Complex interference-oriented large model collaborative robust wildfire identification method and device

Abstract

This invention discloses a robust wildfire identification method and apparatus for complex interference using a small-model collaborative approach, relating to the field of fire early warning technology. The method includes: training a robust wildfire identification model to be trained using training samples to obtain a trained robust wildfire identification model; acquiring an image and text to be identified, and inputting them into the trained robust wildfire identification model; obtaining visual features through the image to be identified, a small-model image encoder, and a large-model image encoder; obtaining text features through the text to be identified and a large-model text encoder; obtaining cross-modal fusion features of visual and verbal matching through a visual-language fusion module; and obtaining the wildfire identification result through the cross-modal fusion features and a fine-grained wildfire target detection network. Using this invention can improve robustness, stability, and recognition accuracy in smoke scenes under varying weather conditions, thereby improving the accuracy of wildfire target detection.

Description

Technical Field

[0001] This invention relates to the field of fire early warning technology, and in particular to a robust wildfire identification method and device for large and small model collaboration in the face of complex interference. Background Technology

[0002] Wildfire identification is a core application of computer vision in the field of public safety, focusing on the early identification and localization of flames and smoke to achieve disaster warning and monitoring. In recent years, the technology has shifted towards deep learning-led visual intelligent wildfire identification. In computer vision task processing, wildfire identification is a typical target detection task.

[0003] In real-world scenarios, surveillance cameras deployed on power transmission line towers inevitably suffer from various external interferences during long-term operation, significantly degrading the quality of the acquired images and posing a major challenge to wildfire detection. These interferences can be broadly categorized into two types: environmental interference and interference from similar targets. Environmental interference includes reduced visibility due to factors such as heavy fog, tower obstruction, low light conditions at night, and sunlight glare. These meteorological and physical conditions not only weaken image clarity and contrast but also blur or partially obscure key features such as wildfire smoke. Interference from similar targets originates from non-fire smoke targets with similar appearance characteristics, such as factory exhaust, residential smoke, and low clouds in the mountains. These targets closely resemble real wildfire smoke in shape, texture, and dynamic features, easily leading to false alarms.

[0004] However, most existing wildfire detection technologies employ lightweight deep learning detection models (such as the YOLO series). While these models offer high real-time performance, they largely rely on the quality and completeness of training data and lack sufficient scene understanding capabilities, thus exhibiting significant limitations in real-world, complex environments. On one hand, existing models cannot effectively handle issues such as image blurring, color degradation, and detail loss under conditions of heavy fog, rain, low light, and occlusion, leading to unstable recognition results and insufficient robustness. On the other hand, due to a lack of understanding of smoke generation mechanisms, environmental background, and semantic context, existing models struggle to reliably distinguish between fine-grained similar targets such as low clouds, fog, industrial exhaust, and residential cooking smoke, often resulting in high false alarm and false negative rates in complex smoke scenarios. Summary of the Invention

[0005] To address the technical problems of unstable recognition results, insufficient robustness, and high false alarm rate in complex smoke scenarios in existing technologies, this invention provides a robust wildfire recognition method and apparatus using a combination of large and small models for complex interference. The technical solution is as follows:

[0006] On the one hand, a robust wildfire identification method for large and small model collaboration oriented to complex interference is provided. This method is implemented by a robust wildfire identification device for large and small model collaboration oriented to complex interference, and includes: S1. Obtain the robust wildfire recognition model to be trained and the training samples; S2. Train the robust wildfire recognition model to be trained based on the training samples to obtain a trained robust wildfire recognition model. S3. Obtain the image to be identified and the text to be identified, and input the image to be identified and the text to be identified into the trained robust wildfire recognition model; The trained robust wildfire recognition model includes a cross-modal feature extraction module and a fine-grained wildfire target detection network. The cross-modal feature extraction module includes a small-model image encoder, a large-model image encoder, a large-model text encoder, and a visual-language fusion module. S4. Obtain visual features through the image to be identified, the small model image encoder, and the large model image encoder; S5. Obtain text features through the text to be identified and the large model text encoder; S6. Input visual features and text features into the visual-language fusion module to obtain cross-modal fusion features for visual and language matching; S7. Wildfire identification results are obtained by using cross-modal fusion features and a fine-grained wildfire target detection network.

[0007] On the other hand, a robust wildfire identification device for large-scale model collaboration under complex interference is provided. This device is applied to the robust wildfire identification method for large-scale model collaboration under complex interference. The device includes: The acquisition unit is used to acquire the robust wildfire recognition model to be trained and the training samples; The training unit is used to train the robust wildfire recognition model to be trained based on the training samples, so as to obtain the trained robust wildfire recognition model. The input unit is used to acquire the image to be identified and the text to be identified, and input the image to be identified and the text to be identified into the trained robust wildfire recognition model; The trained robust wildfire recognition model includes a cross-modal feature extraction module and a fine-grained wildfire target detection network. The cross-modal feature extraction module includes a small-model image encoder, a large-model image encoder, a large-model text encoder, and a visual-language fusion module. The first feature extraction unit is used to obtain visual features through the image to be identified, the small model image encoder, and the large model image encoder; The second feature extraction unit is used to obtain text features through the text to be identified and the large model text encoder; The fusion unit is used to input visual features and text features into the visual-language fusion module to obtain cross-modal fusion features that match visual and linguistic features; The identification unit is used to obtain wildfire identification results by fusing cross-modal features and a fine-grained wildfire target detection network.

[0008] On the other hand, a size-model cooperative robust wildfire identification device for complex interference is provided. The size-model cooperative robust wildfire identification device for complex interference includes: a processor; a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, they implement any of the methods described above for size-model cooperative robust wildfire identification for complex interference.

[0009] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described methods of the size-model cooperative robust wildfire identification method for complex interference.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention combines scene-level knowledge provided by a large model with the fine-grained wildfire recognition capabilities of a small model to construct a multi-layered, multi-modal recognition system resistant to complex interference. This effectively improves the accuracy and robustness of wildfire recognition, especially in dynamic natural environments. Through efficient fusion of visual and linguistic information, adversarial interference removal, and the joint use of a fine-grained target detection network, the model maintains accurate recognition of wildfire smoke targets in real and complex natural environments, thereby significantly improving target detection accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a large-scale model collaborative robust wildfire identification method for complex interference provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of a trained robust wildfire identification model provided in an embodiment of the present invention; Figure 3 This is a flowchart of a visual feature generation process provided by an embodiment of the present invention; Figure 4This is a flowchart of a text feature generation process provided by an embodiment of the present invention; Figure 5 This is a flowchart illustrating a fine-grained wildfire target detection network provided in an embodiment of the present invention; Figure 6 This is a block diagram of a large-scale model collaborative robust wildfire identification device for complex interference provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a large-scale model collaborative robust wildfire identification device for complex interference provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0014] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0015] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0016] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0017] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0018] This invention provides a collaborative robust wildfire detection method using large and small models to handle complex interference. This method can be implemented using a collaborative robust wildfire detection device, which can be a terminal or a server. The core idea of ​​the collaborative robust wildfire detection method (CRWD) proposed in this invention is that a large model provides general scene understanding, while a small model is responsible for fine-grained wildfire detection. Figure 1 The flowchart shown is for a large-scale model collaborative robust wildfire identification method for complex interference. The processing flow of this method may include the following steps:

[0019] S1. Obtain the robust wildfire recognition model to be trained and the training samples.

[0020] The robust wildfire recognition model to be trained includes a small-model image encoder to be updated, a pre-trained large-model image encoder, a pre-trained large-model text encoder, a visual-language fusion module to be updated, an adversarial interference removal network to be updated, and a fine-grained wildfire target detection network to be updated.

[0021] The training samples include sample images, sample text, and ground truth labels. The sample images may or may not include wildfire images. The sample text is the textual description of the sample images. The ground truth labels are the results of whether the sample images include wildfire images.

[0022] S2. Train the robust wildfire recognition model to be trained based on the training samples to obtain the trained robust wildfire recognition model.

[0023] In one feasible implementation, in order to better train the model, an adversarial interference removal network is introduced during model training. After training is completed, the adversarial interference removal network is removed to obtain a well-trained robust wildfire recognition model.

[0024] Optionally, the specific operation of S2 may include the following steps S21-S26: S21. Input the sample images from the training samples into the small model image encoder to be updated and the pre-trained large model image encoder respectively to obtain the first feature and the second feature. Calculate the self-supervised loss function of the small model image encoder based on the feature consistency between the first feature and the second feature. Obtain the predicted visual features based on the first feature and the second feature.

[0025] In one feasible implementation, in order to constrain the semantic discrimination ability of the small model image encoder for foreground and background, this embodiment of the invention proposes to use self-supervised loss to constrain the feature consistency between the small model image encoder and the large model image encoder, as shown in the following equation (1): (1) in, This represents the first feature obtained by the small model image encoder. This represents the second feature obtained by the large model image encoder. Indicates mean square error. This indicates a loss for Dais.

[0026] S22. Input the sample text from the training samples into the pre-trained large model text encoder to obtain the predicted text features.

[0027] In one feasible implementation, the large-model image encoder utilizes a large-scale image dataset for self-supervised pre-training, learning a deep correspondence between image features and text descriptions. This gives the model a powerful zero-shot learning capability. Even without targeted training or fine-tuning on specific downstream tasks, it can still handle and solve a variety of tasks. Although the large-model image encoder performs excellently with zero-shot performance on general natural scene images, its performance will decrease to some extent when facing specialized domains with professional knowledge or handling more complex tasks. Therefore, when directly applying the large-model image encoder for natural scenes to the image wildfire recognition task, the spectral features obtained by the large-model image encoder will inevitably contain false activations.

[0028] S23. Input the predicted visual features and predicted text features into the visual-language fusion module to be updated to obtain the predicted cross-modal fusion features.

[0029] S24. Input the visual-language fusion features into the adversarial interference removal network to be updated. The adversarial interference removal network uses the adversarial loss function to constrain the adversarial discriminator to evaluate the severity of the sample being interfered with by the external environment.

[0030] In one feasible implementation, to address data interference in real-world environments (such as complex weather conditions, obstructions, and cluttered backgrounds), this invention proposes an adversarial interference removal network. Drawing inspiration from generative adversarial networks, the adversarial interference removal network aims to maintain the cross-modal fusion features of visual-language matching, ensuring they contain rich semantic information about wildfires to minimize detection errors and eliminate environmental interference affecting visual-language fusion features. It also possesses domain invariance to remove environmental interference such as fog, clouds, and false smoke. The adversarial interference removal network only participates in the training process and needs to be removed during the testing and practical application phases.

[0031] The adversarial interference removal network consists of a generator, a discriminator, and a gradient reversal layer. The generator G(·) is implemented using a multilayer perceptron to ensure a lightweight model while maintaining the original feature distribution. The gradient reversal layer (GRL) preserves the original features of the data during forward propagation but intentionally conveys the opposite features during backward propagation. The discriminator D(·) determines the degree of interference in the current input data based on cross-modal fusion features. Through multiple rounds of training and game theory, the visual-language fusion module is forced to learn to remove environmental interference information, retaining only the core features of the wildfire.

[0032] The loss is implemented using the standard generative adversarial network loss, as shown in equation (2): (2) in, Represents the true feature distribution, This indicates the characteristics of a normal distribution. This represents the cross-modal fusion features of the input prediction. Expressing expectations, This indicates sampling from the real data distribution. This indicates sampling from the noise distribution.

[0033] This network learns to distinguish between "interference features" and "wildfire features" through adversarial training, effectively eliminating feature interference caused by factors such as fog, clouds, and other non-wildfire smoke, ensuring purer generated features for accurate target identification. The network's design borrows from the discriminator mechanism in generative adversarial networks, effectively constraining the features learned by the model to be true features consistent with the wildfire identification task.

[0034] S25. Input the predicted cross-modal fusion features into the fine-grained wildfire target detection network to be updated to obtain the predicted recognition results. Calculate the multi-class cross-entropy loss function by combining the predicted recognition results with the ground truth labels.

[0035] In one feasible implementation, the fine-grained target detection loss is achieved using a multi-class cross-entropy loss function. It can be expressed as the following formula (3): (3) in, Indicates the number of categories. Indicates the truth label, This indicates the model's prediction results.

[0036] S26. Based on the self-supervised loss function, adversarial loss function, and multi-class cross-entropy loss function, iteratively update the parameters of the small model image encoder to be updated, the visual-language fusion module to be updated, the adversarial interference removal network to be updated, and the fine-grained wildfire target detection network to be updated until the training stopping condition is met, and end the iterative training to obtain the trained robust wildfire recognition model.

[0037] In one feasible implementation, the small model image encoder E, the adversarial interference removal network, and the fine-grained wildfire target detection network S can be updated alternately under a unified target, as shown in equation (4): (4) in, It is the self-supervised loss of the small model image encoder. It is a generator network and discriminator The losses incurred in the fight against the epidemic. and All are implemented using multilayer perceptrons to ensure lightweight models. It is a fine-grained wildfire target detection network Target detection loss.

[0038] It should be noted that in the above steps, the structure and generation mechanism of the small model image encoder to be updated, the pre-trained large model image encoder, the pre-trained large model text encoder, the visual-language fusion module to be updated, and the fine-grained wildfire target detection network to be updated are basically the same as those used in actual applications. Please refer to the corresponding descriptions in S4-S8 below. The embodiments of the present invention will not be elaborated here.

[0039] S3. Obtain the image and text to be identified, and input the image and text to be identified into the trained robust wildfire recognition model.

[0040] The trained robust wildfire recognition model includes a cross-modal feature extraction module and a fine-grained wildfire target detection network. The cross-modal feature extraction module includes a small-model image encoder, a large-model image encoder, a large-model text encoder, and a visual-language fusion module.

[0041] One feasible implementation method is, for example Figure 2 As shown, the core idea of ​​this invention is that a large model provides general scene understanding, while a small model is responsible for fine-grained wildfire identification. The robust wildfire identification model structure includes two stages: cross-modal feature extraction and fine-grained wildfire identification.

[0042] (1) Cross-modal feature extraction stage First, a large-scale image encoder and a large-scale text encoder are used to extract high-level semantic features from the input images and their auxiliary text descriptions to construct an overall understanding of the current scene (such as weather conditions, terrain features, and background structure). Simultaneously, a small-scale image encoder specifically designed for wildfire identification is used to extract fine-grained visual features of smoke and flames from the images. Then, a visual-language fusion module is used to integrate the general knowledge of the large-scale model with the task-related fine-grained features, generating a comprehensive and robust multimodal representation.

[0043] (2) Fine-grained wildfire identification stage The fused cross-modal features are first input into an adversarial interference removal network. By introducing a discriminator mechanism from an adversarial generative network, interference factors such as fog, clouds, and fake smoke in the real environment are effectively removed, making the features more closely resemble the distribution of wildfires under conditions without external environmental interference. Simultaneously, the cross-modal fused features are input into a fine-grained wildfire target detection network to locate and classify wildfire smoke targets, while excluding other visually similar interference targets, thus obtaining the final wildfire identification result. To ensure the coherence of feature learning, the feature encoder and the fine-grained wildfire target detection network maintain the same lightweight network structure design.

[0044] Through the synergistic effect of the above stages, the embodiments of the present invention achieve an organic combination of large model knowledge understanding and small model target detection capabilities, significantly improving the robustness and accuracy of wildfire identification.

[0045] S4. Visual features are obtained through the image to be identified, the small model image encoder, and the large model image encoder.

[0046] In one feasible implementation, to transfer image-level features from a large-model image encoder to a dense wildfire identification task, the feature encoder employs the SAN reparameterization technique to perform visual feature fusion from multiple dimensions, extracting spectral features and semantic category predictions from the input image. This embodiment of the invention selects ViT as the backbone network to construct the feature encoder.

[0047] Optionally, the small model image encoder can be an 8-layer ViT-B (Vision Transformer Base) feature encoder, and the large model image encoder can be a 12-layer ViT-B CLIP model.

[0048] The specific operation of S4 may include the following steps S41-S44: S41. Divide the image to be recognized into 8×8 image blocks and map them into visual tokens through a linear mapping layer.

[0049] One feasible implementation method is, for example Figure 3 As shown, the input image is first divided into "8×8" image blocks and mapped to visual tokens through a linear embedding layer to preserve local visual features.

[0050] S42. Link the visual token with the query token.

[0051] In one feasible implementation, these visual tokens are concatenated with N learnable query tokens and input together into a subsequent encoding layer for visual feature extraction.

[0052] S43. Input the concatenated features into the ViT-B feature encoder and the ViT-B CLIP model respectively, and fuse the {stem, 1, 2} layer features of the ViT-B feature encoder (i.e., the first 3 layers of the ViT-B feature encoder) with the {stem, 3, 6} layer features of the ViT-B CLIP model (i.e., the 1st, 4th, and 7th layers of the ViT-B CLIP model) to obtain the smoke and flame visual features output by the ViT-B feature encoder and the image scene features output by the ViT-B CLIP model.

[0053] In one feasible implementation, considering the differences in feature dimension and spatial resolution between the large-model image encoder and the small-model image encoder, "1×1" convolution and bilinear interpolation are used to adaptively adjust the spatial resolution of the input image. In the shallow network, the features of the first three layers {stem, 1, 2} of the small-model image encoder network are spatially adapted and then fused with the low-scale visual texture information and mid-scale regional spatial information of the {stem, 3, 6} layers of the large-model image encoder. Specifically, the high-dimensional self-attention features of the deepest layer of the large-model image encoder are fused with the deepest features of the small-model image encoder by element-wise addition, constraining the consistency of their semantic representation to ensure consistent transformation of spatial features from local to global, while ensuring the independence of feature learning in the remote sensing image wildfire recognition task, thereby enhancing the model's ability to learn local spatial transformations.

[0054] S44. Combine the visual features of smoke and flame with the image scene features to obtain the visual features.

[0055] In one feasible implementation, the visual features can be obtained by simply concatenating the outputs of the large-model image encoder and the small-model image encoder. .

[0056] In this embodiment of the invention, the problem of wildfire identification in complex environments is solved by combining the scene-level knowledge of a large model with the fine-grained wildfire identification capabilities of a small model. The large model provides scene-level knowledge about the environmental background, weather, and terrain, ensuring a comprehensive understanding of changing scenarios; while the small model focuses on fine-grained feature extraction of the wildfire target, ensuring accurate identification of wildfires in complex backgrounds. This scheme, through the collaborative work of the two, can handle complex background information while focusing on the characteristics of the wildfire target itself during the identification process, thereby improving the accuracy and robustness of identification.

[0057] S5. Obtain text features through the text to be identified and the large model text encoder.

[0058] Alternatively, the specific operation of S5 can be as follows: The text to be identified is input into the large model text encoder to obtain the wildfire semantic features and background semantic features respectively.

[0059] One feasible implementation method is, for example Figure 4 The flowchart shown illustrates the text feature generation process. To further mitigate feature mismatch issues caused by domain differences, this invention generates categorical language descriptions of the current image from multiple perspectives, including image scene, smoke targets, power scene, meteorological conditions, and terrain. Using text encoders from mainstream basic models such as CLIP, GPT, and Germini, image scene descriptions and macroscopic semantic backgrounds are extracted. This information helps the model better understand image content, especially in recognizing wildfire targets against complex backgrounds. In this embodiment, the large-scale text encoder can employ the CLIP text encoder.

[0060] CRWD focuses on exploring how to transform text descriptions into conceptual supervised information about wildfires and their context to optimize wildfire recognition networks. Specifically, CRWD uses a pre-trained large-model text encoder to conceptualize text descriptions into global semantic guidance about wildfires and their context (i.e., wildfire semantic features and context semantic features), which are then labeled as follows: and .

[0061] S6. Input visual features and text features into the visual-language fusion module to obtain cross-modal fusion features for visual and language matching.

[0062] In one feasible implementation, the visual-language fusion module integrates conceptualized text features. and With visual features Integration is used to enhance consistency between spatial activation and semantic categories within the region.

[0063] Optionally, the specific operation of S6 may include the following steps S61-S63: S61. Evaluate the first similarity between visual features and wildfire semantic features, and the second similarity between visual features and background semantic features, respectively.

[0064] In one feasible implementation, the visual-language fusion module first evaluates the similarity between visual semantic category prediction and category text features to further clarify the semantic category of the current image, as shown in formula (5): , (5) in, and These represent the dot product operator and the L2 norm, respectively. Characterizing the first similarity, Characterize the second similarity.

[0065] S62. Determine whether the image region contains wildfire pixels based on the first similarity and the second similarity. Based on the determination result, select wildfire semantic features and visual features to fuse to obtain wildfire fused features, and select background semantic features and visual features to fuse to obtain background fused features.

[0066] In one feasible implementation, the visual language fusion module uses the high-level semantic information of clouds and background obtained by the CLIP text encoder to divide the input image into regions, obtaining the spectral features of the clouded region as shown in Equation (6) and the spectral features of the background region as shown in Equation (7). (6) (7) in, This represents an indicator function that determines the global semantic category of the current image based on similarity weights. If... If the probability is positive, it indicates that the current region is very likely to contain wildfire pixels, which can be used to guide wildfire spectral feature analysis; otherwise, it means that the current region mainly consists of background. In other words, the indicator function selects the value with the higher probability as the current value.

[0067] S63. The wildfire fusion feature is fused with the background fusion feature to obtain the fusion feature of visual features and text features.

[0068] In one feasible implementation, the foreground should retain as much original visual information as possible, and the background area should be averaged to reduce the interference of complex background on the wildfire identification task, as described in formula (8): (8) in, As a balancing factor, it can be determined through human experience.

[0069] S7. Wildfire identification results are obtained by using cross-modal fusion features and a fine-grained wildfire target detection network.

[0070] In one feasible implementation, to address the problem of false alarms caused by the high similarity between factory exhaust, residential smoke, and low-lying clouds in the mountains and wildfire smoke in appearance, texture, and diffusion patterns, a fine-grained wildfire target detection network for wildfire identification is designed. This network, based on a conventional target detection framework, further introduces multi-dimensional discrimination mechanisms such as morphological representation, texture modeling, and scene semantics, thereby establishing clearer category boundaries between similar interfering targets and improving the reliability and accuracy of wildfire identification.

[0071] Optionally, the fine-grained wildfire target detection network includes a feature extraction network with a pyramid structure, a region proposal network, a fine-grained wildfire discriminator, and a fine-grained feature fusion head. The fine-grained wildfire discriminator includes three branches: a morphological branch, a texture branch, and a contextual semantic branch.

[0072] The specific operation of S7 may include the following steps S71-S74: S71. Input the cross-modal fusion features into the feature extraction network, and use the feature extraction network to perform multi-scale feature encoding on the cross-modal fusion features to obtain multi-level feature representations.

[0073] One feasible implementation method is, for example Figure 5 As shown, a feature extraction network with a pyramid structure (such as ResNet+FPN or Swin Transformer+FPN) is used to perform multi-scale feature encoding on the input cross-modal fusion features to obtain multi-level feature representations from low-level edge textures to high-level semantic regions, as shown in Equation (9): (9) in, This represents a feature extraction network with a pyramid structure. Representing multi-scale features, Indicates the number of network layers. This represents the cross-modal fusion features of the input.

[0074] S72. Input the multi-level feature representation into the region proposal network to obtain wildfire candidate regions.

[0075] In one feasible implementation, preliminary candidate boxes for suspected smoke regions are generated using a Region Proposal Network (RPN) framework. This step allows for the initial screening of all regions in the entire image that are highly likely to contain smoke targets. However, this stage does not distinguish whether the smoke is from a wildfire, thus providing candidate region input for subsequent fine-grained discrimination, as shown in equation (10).

[0076] (10) in, This represents the set of candidate bounding boxes detected by the network. This indicates the number of candidate bounding boxes.

[0077] S73. Input the candidate wildfire regions into the three branches of the fine-grained wildfire discriminator. Obtain geometric morphological information through the morphological branch, obtain texture features through the texture branch, and obtain category discrimination through the contextual semantic branch.

[0078] In one feasible implementation, after initial screening of candidate regions, this invention adds a fine-grained wildfire discriminator on top of the target detection backbone to achieve accurate differentiation between wildfire smoke and visually similar targets. This discriminator constructs morphological, textural, and contextual semantic branches based on feature blocks of candidate regions. Finally, a fine-grained feature fusion head is used to fuse the features from these three branches, forming a fine-grained category discrimination result.

[0079] 1) The ShapeNeck branch extracts geometric morphological information such as smoke contour, edge sharpness, and diffusion direction by performing convolutional encoding on the spatial structure of the candidate region, as shown in the following equation (11): (11) Since wildfire smoke typically exhibits a vertical diffusion structure from bottom to top, and its boundaries are irregularly shaped due to thermal flow disturbances, morphological branching can effectively capture the differences in diffusion path and contour stability between wildfire smoke and targets such as low clouds and factory smoke.

[0080] 2) Texture feature branches Texture features within the candidate region are extracted using channel attention mechanisms and local frequency domain filtering, including smoke particle size, density gradient, and local dynamic perturbation patterns, as shown in Equation (12): (12) Wildfire smoke, influenced by thermal convection, typically exhibits a coarser internal texture and a distinct turbulent structure, while factory exhaust and residential cooking smoke show more uniform, banded textures; low clouds, on the other hand, are characterized by a loose structure and smooth texture. Therefore, texture feature branches can further enhance the model's ability to distinguish between different smoke sources.

[0081] 3) Contextual semantic branches The candidate regions are classified, and the category set includes wildfire smoke, factory exhaust, residential cooking smoke, low clouds in the mountains and other background categories, as shown in the following formula (13): (13) By adopting this explicit multi-class modeling approach, the network no longer simply treats all non-wildfire smoke as background, but learns the independent distribution of various visually similar interference targets. This allows it to learn more detailed category boundary features when making classification decisions, significantly reducing the false detection rate.

[0082] S74. Through a multilayer perceptron with a fine-grained feature fusion head, geometric morphology information, texture features, and category discrimination are fused to obtain the wildfire category prediction probability.

[0083] In one feasible implementation, the FineHead fine-grained feature fusion head fuses the output features of the above three branches using a simple multilayer perceptron, and jointly determines the source of smoke from three dimensions: morphology, texture, and semantics, and finally obtains the fine-grained wildfire category prediction probability, thereby improving the accuracy of wildfire identification results, as shown in the following formula (14): (14) here, These represent the category determination result and the bounding box, respectively.

[0084] In this step, by designing three independent branches—morphology, texture, and contextual semantics—this invention can effectively extract fine-grained differences in morphology, texture, and diffusion patterns between wildfire smoke and similar interfering targets. Finally, these three types of features are fused using a fine-grained feature fusion head, thereby achieving accurate differentiation between wildfire targets and visually similar interfering targets, significantly improving the accuracy and robustness of wildfire identification.

[0085] To address the technical shortcomings of existing detection models that significantly reduce recognition performance under conditions of heavy fog, rain, low light at night, and occlusion, this invention introduces the scene semantic understanding capabilities of a large model to provide a small model with high-order semantic priors such as weather, lighting, and occlusion, thereby improving the robustness and stability of the overall system under varying weather conditions. To overcome the technical bottleneck of existing small models' inability to effectively distinguish between similar targets such as low-level fog, factory exhaust, and residential cooking smoke, this invention constructs a large-small model collaboration mechanism. This mechanism allows the large model to utilize its general knowledge and strong semantic reasoning capabilities to assist the small model in achieving fine-grained feature understanding, significantly improving recognition accuracy in similar smoky scenes.

[0086] This invention combines scene-level knowledge provided by a large model with the fine-grained wildfire recognition capabilities of a small model to construct a multi-layered, multi-modal recognition system resistant to complex interference. This effectively improves the accuracy and robustness of wildfire recognition, especially in dynamic natural environments. Through efficient fusion of visual and linguistic information, adversarial interference removal, and the joint use of a fine-grained target detection network, the model maintains accurate recognition of wildfire smoke targets in real and complex natural environments, thereby significantly improving target detection accuracy.

[0087] Figure 6 This is a block diagram of a large-scale model collaborative robust wildfire identification device for complex interference, provided by an embodiment of the present invention. This device is used for a large-scale model collaborative robust wildfire identification method for complex interference. (Refer to...) Figure 6 The device includes an acquisition unit 610, a training unit 620, an input unit 630, a first feature extraction unit 640, a second feature extraction unit 650, a fusion unit 660, and a recognition unit 670. Wherein:

[0088] The acquisition unit 610 is used to acquire the robust wildfire recognition model to be trained and the training samples; Training unit 620 is used to train the robust wildfire recognition model to be trained based on training samples, so as to obtain a trained robust wildfire recognition model. Input unit 630 is used to acquire the image to be identified and the text to be identified, and input the image to be identified and the text to be identified into a trained robust wildfire recognition model. The trained robust wildfire recognition model includes a cross-modal feature extraction module and a fine-grained wildfire target detection network. The cross-modal feature extraction module includes a small-model image encoder, a large-model image encoder, a large-model text encoder, and a visual-language fusion module. The first feature extraction unit 640 is used to obtain visual features through the image to be identified, the small model image encoder, and the large model image encoder; The second feature extraction unit 650 is used to obtain text features through the text to be identified and the large model text encoder; The fusion unit 660 is used to input visual features and text features into the visual-language fusion module to obtain cross-modal fusion features that match visual and language features; The identification unit 670 is used to obtain wildfire identification results by fusing cross-modal features and a fine-grained wildfire target detection network.

[0089] Figure 7 This is a schematic diagram of the structure of a large-scale model collaborative robust wildfire identification device for complex interference provided in an embodiment of the present invention, as shown below. Figure 7As shown, a large-scale model-coordinated robust wildfire identification device for complex interference can include the above-mentioned... Figure 6 The illustrated size-model cooperative robust wildfire identification device for complex interference is shown. Optionally, the size-model cooperative robust wildfire identification device 710 for complex interference may include a first processor 2001.

[0090] Optionally, the large-scale model collaborative robust wildfire identification device 710 for complex interference may also include a memory 2002 and a transceiver 2003.

[0091] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0092] The following is combined Figure 7 A detailed introduction to the various components of the 710 large-scale model collaborative robust wildfire identification device for complex interference is provided below: The first processor 2001 is the control center of the large-scale model collaborative robust wildfire identification device 710 for complex interference. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0093] Optionally, the first processor 2001 can perform various functions of the size model collaborative robust wildfire identification device 710 for complex interference by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0094] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 7 CPU0 and CPU1 are shown in the diagram.

[0095] In a specific implementation, as one example, the large-scale model collaborative robust wildfire identification device 710 for complex interference can also include multiple processors, for example... Figure 7The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0096] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0097] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or exist independently, and may be integrated with the interface circuit of the robust wildfire identification device 710 through a size model for complex interference. Figure 7 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0098] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0099] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 7 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0100] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be used in conjunction with the interface circuit of the robust wildfire identification device 710 through a size model for complex interference. Figure 7 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0101] It should be noted that, Figure 7 The structure of the size-model cooperative robust wildfire identification device 710 for complex interference shown in the figure does not constitute a limitation on the router. Actual size-model cooperative robust wildfire identification devices for complex interference may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0102] Furthermore, the technical effectiveness of the size model collaborative robust wildfire identification device 710 for complex interference can be referred to the technical effectiveness of the size model collaborative robust wildfire identification method for complex interference described in the above method embodiments, and will not be repeated here.

[0103] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or it may be any conventional processor, etc.

[0104] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0105] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0106] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0107] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0108] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0109] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. 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 implementations should not be considered beyond the scope of this invention.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0111] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units 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 device, 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 devices or units may be electrical, mechanical, or other forms.

[0112] The units described 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.

[0113] In addition, the functional units in the various embodiments of the present invention 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.

[0114] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, 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 invention. 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.

[0115] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robust wildfire identification method using a large-scale model collaborative approach to complex interference, characterized in that, The method includes: S1. Obtain the robust wildfire recognition model to be trained and the training samples; S2. Train the robust wildfire recognition model to be trained based on the training samples to obtain a trained robust wildfire recognition model. S3. Obtain the image to be identified and the text to be identified, and input the image to be identified and the text to be identified into the trained robust wildfire recognition model; The trained robust wildfire recognition model includes a cross-modal feature extraction module and a fine-grained wildfire target detection network. The cross-modal feature extraction module includes a small-model image encoder, a large-model image encoder, a large-model text encoder, and a visual-language fusion module. S4. Obtain visual features through the image to be identified, the small model image encoder, and the large model image encoder; S5. Obtain text features through the text to be recognized and the large model text encoder; S6. Input visual features and text features into the visual-language fusion module to obtain cross-modal fusion features for visual and language matching; S7. Wildfire identification results are obtained by using cross-modal fusion features and a fine-grained wildfire target detection network.

2. The method for robust wildfire identification using a large-scale model in response to complex interference as described in claim 1, characterized in that, The robust wildfire recognition model to be trained includes a small model image encoder to be updated, a pre-trained large model image encoder, a pre-trained large model text encoder, a visual-language fusion module to be updated, an adversarial interference removal network to be updated, and a fine-grained wildfire target detection network to be updated. The step S2 trains the robust wildfire recognition model to be trained based on the training samples to obtain a trained robust wildfire recognition model, including: S21. Input the sample images from the training samples into the small model image encoder to be updated and the pre-trained large model image encoder respectively to obtain the first feature and the second feature. Calculate the self-supervised loss function of the small model image encoder based on the feature consistency between the first feature and the second feature. Obtain the predicted visual features based on the first feature and the second feature. S22. Input the sample text from the training samples into the pre-trained large model text encoder to obtain the predicted text features; S23. Input the predicted visual features and predicted text features into the visual-language fusion module to be updated to obtain the predicted cross-modal fusion features. S24. Input the visual language fusion features into the adversarial interference removal network to be updated. The adversarial interference removal network uses the adversarial loss function to constrain the adversarial discriminator to evaluate the severity of the sample being interfered with by the external environment. S25. Input the predicted cross-modal fusion features into the fine-grained wildfire target detection network to be updated to obtain the prediction and recognition results. Calculate the multi-class cross-entropy loss function by combining the prediction and recognition results with the ground truth labels. S26. Based on the self-supervised loss function, adversarial loss function, and multi-class cross-entropy loss function, iteratively update the parameters of the small model image encoder to be updated, the visual-language fusion module to be updated, the adversarial interference removal network to be updated, and the fine-grained wildfire target detection network to be updated until the training stopping condition is met, and end the iterative training to obtain the trained robust wildfire recognition model.

3. The method for robust wildfire identification using a large-scale model in response to complex interference as described in claim 1, characterized in that, The small model image encoder is an 8-layer ViT-B feature encoder, and the large model image encoder is a 12-layer ViT-B CLIP model. The step S4 obtains visual features through the image to be identified, the small model image encoder, and the large model image encoder, including: S41. Divide the image to be recognized into 8×8 image blocks and map them into visual tokens through a linear mapping layer; S42. Connect the visual token and the query token together; S43. Input the concatenated features into the ViT-B feature encoder and the ViT-B CLIP model respectively, and fuse the {stem, 1, 2} layer features of the ViT-B feature encoder with the {stem, 3, 6} layer features of the ViT-B CLIP model to obtain the smoke and flame visual features output by the ViT-B feature encoder and the image scene features output by the ViT-B CLIP model. S44. The visual features of the smoke and flame and the image scene features are spliced ​​together to obtain visual features.

4. The method for robust wildfire identification using a large-scale model in response to complex interference as described in claim 3, characterized in that, The S5 process obtains text features through the text to be identified and the large-model text encoder, including: The text to be identified is input into the large model text encoder to obtain the wildfire semantic features and background semantic features respectively.

5. The method for robust wildfire identification using a large-scale model in response to complex interference as described in claim 4, characterized in that, The step S6 inputs visual and textual features into the visual-language fusion module to obtain cross-modal fusion features for visual and linguistic matching, including: S61. Evaluate the first similarity between visual features and wildfire semantic features, and the second similarity between visual features and background semantic features, respectively. S62. Determine whether the image region contains wildfire pixels based on the first similarity and the second similarity. Based on the determination result, select wildfire semantic features and visual features to fuse to obtain wildfire fused features, and select background semantic features and visual features to fuse to obtain background fused features. S63. The wildfire fusion feature is fused with the background fusion feature to obtain the cross-modal fusion feature for visual and text matching.

6. The method for robust wildfire identification using a large-scale model in response to complex interference as described in claim 1, characterized in that, The fine-grained wildfire target detection network includes a feature extraction network with a pyramid structure, a region proposal network, a fine-grained wildfire discriminator, and a fine-grained feature fusion head. The fine-grained wildfire discriminator includes three branches: a morphological branch, a texture branch, and a contextual semantic branch. The S7 obtains wildfire identification results through cross-modal feature fusion and a fine-grained wildfire target detection network, including: S71. Input the cross-modal fusion features into the feature extraction network, and use the feature extraction network to perform multi-scale feature encoding on the cross-modal fusion features to obtain multi-level feature representations; S72. Input the multi-level feature representation into the region proposal network to obtain wildfire candidate regions; S73. Input the wildfire candidate regions into the three branches of the fine-grained wildfire discrimination head respectively. Obtain geometric morphological information through the morphological branch, obtain texture features through the texture branch, and obtain category discrimination through the contextual semantic branch. S74. Through a multilayer perceptron with a fine-grained feature fusion head, geometric morphology information, texture features, and category discrimination are fused to obtain the wildfire category prediction probability.

7. A robust wildfire identification device for large-scale model collaboration under complex interference, wherein the robust wildfire identification device for large-scale model collaboration under complex interference is used to implement the robust wildfire identification method for large-scale model collaboration under complex interference as described in any one of claims 1-6, characterized in that, The device includes: The acquisition unit is used to acquire the robust wildfire recognition model to be trained and the training samples; The training unit is used to train the robust wildfire recognition model to be trained based on the training samples, so as to obtain a trained robust wildfire recognition model. The input unit is used to acquire the image to be identified and the text to be identified, and input the image to be identified and the text to be identified into the trained robust wildfire recognition model. The trained robust wildfire recognition model includes a cross-modal feature extraction module and a fine-grained wildfire target detection network. The cross-modal feature extraction module includes a small-model image encoder, a large-model image encoder, a large-model text encoder, and a visual-language fusion module. The first feature extraction unit is used to obtain visual features through the image to be identified, the small model image encoder, and the large model image encoder; The second feature extraction unit is used to obtain text features through the text to be identified and the large model text encoder; The fusion unit is used to input visual features and text features into the visual-language fusion module to obtain cross-modal fusion features that match visual and linguistic features; The identification unit is used to obtain wildfire identification results by fusing cross-modal features and a fine-grained wildfire target detection network.

8. The large-scale model collaborative robust wildfire identification device for complex interference as described in claim 7, characterized in that, The robust wildfire recognition model to be trained includes a small model image encoder to be updated, a pre-trained large model image encoder, a pre-trained large model text encoder, a visual-language fusion module to be updated, an adversarial interference removal network to be updated, and a fine-grained wildfire target detection network to be updated. The training unit is used for: S21. Input the sample images from the training samples into the small model image encoder to be updated and the pre-trained large model image encoder respectively to obtain the first feature and the second feature. Calculate the self-supervised loss function of the small model image encoder based on the feature consistency between the first feature and the second feature. Obtain the predicted visual features based on the first feature and the second feature. S22. Input the sample text from the training samples into the pre-trained large model text encoder to obtain the predicted text features; S23. Input the predicted visual features and predicted text features into the visual-language fusion module to be updated to obtain the predicted cross-modal fusion features. S24. Input the visual language fusion features into the adversarial interference removal network to be updated. The adversarial interference removal network uses the adversarial loss function to constrain the adversarial discriminator to evaluate the severity of the sample being interfered with by the external environment. S25. Input the predicted cross-modal fusion features into the fine-grained wildfire target detection network to be updated to obtain the prediction and recognition results. Calculate the multi-class cross-entropy loss function by combining the prediction and recognition results with the ground truth labels. S26. Based on the self-supervised loss function, adversarial loss function, and multi-class cross-entropy loss function, iteratively update the parameters of the small model image encoder to be updated, the visual-language fusion module to be updated, the adversarial interference removal network to be updated, and the fine-grained wildfire target detection network to be updated until the training stopping condition is met, and end the iterative training to obtain the trained robust wildfire recognition model.

9. A robust wildfire identification device for large and small model collaboration accommodating complex interference, characterized in that, The large-scale model collaborative robust wildfire identification device for complex interference includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 6.

Images