upsampled third-scale feature map is laterally connected and feature fused with the second-scale feature map to obtain the second-scale fused feature map. The second scale fusion feature map is upsampled through the second upsampling module, and the upsampled second scale fusion feature map is horizontally connected and feature fused with the first scale feature map to obtain the first scale fusion feature map. The third-scale feature map, the second-scale fused feature map, and the first-scale fused feature map are convolved and integrated through the corresponding second C3K2 module, multi-scale context enhancement is performed through the corresponding second MSCA module, and convolution operation and scale alignment are performed through the corresponding second DSConv module to obtain a multi-scale feature pyramid.

7. The flame target detection method based on the improved YOLOv11 according to claim 6, characterized in that, The step of filtering each candidate box based on the confidence level corresponding to each category to obtain the flame target detection result includes: Sort the candidate boxes for each category in descending order of confidence to obtain a candidate box list; Remove candidate boxes from the candidate box list whose confidence level is lower than the first preset threshold to obtain a preliminary candidate box list; Non-maximum suppression is performed on each category in the initial candidate box list, and the box with the highest confidence is selected as the retained box in turn. Calculate the intersection-union ratio (IUU) of the retained box with the remaining candidate boxes of the same category, and delete candidate boxes whose IUU is greater than a second preset threshold to obtain a deduplicated set of flame detection boxes; Based on the deduplicated set of flame detection frames, the target detection result is obtained.

8. A flame target detection device based on an improved YOLOv11, characterized in that, The device includes: The image acquisition module is used to acquire flame images; The framework acquisition module is used to acquire the framework of the YOLOv11 model, which includes a backbone network, a neck network, and a detection head. The computational complexity statistics module is used to calculate the computational complexity of each convolutional module in the backbone network and the neck network. The replacement position determination module is used to determine the replacement position and number of convolutional modules to be replaced in the backbone network and the neck network based on the computational load. The first improvement module is used to replace the convolutional modules at corresponding positions in the backbone network and the neck network with DSConv modules according to the replacement position and the replacement quantity, so as to obtain the first improved backbone network and the first improved neck network. The second improved module is used to embed a first MSCA module at the end of the residual connection branch of the C2PSA module in the first improved backbone network to obtain a second improved backbone network. Multiple second MSCA modules are embedded after the fusion position and / or downsampling transition position corresponding to several C3K2 modules in the first improved neck network to obtain a second improved neck network. An improved model generation module is used to obtain an improved YOLOv11 model based on the second improved backbone network, the second improved neck network, and the detection head; The target detection module is used to input the flame image into the improved YOLOv11 model for detection and obtain the flame target detection result.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the flame target detection method based on the improved YOLOv11 as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the flame target detection method based on the improved YOLOv11 as described in any one of claims 1 to 7.

Images