Forest fire image detection method based on semi-supervised learning

Abstract

The invention relates to the field of forest fire detection, in particular to a forest fire image detection method based on semi-supervised learning, and the method comprises the steps: constructing a teacher-student model as an initial model, and carrying out the initialization and preheating training of the model; introducing a mixed attention mechanism to the model, and respectively building a space attention module and a channel attention module to enhance feature extraction; setting a pseudo label reliability evaluation module based on the teacher model, and generating and evaluating pseudo label reliability; performing student model training based on a separation strategy; and performing experimental evaluation verification. According to the method, the detection precision and robustness of the model under the condition of a small amount of labeled data are remarkably improved, particularly the recognition capability of small-scale fire points and fuzzy smoke is improved, and a core technical support is provided for constructing a real-time, accurate and efficient intelligent forest fire monitoring and early warning system.

Description

Technical Field

[0001] This invention relates to the field of forest fire detection, specifically a forest fire image detection method based on semi-supervised learning. Background Technology

[0002] Against the backdrop of escalating global climate change and frequent extreme weather events, forest fire prevention and early monitoring are of great significance for protecting ecological security, people's lives and property, and social stability. With the popularization of remote sensing and drone technologies, computer vision-based automatic fire detection technology has become a key means to achieve the prevention and control goals of "early detection, early suppression, and early extinguishment."

[0003] In recent years, while deep learning-based fire image detection methods have made some progress, they still face the following key challenges: First, existing supervised learning methods heavily rely on large amounts of high-quality labeled data, while forest fire image samples are scarce, labeling costs are high, and the uneven distribution of samples is prominent, resulting in insufficient generalization ability of the model in complex real-world scenarios. Second, most methods lack a reliable evaluation mechanism for the quality of pseudo-labels when directly applying them for self-training, which easily introduces noisy labels, reducing the model's convergence stability and detection accuracy. Third, traditional semi-supervised methods have limited efficiency in utilizing unlabeled data and struggle to effectively distinguish between high-confidence difficult samples and low-confidence noisy samples, limiting the model's recognition performance for difficult cases such as small target fire points and blurred smoke. Fourth, existing methods mostly use a single threshold or fixed strategy for pseudo-label selection and allocation, which is difficult to adapt to the dynamic feature changes of multi-scale and multi-morphological targets in forest fire images, limiting the robustness and practicality of the detection system in real-world environments. Summary of the Invention

[0004] The purpose of this invention is to provide a forest fire image detection method based on semi-supervised learning. By constructing a pseudo-label reliability assessment module and a separate sample allocation strategy, this method accurately selects high-quality pseudo-labels and optimizes their learning process while effectively utilizing massive amounts of unlabeled image data. This method can significantly improve the detection accuracy and robustness of the model under conditions of limited labeled data, especially its ability to identify small-scale fire spots and blurred smoke, providing core technical support for building a real-time, accurate, and efficient intelligent forest fire monitoring and early warning system.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A forest fire image detection method based on semi-supervised learning, the method includes:

[0007] S100. Construct a teacher-student model as the initial model, and perform model initialization and warm-up training;

[0008] S200. A hybrid attention mechanism is introduced into the model, with spatial attention module and channel attention module built separately to enhance feature extraction;

[0009] S300: Based on the teacher model, a pseudo-label reliability assessment module is set up to generate and assess the reliability of pseudo-labels;

[0010] S400, Student model training based on a split strategy;

[0011] S500, conduct experimental evaluation and verification.

[0012] Preferably, model initialization and preheating training are performed in S100. This step lays the foundation for the entire semi-supervised training process, with the goal of obtaining an initial model with preliminary fire detection capabilities, including:

[0013] S101. Prepare and process data, acquire forest fire images, and construct a dataset;

[0014] The dataset consists of two parts: labeled data and unlabeled data; the labeled data is limited in quantity and contains precisely labeled "flame" and "smoke" bounding boxes; the unlabeled data is much larger than the labeled data and has no manual annotations.

[0015] S102. Obtain labeled data and perform fully supervised training on the one-stage object detector using a fully convolutional one-stage object detection algorithm.

[0016] During training, a standard detection loss function is used, in which the focus loss function can be used to learn the visual features of the bright red flames and the grayish-white texture of the smoke, and the generalized intersection and union loss function is used to optimize the bounding box regression to accurately locate the fire target.

[0017] S103. Initialize the teacher-student model. After fully supervised training, copy the trained model parameters. One copy is used for the student model, which is actively updated via gradient descent during subsequent training. The other copy is used for the teacher model, whose parameters are initially identical to those of the student model, but are not directly updated via gradient descent to maintain the stability of the generated pseudo-labels. Specifically, this includes:

[0018] The teacher-student models collaborate alternately to utilize unlabeled data. In each iteration, the teacher model first infers from unlabeled forest fire images. To ensure prediction stability, the input image undergoes weak enhancement operations (e.g., random flipping, scaling, slight color perturbation), and the teacher model's output bounding boxes and categories are used as pseudo-labels. Next, the student model receives a strongly enhanced version of the same image (including random cropping, rotation, blurring, strong color perturbation, etc.) and is trained under the guidance of pseudo-labels. During training, the student model's total loss function is: ;

[0019] in, This represents the supervised loss for labeled samples, including classification loss and bounding box regression loss. The pseudo-label loss, representing unlabeled samples, is used to guide the student model to learn on unlabeled images. This represents the balance coefficient, used to adjust the relative importance of supervised and unsupervised losses, and is set to 1 in this invention;

[0020] To ensure the long-term stability of the teacher model output, the model parameters are not updated directly through backpropagation, but rather iterated through the exponential moving average of the student model parameters: ;

[0021] in, and These represent parameters for teachers and students, respectively. This represents the attenuation coefficient (usually set to 0.999).

[0022] This strategy ensures that the teacher model's predictions are not affected by the instantaneous fluctuations of the student model, thus providing stability for pseudo-label generation.

[0023] Preferably, a hybrid attention mechanism is introduced into the model in S200. This step aims to improve the model's ability to extract features from "small-scale fire points" and "blurred smoke," thereby improving the quality of subsequent pseudo-labels from the source. This includes:

[0024] S201. Obtain the input feature map The input feature map is then normalized.

[0025] S202. Generate a shared query matrix through linear transformation. Bond matrix The specific formula is as follows: , ;

[0026] in, and Represents the shared linear transformation weight matrix;

[0027] S203, Construct a spatial attention module, and spatial value layer Projecting each onto a lower-dimensional space of dimension p, we obtain... and The spatial attention map is then calculated according to the formula. ;

[0028] in, , and These represent the shared query, the projected key matrix, and the spatial value layer, respectively. This represents the dimension of each vector. The projection weights represent the spatial values;

[0029] S204. Build the channel attention module. This module mainly captures the dependencies between channels through dot product operations along the channel dimension. Similar to the spatial attention module, it uses the same... and To learn complementary features and obtain channel value layers Its dimensions are The channel attention map is then calculated using the formula: ;

[0030] in, , and These represent the channel value layer, shared query, and shared key matrix, respectively. Indicates the projected weights of the channel values;

[0031] S205. The outputs of the two attention modules are added together and further processed through a convolutional layer to obtain an enhanced feature representation; the final output of the hybrid attention module is: ;

[0032] in, and These represent spatial and channel attention maps, respectively. and They represent and The convolution operation.

[0033] Preferably, the pseudo-label reliability assessment module in S300 aims to resolve "selection ambiguity," that is, to ensure that the pseudo-labels generated by the teacher model are both correctly classified and accurately located. The specific steps are as follows:

[0034] The pseudo-label reliability assessment module introduces a dual-branch structure in the detection head, where one classification branch predicts the target category and the other localization branch evaluates the geometric consistency of the bounding box. The outputs of both branches are jointly calculated to determine the final pseudo-label confidence score. ;

[0035] in, Represents the category score, This represents the localization quality score; this joint metric can effectively eliminate pseudo-labels that have "high classification scores but severely offset bounding boxes", reducing the participation of low-quality pseudo-labels in training;

[0036] In the dual-branch structure, the supervision of the classification branch adopts a unified classification objective, and the specific formula is as follows: ;

[0037] in, Indicates focal loss, monitoring signal The definitions for labeled and unlabeled samples are as follows: ;

[0038] In labeled samples, IoU represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box. In unlabeled samples, the maximum response of the teacher model in each category is used as the supervision signal to reduce the interference caused by false label noise.

[0039] In addition, to ensure that the localization branch focuses on predicting geometric consistency, an additional IoU regression loss is introduced: ;in, This represents the binary cross-entropy loss;

[0040] This module employs a dual-branch structure for joint training and utilizes the IoU (Intersection over Union) between predicted bounding boxes and their corresponding ground truth bounding boxes to generate IoU-based soft labels. This comprehensively evaluates the reliability of false labels in both classification and localization. This effectively reduces false label misselection in forest fire detection caused by blurred boundaries or confusion between smoke and clouds at small-scale fire points.

[0041] Preferably, in S400, student model training is based on a separation strategy. The purpose of this step is to resolve "assignment ambiguity," that is, to utilize the filtered pseudo-labels for supervised learning more intelligently and robustly during the student model training phase, paying particular attention to the recall of small and fuzzy targets. The specific steps are as follows:

[0042] S401. Input student model and perform strong enhancement: Apply strong enhancement (such as random cropping, color dithering, blurring, etc.) to the same batch of unlabeled images, and then input them into the student model;

[0043] S402. Perform separate sample allocation based on the pseudo-labels obtained in S300 and their joint confidence scores. Set two thresholds and adaptive Candidate samples are divided into the following three categories: ;

[0044] in, The following is adaptively determined based on the sample distribution: ;

[0045] In classification tasks, all candidate samples participate in consistency learning to better mimic the classification output of the teacher model; while in localization tasks, candidate samples must meet more stringent conditions: they must be in the same class as the positive sample, have an IoU greater than 0.6, and have their center point located within the positive sample bounding box in order to be identified as potential positive samples.

[0046] For a potential positive sample that meets the criteria, its supervision box is obtained by the weighted average of the matching positive sample boxes: ;

[0047] in, This represents the positive sample boxes predicted by the teachers. This indicates the number of positive samples that match it.

[0048] This weighting strategy can reduce the adverse effects of individual noisy pseudo-boxes on training;

[0049] The final unsupervised loss consists of three parts:

[0050] ;

[0051] in, and These represent focus loss and IoU regression loss, respectively. This represents the generalized intersection-union loss, with weight parameters. The value is set to 0.5. This strategy avoids cross-interference from false label noise between tasks by identifying positive samples separately in the classification and localization tasks.

[0052] Preferably, experimental evaluation and verification are performed in S500, including:

[0053] S501. FCOS is selected as the baseline detector;

[0054] The S502 experimental platform is built based on PyTorch and runs on Windows 10. The hardware configuration includes two NVIDIA GeForce RTX 3090 GPUs and one Intel Core i7-11700K @ 3.60 GHz CPU.

[0055] S503. The stochastic gradient descent algorithm is used for model optimization. The specific parameter configuration is as follows: momentum parameter Weight decay rate Initial learning rate ;

[0056] S504. Configure the learning rate scheduling strategy, adopting a stepped decay strategy as the learning rate update strategy. Set the learning rate to decrease to 1 / 10 of its current value after every 30,000 iterations. The learning rate update formula is: ;in, Indicates the scheduling period;

[0057] S505. The overall training process is divided into two stages: a warm-up stage and a semi-supervised self-training stage. The first 100 complete learning cycles are used to obtain stable initial weights, and the subsequent 900 training cycles are used to fully optimize the model performance. The total training time is 1000 training cycles to ensure that the model fully converges.

[0058] S506. Set basic evaluation indicators, using precision and recall as the main evaluation indicators;

[0059] Precision measures the accuracy of predictions, while recall measures the model's ability to identify positive samples. The specific values ​​are as follows:

[0060] , ;

[0061] in, This represents the number of samples correctly identified as positive by the model. This represents the number of samples that were actually negative but were incorrectly identified as positive. This represents the number of samples that are actually positive but were not identified by the model.

[0062] S507. Set comprehensive performance indicators, and introduce average accuracy and mean of average accuracy as comprehensive evaluation indicators.

[0063] Among them, mean precision is used to evaluate the detection performance of a single class by calculating the area under the precision-recall curve, while the mean precision is calculated by averaging the AP of all classes, comprehensively reflecting the overall detection performance of the model. The specific details are as follows: , ;

[0064] in, This represents the total number of target categories (in forest fire detection tasks, n=2, corresponding to the two target categories of flame and smoke respectively).

[0065] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0066] This invention significantly improves the detection performance and generalization ability of the model under conditions of limited labeled samples by constructing a pseudo-label reliability assessment module (PREM) and a separate sample allocation strategy (DSAS), effectively utilizing massive amounts of unlabeled forest fire image data. By introducing a pseudo-label quality assessment mechanism based on multi-dimensional consistency, high-confidence pseudo-labels can be accurately selected for supervised learning, while low-confidence samples are used for feature representation enhancement, effectively avoiding interference from noisy labels on model training and significantly improving the convergence stability and robustness of the detection algorithm. Experimental results show that on forest fire datasets containing complex backgrounds, multi-scale fire points, and blurred smoke, this method significantly improves the mean detection accuracy (mAP), greatly increases the recall rate of small-scale fire points, and significantly reduces the false alarm rate compared to traditional semi-supervised methods. In addition, it demonstrates good adaptability and generalization ability in cross-scene testing, effectively handling fire identification tasks under different forest environments, lighting conditions, and meteorological interference. By improving the accuracy and real-time performance of fire identification, this invention provides a reliable technical guarantee for achieving early detection, early suppression, and early extinguishment of forest fires. It helps to significantly shorten fire response time, improve firefighting efficiency, and reduce disaster losses. This method has significant advantages in detection accuracy, environmental adaptability, and engineering deployability, providing core technical support for building an intelligent, all-weather forest fire monitoring and early warning system, and has significant social benefits and application value. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of the strategy in this invention.

[0069] Figure 2 This is an overall framework diagram of the method in this invention.

[0070] Figure 3 This is a diagram of the hybrid attention module in this invention.

[0071] Figure 4 This is a diagram of the pseudo-tag reliability assessment module in this invention.

[0072] Figure 5 These are example images of different fire scenarios in the dataset of this invention.

[0073] Figure 6 This is an example diagram of different fire smoke conditions in the dataset of this invention.

[0074] Figure 7 This is a visualization of the ablation experiment results in this invention.

[0075] Figure 8 This is a visualization of the experimental results comparing the method of this invention with other detection methods. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figures 1-8 The present invention provides the following technical solution:

[0078] Example 1: A forest fire image detection method based on semi-supervised learning, comprising:

[0079] in, Figure 1 The flowchart illustrates a forest fire image detection method based on semi-supervised learning, as described in this invention. Figure 2 The overall framework diagram of a forest fire image detection method based on semi-supervised learning in this invention is shown here. Figure 1 , Figure 2 This will illustrate the entire working process of the technical solution of the present invention.

[0080] Step 1: Perform model initialization and warm-up training:

[0081] This step lays the foundation for the entire semi-supervised training process, with the goal of obtaining an initial model with preliminary fire detection capabilities.

[0082] S101. Prepare and process the data, collect forest fire images, and construct a dataset. The dataset consists of two parts: ① labeled data, which is limited in quantity and contains precisely labeled "flame" and "smoke" bounding boxes. ② unlabeled data, which is much larger than the labeled data and contains no manual annotations.

[0083] S102. Perform model warm-up training using labeled data on a one-stage target detector (in this invention, a fully convolutional one-stage target detection algorithm is used for fully supervised training). The training process uses a standard detection loss function, in which the focus loss function can be used to learn visual features such as the bright red of flames and the gray-white texture of smoke, and the generalized intersection-union loss function is used to optimize the bounding box regression to accurately locate the fire target.

[0084] S103. Initialize the teacher-student model. After the pre-training, copy the trained model parameters. One copy is used for the student model, which is actively updated via gradient descent during subsequent training. The other copy is used for the teacher model, whose parameters are initially identical to those of the student model, but are not directly updated via gradient descent to maintain the stability of the generated pseudo-labels. Specifically:

[0085] The teacher-student models collaborate alternately to utilize unlabeled data. In each iteration, the teacher model first infers from unlabeled forest fire images. To ensure prediction stability, the input image undergoes weak enhancement operations (e.g., random flipping, scaling, slight color perturbation), and the teacher model's output bounding boxes and categories serve as pseudo-labels. Subsequently, the student model receives a strongly enhanced version of the same image (including random cropping, rotation, blurring, strong color perturbation, etc.) and is trained under the guidance of the pseudo-labels. During training, the student model's total loss function consists of the following two parts: ;

[0086] in, This represents the supervised loss for labeled samples, including classification loss and bounding box regression loss. The pseudo-label loss, representing unlabeled samples, is used to guide the student model to learn on unlabeled images. The represents the balance coefficient, used to adjust the relative importance of supervised and unsupervised losses, and is set to 1 in this invention.

[0087] To ensure the long-term stability of the teacher model's output, its parameters are not updated directly through backpropagation, but rather iterated using the exponential moving average (EMA) of the student model's parameters:

[0088] ;

[0089] in, and These represent parameters for teachers and students, respectively. This represents the attenuation coefficient. This strategy ensures that the teacher model's predictions are not affected by the instantaneous fluctuations in the student model, thus providing stability for pseudo-label generation.

[0090] The second step is to introduce a hybrid attention mechanism to enhance feature extraction:

[0091] This step aims to improve the model's ability to extract features from "small-scale fire points" and "smoke with blurred boundaries," thereby improving the quality of subsequent false labels from the source. The specific steps are as follows:

[0092] S201. Obtain the input feature map The input feature map is then normalized.

[0093] S202. Generate a shared query matrix through linear transformation. Bond matrix The specific formula is as follows: , ;

[0094] in, and Represents the shared linear transformation weight matrix;

[0095] S203, Construct a spatial attention module, and spatial value layer Projecting them onto the lower-dimensional space respectively, we get and The spatial attention map is then calculated according to the formula. ;

[0096] in, , and These represent the shared query, the projected key matrix, and the spatial value layer, respectively. This represents the dimension of each vector. The projection weights represent the spatial values;

[0097] S204. Build the channel attention module, using the same... and To learn complementary features and obtain channel value layers Its dimensions are The channel attention map is then calculated using the formula: ;

[0098] in, , and These represent the channel value layer, shared query, and shared key matrix, respectively. Indicates the projected weights of the channel values;

[0099] S205. The outputs of the two attention modules are added together and further processed through a convolutional layer to obtain an enhanced feature representation, such as... Figure 3 As shown; the final output of the hybrid attention module is: ;

[0100] in, and These represent spatial and channel attention maps, respectively. and They represent and The convolution operation.

[0101] Step 3: Generate and evaluate the reliability of pseudo-labels based on the teacher model:

[0102] This step aims to resolve "selection ambiguity," that is, to ensure that the pseudo-labels generated by the teacher model are both correctly classified and accurately located. The specific steps are as follows:

[0103] A pseudo-label reliability assessment module (PREM) is configured, which introduces a two-branch structure into the detection head. One branch predicts the target class, and the other evaluates the geometric consistency of the bounding box. The outputs of both are jointly calculated to determine the final pseudo-label confidence score. ;

[0104] in, Represents the category score, Indicates the positioning quality score;

[0105] In the dual-branch structure, the supervision of the classification branch adopts a unified classification objective, and the specific formula is as follows: ;

[0106] in, Indicates focal loss, monitoring signal The definitions for labeled and unlabeled samples are as follows: ;

[0107] In labeled samples, IoU represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box. In unlabeled samples, the maximum response of the teacher model in each category is used as the supervision signal to reduce the interference caused by false label noise.

[0108] In addition, to ensure that the localization branch focuses on predicting geometric consistency, an additional IoU regression loss is introduced: ;in, This represents the binary cross-entropy loss.

[0109] This module adopts, as follows: Figure 4 The dual-branch structure shown in the diagram enables joint training, while simultaneously generating IoU-based soft labels using the IoU between the predicted bounding boxes and their corresponding ground truth bounding boxes. This comprehensively measures the reliability of false labels in both classification and localization. This approach effectively reduces false label misselection in forest fire detection caused by blurred boundaries or confusion between smoke and clouds around small-scale fire points.

[0110] Step 4: Train the student model based on a separation strategy:

[0111] The purpose of this step is to resolve "assignment ambiguity," that is, to utilize the filtered pseudo-labels for supervised learning more intelligently and robustly during the student model training phase, with a particular focus on recalling small and fuzzy targets. The specific steps are as follows:

[0112] S401. Input and enhance the student model: Apply strong enhancement to the same batch of unlabeled images, and then input them into the student model;

[0113] S402. Perform separate sample allocation based on the pseudo-labels obtained in S300 and their joint confidence scores. Set two thresholds and adaptive Candidate samples are divided into the following three categories: ;

[0114] in, The following is adaptively determined based on the sample distribution: ;

[0115] In classification tasks, all candidate samples participate in consistency learning to better mimic the classification output of the teacher model; while in localization tasks, candidate samples must meet more stringent conditions: they must be in the same class as the positive sample, have an IoU greater than 0.6, and have their center point located within the positive sample bounding box in order to be identified as potential positive samples.

[0116] For a potential positive sample that meets the criteria, its supervision box is obtained by the weighted average of the matching positive sample boxes: ;

[0117] in, This represents the positive sample boxes predicted by the teachers. This indicates the number of positive samples that match it.

[0118] The final unsupervised loss consists of three parts:

[0119] ;

[0120] in, and These represent focus loss and IoU regression loss, respectively. This represents the generalized intersection-union loss, with weight parameters. Set it to 0.5.

[0121] Step 5: Conduct experimental evaluation and verification:

[0122] S501, FCOS is selected as the baseline detector.

[0123] The S502 experimental platform is built based on PyTorch 3.8 and runs on Windows 10. The hardware configuration includes two NVIDIA GeForce RTX 3090 GPUs and one Intel Core i7-11700K @ 3.60 GHz CPU.

[0124] S503. The stochastic gradient descent (SGD) algorithm is used for model optimization, with the following specific parameter configurations: momentum parameter. Weight decay rate Initial learning rate ;

[0125] S504. Configure the learning rate scheduling strategy, adopting a step-decay strategy for learning rate updates. Set the learning rate to 1 / 10 of its current value after every 30,000 iterations. The learning rate update formula is as follows:

[0126] ;in, Indicates the scheduling period;

[0127] S505. The overall training process is divided into two stages: a warm-up stage and a semi-supervised self-training stage. The first 100 complete learning cycles are used to obtain stable initial weights, and the subsequent 900 training cycles are used to fully optimize the model performance. The total training time is 1000 training cycles to ensure that the model fully converges.

[0128] S506. Set basic evaluation indicators, using precision and recall as the main evaluation indicators;

[0129] Precision measures the accuracy of predictions, while recall measures the model's ability to identify positive samples. The specific values ​​are as follows:

[0130] , ;

[0131] in, This represents the number of samples correctly identified as positive by the model. This represents the number of samples that were actually negative but were incorrectly identified as positive. This represents the number of samples that are actually positive but were not identified by the model.

[0132] S507. Set comprehensive performance indicators, and introduce average accuracy and mean of average accuracy as comprehensive evaluation indicators.

[0133] Among them, mean precision is used to evaluate the detection performance of a single class by calculating the area under the precision-recall curve, while the mean precision is calculated by averaging the AP of all classes, comprehensively reflecting the overall detection performance of the model. The specific details are as follows: , ;

[0134] in, This indicates the total number of target categories.

[0135] Example 2: Dataset Support:

[0136] To fully evaluate this invention and support the evaluation of fire detection models, the FSDD (Fire Smoke Detection Dataset) dataset was constructed, containing 12,000 forest fire images. All images were finely annotated using the LabelImg tool, covering two key targets: "fire" and "smoke".

[0137] The dataset was divided into training and testing sets in a 9:1 ratio. A further 10% of the training set was randomly selected as a validation set for model selection and parameter tuning. The images have a resolution of 1920×1080 pixels, suitable for high-resolution fire detection tasks. Images were acquired from multiple high-definition cameras deployed on different monitoring platforms, including fixed and mobile devices, covering various perspectives to simulate camera angle changes in actual fire monitoring. Specifically, the perspective distribution is as follows: vertical perspective (approximately 30%), suitable for high-altitude and rooftop monitoring; horizontal perspective (approximately 50%), suitable for ground and low-level monitoring; and oblique perspective (approximately 20%), suitable for dynamic monitoring and large-scale fire scenarios.

[0138] The image acquisition interval was set to 1 frame per second to capture the dynamic changes in the fire scene. The dataset contains images from different time periods, with 70% being daytime scenes and 30% being nighttime scenes, which helps to evaluate the model's performance under low light and complex lighting conditions. In addition, the dataset also covers a variety of weather conditions, including sunny (45%), cloudy (35%), and hazy (20%) weather, to reflect the changes in the visual characteristics of flames and smoke under different meteorological conditions.

[0139] To increase the challenge of the detection task, about 40% of the images contain occlusions, such as trees and buildings partially obscuring flames or smoke, simulating a scenario where the visibility of fire targets is limited in a real environment.

[0140] Figure 5 The exhibition showcases example images of fires of varying intensities, including small, medium, and large-scale fires, demonstrating the visual representation of flames and smoke against complex backgrounds and under obstructed conditions.

[0141] exist Figure 5 In the middle, (a) represents smoke obscuring / initial fire: (a) mainly shows dense smoke, with flames not obvious or obscured by smoke, demonstrating the model's ability to identify smoke.

[0142] (b) Medium-scale / complex background: dense forest, visible flames, demonstrating the identification of the mixed state of flames and smoke against a complex forest background.

[0143] (c) Steep terrain / oblique angle view: The fire occurred on a hillside, showing that the dataset contains different shooting angles (oblique views).

[0144] (d) is a large-scale fire: fires spread across the mountains and fields, with smoke filling the entire sky, representing a large-scale fire scene.

[0145] (e) Surface fire / local small fire: The flames are mainly concentrated in the ground shrub layer, which are small-scale fire points. This is the key challenge of "small target detection" that this patent aims to solve.

[0146] (f) Close-up / obstructed fire: This may be a close-up of the flames or a fire obstructed by trees, showing the characteristics of a partially obstructed fire.

[0147] The characteristics and advantages of the FSDD dataset can be summarized as follows:

[0148] 1. Multi-view coverage: Includes vertical, horizontal and oblique views, enhancing the model's adaptability to changes in camera position;

[0149] 2. Multiple time periods and multiple weather conditions: Covers different lighting conditions during the day and night and various weather scenarios, supporting model evaluation in complex environments;

[0150] 3. Occlusion and complex background: Introduce occlusion and complex background images to improve the model's detection robustness in partially visible or interfering environments;

[0151] 4. High-resolution images: All images are 1920×1080 pixels, suitable for high-precision detection tasks;

[0152] 5. Standard partitioning method: The training, validation and test sets are divided according to the common proportion, which facilitates model comparison and evaluation.

[0153] This dataset aims to provide a realistically challenging benchmark task for fire detection research, particularly for evaluating model performance in scenarios with obstructions, complex backgrounds, and various environmental disturbances. The plan is to release the FSDD dataset, along with usage instructions, after further refinement, to support the further development of fire detection technology.

[0154] To further validate the strategy proposed in this invention, in addition to using the FSDD dataset for validation, this invention also uses the publicly available dataset FIGLib (HPWREN Fire Ignition Image Library) for instance validation. This dataset contains wildfire image sequences captured by fixed-view surveillance cameras deployed on remote mountaintops in Southern California (belonging to the High Performance Wireless Research and Education Network, HPWREN). As of September 2025, FIGLib contains 488 fire sequences from 101 cameras at 30 monitoring sites between June 2016 and September 2025. Each sequence typically covers images from 40 minutes before to 40 minutes after the appearance of the visible smoke plume, and provides a binary label (smoke present / no smoke) for each image. The image acquisition interval is approximately 60 seconds, and each fire sequence contains a total of 81 images, but on average, about 6.6 images are missing per fire, distributed at random positions at the beginning, end, or middle of the sequence.

[0155] The dataset contains a total of 36,600 high-resolution images, with resolutions of 1536×2048 pixels or 2048×3072 pixels, depending on the camera model used. Pre-ignition monitoring and multi-angle coverage are achieved through a cluster of cameras, with four perspectives spanning over 90 degrees and remaining stable for many years, collectively covering a 360-degree area around the mountaintop. Figure 2 This presentation showcases several representative complex fire and smoke scenarios in FIGLib, including cloud-free forest fires, single or multiple fires at night, visible light images of smoke under cloud interference, and daytime and nighttime fire observations under infrared imaging conditions. Specifically, [details omitted]. Figure 6 As shown. These diverse scenarios help to evaluate the model's adaptability under different lighting conditions (daytime / nighttime), imaging spectra (visible / infrared), and weather conditions (sunny / cloudy interference).

[0156] In such Figure 6 In the middle, (a) and (d) are daytime visible light scenes: these two images are typical daytime surveillance perspectives, with sufficient light, and distant mountains, sky and smoke / fire points can be seen.

[0157] (a) shows the spread of smoke during the day. (d) shows a wildfire at a distance during the day.

[0158] (b) and (e) are nighttime visible light scenes: the backgrounds of these two images are black, but there are obvious light sources.

[0159] (b) shows multiple fires at night (with a clear band of firelight in the distance) or fire points under the interference of city lights. (e) shows a single point or small-scale fire in a dark night environment. These are used to verify whether the model can still capture fire points under extremely low light conditions.

[0160] (c) and (f) are infrared / grayscale imaging scenes: these two images are black and white, which is the infrared (IR) or grayscale mode unique to the FIGLib dataset.

[0161] Image (c) appears to be an infrared image taken at night or in low light, with the bright areas representing heat sources (fire). Image (f) appears to be a grayscale or near-infrared image taken during the day, where the smoke pattern is more pronounced in black and white contrast. This is used to verify whether the model has cross-modal generalization ability, meaning it can recognize not only color (red fire) but also shape and texture.

[0162] Example 3: Experimental Verification:

[0163] Validation was performed using a self-built FSDD dataset (12,000 high-resolution 1920×1080 images) and the publicly available FigLib dataset (36,600 images). Based on the FCOS detector, three annotation ratios (1 / 16, 1 / 8, and 1 / 4) were used to simulate low-label data scenarios. Mean AP (mAP) was used as the core metric, supplemented by recall and precision as evaluation metrics.

[0164] (1) Ablation experiment:

[0165] Ablation experiments were conducted on the FSDD and FIGLib datasets, and the results are shown in Table 1.

[0166] Table 1 Ablation test combinations and results

[0167]

[0168] Table 1 presents the ablation experiment results of the semi-supervised framework on the FSDD dataset, which clearly reflects the contribution of different modules to the overall performance.

[0169] In Table 1, (a) Baseline FCOS represents the baseline model (FCOS) trained using only semi-supervised training.

[0170] (b) +HAM (Hybrid Attention Module) means introducing a hybrid attention module (HAM) on the baseline model;

[0171] (c)+PREM (Pseudo-tag Reliability Assessment) means that a pseudo-tag reliability assessment module (PREM) is added on the basis of (b).

[0172] (d)+DSAS (excluding potential positive sample mining) means that, based on (c), a separate sample allocation strategy (DSAS) is adopted.

[0173] (e)+DSAS (including potential positive sample mining) means that potential positive sample mining is enabled based on (d).

[0174] Based on this, under the baseline FCOS with only semi-supervised training, the model's mAP at labeling ratios of 1 / 16, 1 / 8, and 1 / 4 were 76.2%, 79.5%, and 82.0%, respectively. This performance is limited by the complexity of forest fire scenarios: fire points are usually extremely small in scale, and smoke boundaries are blurred and highly similar to clouds and fog, leading to situations where the model often has high classification confidence but inaccurate localization during pseudo-label generation, thus weakening the effectiveness of pseudo-labels. After introducing a Hybrid Attention Module (HAM), the performance improved to 78.0%, 81.2%, and 83.6%, respectively. This improvement enhances the expression of key features through joint spatial-channel modeling, making the representation of small-scale fire points and blurred smoke more prominent in the feature space, thus improving pseudo-label generation and positive sample allocation from the source. Further adding a Pseudo-Label Reliability Evaluation Module (PREM) further improved the mAP to 80.5%, 83.1%, and 85.2%. This module, by jointly considering classification scores and localization quality, avoids the one-sidedness of relying solely on classification confidence, effectively mitigating selection ambiguity and making pseudo-labels more reliable while reducing the adoption of pseudo-boxes with high scores but off-target characteristics. Building on this, a Separate Sample Assignment Strategy (DSAS) is applied to independently separate positive samples for classification and localization tasks, reducing interference between the two tasks. Without latent positive sample mining, performance reaches 81.7%, 84.2%, and 86.3%, demonstrating that separating classification and localization tasks allows the model to utilize pseudo-labels more stably. Further enabling latent positive sample mining improves mAP to 83.0%, 85.5%, and 87.4%, indicating that stricter conditional screening of candidate samples and weighted generation of supervision boxes better uncovers overlooked small fire points and vaguely defined smoke targets, significantly alleviating assignment ambiguity. Ultimately, when the Hybrid Attention (HAM) module, the pseudo-label reliability assessment module, and the separate sample allocation strategy work together, the method of this invention achieves mAP of 84.1%, 86.7%, and 88.3% at labeling ratios of 1 / 16, 1 / 8, and 1 / 4, respectively, representing an improvement of over 7 percentage points compared to the baseline. Overall, the results show that HAM lays the foundation for feature enhancement, PREM provides assurance in pseudo-label selection, and DSAS, especially in latent positive sample mining, plays a crucial role in improving recall and boundary localization. The synergy of these three modules significantly improves the accuracy and robustness of forest fire detection.

[0175] To verify the key role of the pseudo-label reliability assessment module (PREM) in forest fire detection tasks and the effectiveness of its phased optimization strategy, further detailed experiments were conducted on the FSDD and FIGLib datasets, and the results are shown in Table 2.

[0176] Table 2. Detailed experimental results of the fake tag reliability assessment module.

[0177]

[0178] Table 2 presents a detailed experiment of the False Tag Reliability Assessment Module (PREM). Based on the baseline model (FCOS+HAM), smaller components within PREM were added layer by layer, like peeling an onion. This demonstrates that every small design element within the PREM module (adding branches, modifying formulas, setting objectives) makes a substantial contribution.

[0179] In Table 2, (a) Baseline FCOS+HAM represents the baseline model FCOS+HAM;

[0180] (b)+ IoU branch prediction represents the introduction of IoU branch prediction on the basis of the baseline model FCOS+HAM;

[0181] (c) + Joint Supervision (Cls+IoU) means adding a joint supervision mechanism (Cls+IoU) on top of (b);

[0182] (d)+Special Unlabeled Target indicates that a special soft target setting is used on unlabeled data based on (c).

[0183] As shown in Table 2, under the baseline FCOS+HAM conditions, the model's mAP at annotation ratios of 1 / 16, 1 / 8, and 1 / 4 were 78.0%, 81.2%, and 83.6%, respectively, indicating that the hybrid attention module can enhance the feature representation of small-scale fire points and blurred smoke to a certain extent. However, relying solely on classification scores to filter pseudo-labels still has significant limitations; many pseudo-boxes with high scores but inaccurate localization interfere with training. When IoU branch prediction is introduced, the mAP improves to 79.4%, 82.1%, and 84.2%, indicating that explicitly modeling the geometric accuracy of the boxes effectively filters out some low-quality pseudo-labels. Building on this, a joint supervision mechanism (Cls+IoU) is further added, improving performance to 80.1%, 82.7%, and 84.6%. This design binds classification scores to the localization quality of the boxes, fundamentally alleviating the problem of the disconnect between classification and localization branches, significantly reducing ambiguity in pseudo-label selection, and becoming a key factor in improving performance. Finally, by employing a special soft-target setting on unlabeled data, the performance was further improved to 80.5%, 83.1%, and 85.2%, effectively mitigating the instability caused by noisy pseudo-labels in unlabeled samples. Overall, the gradual improvement of this module significantly enhanced the reliability and utilization efficiency of pseudo-labels, making the model more robust in detecting small fire points and blurred smoke in complex scenes.

[0184] Figure 7 The visualization results of target detection for different ablation experiment combinations (a)–(f) on the 1 / 8 labeled FSDD dataset are presented, clearly showing the progressive improvement in detection performance of each module. Figure 7 (a)-(f) correspond one-to-one with (a)-(f) in Table 1; Table 1 is the data proof, while Figure 7 It is visual proof.

[0185] exist Figure 7 In baseline (a), the model relies solely on semi-supervised training, resulting in generally weak responses from fire targets. Canopy reflections and cloud / fog areas generate numerous false detections, and smoke boundaries exhibit fragmentation and incompleteness. After adding a hybrid attention module to (a) to obtain model (b), the brightness and contrast of small-scale fire points in the image are significantly enhanced, some small flame hotspots can be captured, and false alarms caused by background texture are reduced. However, smoke areas still exhibit blurriness and omissions. Further combining a pseudo-label reliability assessment module with (b) to obtain model (c), many false high-confidence boxes are eliminated, target boundaries better match actual fire points and smoke morphology, and the number of false detections decreases significantly. Using a separate sample allocation strategy based on (c) to obtain model (d), the classification and localization of targets become clearer, previously missed small fire points begin to be detected, smoke outlines become more complete, and overall detection results are more stable. Building upon (d), by further introducing latent positive sample mining to obtain model (e), we can see that the model has a more sensitive response to flames and smoke with blurred edges and partial occlusion, the detection box coverage is more comprehensive, and the boundary continuity is further improved. Finally, under the framework (f) proposed by the strategy of this invention, the detection results are the most robust: both the flame target and the smoke region are clearly identified, the predicted box is basically consistent with the real boundary, false responses in complex backgrounds are greatly reduced, and the overall detection effect has achieved a qualitative leap from coarse and scattered to precise and continuous.

[0186] (2) Comparative experiment:

[0187] On the FSDD and FIGLib datasets, the strategy of this invention was compared with the currently mainstream models such as Strong Teacher and Consistent student (STAC), Interactive Self-training with Mean Teachers (ISMT), HumbleTeacher, Unbiased Teacher, Soft Teacher, Pseudo-labeling and Consistency training (PseCo), and DenseTeacher. The experimental results are shown in Table 3.

[0188] Table 3 Comparative Experimental Results of Semi-Supervised Target Detection Methods

[0189]

[0190] Table 3 and Figure 8One-to-one correspondences demonstrate the advancement of this invention from two dimensions: data metrics and visual effects. Table 3 shows that STAC, as the earliest work to introduce self-training into object detection, achieved mAP of 77.0%, 80.1%, and 82.5% at annotation ratios of 1 / 16, 1 / 8, and 1 / 4, respectively, proving the feasibility of improving fire detection using unlabeled samples. However, in scenarios where fire points are generally small-scale and smoke boundaries are blurred, STAC's pseudo-labels generated solely by the teacher model still contain numerous errors, especially prone to producing pseudo-boundaries with high scores but inaccurate localization. ISMT, building upon this, introduces consistency regularization and alleviates some pseudo-label noise through multi-view enhancement, improving performance to 77.8%, 80.6%, and 83.1%, but still struggles to fully guarantee pseudo-label stability when smoke and fog are indistinguishable. Humble Teacher, by using soft pseudo-labels, reduces overconfident erroneous predictions, further improving performance to 78.5%, 81.2%, and 83.8% at the three ratios, demonstrating greater robustness even with blurred smoke boundaries. The Unbiased Teacher effectively mitigates the problem of uneven pseudo-label distribution through a bias-corrected loss function, exhibiting better robustness, especially when small fire points are missed, with performance improvements reaching 79.3%, 82.0%, and 84.2%. The Soft Teacher further optimizes the pseudo-label weighting and bounding box perturbation mechanisms, amplifying the role of high-quality pseudo-labels and maintaining good supervision in images with sparse fire points and complex smoke boundaries, improving mAP to 79.9%, 82.5%, and 84.7%. PseCo, by combining pseudo-labels and contrastive learning, not only utilizes existing supervision information but also mines feature consistency among unlabeled samples, making the model more robust in distinguishing between smoke and cloud interference, with performance improvements reaching 80.7%, 83.3%, and 85.4%. The Dense Teacher further introduces dense prediction consistency, effectively enhancing the efficiency of pixel-level pseudo-label utilization, improving recall for small-scale fire points and the ability to characterize blurred smoke boundaries, achieving 81.4%, 83.9%, and 86.0% at the three scales, respectively.

[0191] In comparison, the method proposed in this invention achieved the best results of 84.1%, 86.7%, and 88.3% under the same conditions on the FSDD dataset, significantly outperforming all comparative methods with an average improvement of approximately 2–3 percentage points. This advantage mainly stems from three innovative design aspects: First, the hybrid attention module enhances the expression of key features through spatial-channel joint modeling, making small-scale fire points easier to distinguish and blurry smoke regions clearer; second, the pseudo-label reliability assessment module jointly measures both classification and localization, effectively eliminating low-quality pseudo-labels with "high classification confidence but boundary offset," alleviating pseudo-label ambiguity caused by the similarity between smoke and fog in forest fire detection; finally, the separate sample allocation strategy, especially the potential positive sample mining mechanism, effectively improves the recall rate of missed fire points and provides more robust localization supervision in cases of blurred smoke boundaries. The overall comparative results fully verify the significant advantages and stability of the method in complex forest fire scenarios, indicating its higher application value in actual fire monitoring tasks.

[0192] Figure 8The visualization results of the proposed strategy on the FSDD dataset with a 1 / 8 annotation ratio are presented, allowing a clear view of the differences in performance among various algorithms in forest fire scenarios. In the STAC detection results, small-scale fire points are often missed, smoke area boundaries are incomplete, and significant false detections occur in canopy and bright light areas. ISMT alleviates this problem to some extent, making some smoke contours more coherent, but its response to weak fire points remains insufficient. Humble Teacher reduces overconfident false predictions through soft pseudo-labels, increasing the number of fire point detections and smoothing smoke boundaries, but false responses still exist in cloud and fog interference areas. Unbiased Teacher is more sensitive to detecting some missed fire points, but the overall boundaries are still not complete enough. Soft Teacher and PseCo further improve visual performance: the former effectively filters out some "false high-confidence" detection boxes, making fire point localization more accurate; the latter performs more robustly in distinguishing between smoke and fog, with significantly improved contour integrity. Dense Teacher enhances its ability to detect small-scale fire points, and smoke areas show higher continuity, but a small number of false detections still occur in complex backgrounds. In comparison, the detection results of the strategy of this invention are the clearest and most stable: small fire points are almost completely eliminated, smoke area boundaries are continuous and complete, false detections in complex canopies and bright areas are significantly reduced, and the predicted bounding boxes highly match the actual fire targets. Overall comparison shows that this invention exhibits the best visual effect and robustness at a 1 / 8 annotation ratio, and can more realistically reflect the distribution of flames and smoke in a fire scene. In-depth analysis reveals that this advantage is due to the role of three improvements at different stages: the hybrid attention module significantly enhances the representation of small-scale fire points and blurred smoke during the feature extraction stage, making pseudo-label generation more reliable; the pseudo-label reliability evaluation module effectively alleviates the selection ambiguity caused by relying solely on classification scores, avoiding the adoption of pseudo-boundaries with "high scores but bias"; and the separate sample allocation strategy combined with potential positive sample mining improves the recall rate of missed fire points during training and enhances the localization accuracy of smoke boundaries. These three designs run through the key stages of feature extraction, pseudo-label screening, and sample allocation, forming a complementary effect, thus enabling the method of this invention to outperform existing methods in both detection accuracy and robustness in complex scenes.

[0193] (3) Performance evaluation based on complex environments:

[0194] To further evaluate the stability and accuracy of the proposed strategy under different environmental conditions, experiments were conducted on the FigLib and FSDD datasets. Three common types of disturbance were introduced into the datasets: random noise, illumination variation, and partial occlusion. These disturbances simulate real-world environmental factors, and the model's performance was tested under challenging conditions. The disturbance types and their descriptions are as follows:

[0195] ① Random noise: Gaussian noise with a mean of 0 and a standard deviation of 0.1 is added to the image to simulate common sensor noise and environmental factors in camera hardware;

[0196] ② Lighting variations: Adjust the image brightness to simulate different lighting conditions, including low light and overexposure.

[0197] ③ Partial Occlusion: Random objects such as trees, buildings, and clouds are superimposed on the image to simulate occlusion caused by environmental elements.

[0198] The tests were conducted using a dataset with added perturbations, and the results were compared with the original, unperturbed data. The results are summarized in Table 4.

[0199] Table 4 Performance evaluation results after adding interference

[0200]

[0201] As shown in the table above, the present invention performs best under the "No Interference" condition, particularly in the mAP@1 / 16 and mAP@1 / 8 scores, where it significantly outperforms other interference conditions. In contrast, the model's performance deteriorates after the addition of random noise and partial occlusion, especially in the mAP@1 / 16 and mAP@1 / 8 metrics. The robustness of the present invention is weaker under high-intensity noise and occlusion conditions, but it still maintains a certain level of performance in challenging environments. Furthermore, the experiments compared performance under different lighting conditions, indicating that the model is highly sensitive to changes in lighting, especially under low-light conditions, where its performance declines. This further verifies the adaptability of the present invention in practical applications, particularly how varying environmental conditions affect its stability and accuracy.

[0202] Overall, this invention demonstrates good adaptability under various environmental interference conditions. In particular, its accuracy remains high even in the absence of strong interference, making it suitable for tasks such as fire monitoring in real-world environments.

[0203] Through systematic analysis of the above experiments, this invention demonstrates its systematic innovation and comprehensive performance advantages for forest fire detection tasks. The framework, through three core designs—Hybrid Attention Module (HAM), Pseudo-Label Reliability Assessment Module (PREM), and Distributed Sample Assignment Strategy (DSAS)—forms a complete technology chain from feature enhancement and label cleanup to sample optimization, effectively addressing key challenges such as difficulties in small-scale fire detection, blurred smoke boundaries, and poor pseudo-label quality. Ablation experiments show that each module has a clear cumulative contribution, gradually increasing mAP from a baseline of 79.5% to 86.7% at a 1 / 8 labeling ratio, with a cumulative gain exceeding 7 percentage points, validating its synergistic design concept. Detailed experiments further reveal that the PREM module significantly alleviates the "selection ambiguity" caused by inconsistencies between classification and localization through mechanisms such as IoU branch prediction and joint supervision. In comparison with various mainstream semi-supervised detection methods, the method of this invention achieves significant superior performance on the FSDD and FIGLib datasets, with its mAP being nearly 3 percentage points better than the best comparison method. Combined with the more complete fire point recognition, clearer smoke boundaries and fewer false detections in the visualization results, it fully demonstrates that the method has stronger robustness, higher detection accuracy and practical application potential in complex scenarios.

[0204] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A forest fire image detection method based on semi-supervised learning, characterized in that: The method comprises: S100, constructing a teacher-student model as an initial model, and performing model initialization and pre-training; S200, introducing a mixed attention mechanism to the model, and respectively building a spatial attention module and a channel attention module to enhance feature extraction; S300, setting a pseudo-label reliability evaluation module based on the teacher model, generating and evaluating the pseudo-label reliability; The pseudo-label reliability evaluation module introduces a double-branch structure in the detection head, one classification branch is used to predict the target class, and the other positioning branch is used to evaluate the geometric consistency of the bounding box, and the results output by the two are jointly calculated as the final pseudo-label confidence; the supervision of the classification branch in the double-branch structure adopts a unified classification target; at the same time, in order to make the positioning branch focus on predicting geometric consistency, an IoU regression loss is additionally introduced; S400, training the student model based on a separation strategy, specifically comprising: applying strong enhancement to the same batch of unlabeled images and inputting the student model; based on the pseudo-label and its joint confidence, the candidate samples are divided into three categories: positive samples, negative samples and ignored samples; in the classification task, all candidate samples participate in consistency learning, while in the positioning task, the candidate samples need to meet the same class as the positive samples, the IoU is higher than 0.6, and the center point is located in the positive sample box, so as to be identified as potential positive samples; for the potential positive samples meeting the conditions, the supervision box is obtained by matching the weighted average of the positive sample box, and the final unsupervised loss is calculated; S500, performing experimental evaluation and verification. 2.The forest fire image detection method based on semi-supervised learning according to claim 1, wherein, The model initialization and pre-training in S100 comprises: S101, acquiring forest fire images and constructing a data set; The data set contains two parts: labeled data and unlabeled data; wherein, the number of labeled data is limited, and contains accurate labeled "flame" and "smoke" bounding boxes; the number of unlabeled data is much larger than that of labeled data, and there is no any artificial labeling; S102, acquiring labeled data, and performing complete supervised training on a one-stage target detector; In the training process, a standard detection loss function is used, wherein, a focal loss function can be used to learn the bright red color of the flame and the gray-white texture visual features of the smoke, and a generalized intersection over union loss function is used to optimize the bounding box regression to accurately locate the fire target; S103, teacher-student model initialization, after the complete supervised training is completed, the trained model parameters are copied, one set of parameters is used for the student model, and the other set of parameters is used for the teacher model, specifically comprising: The teacher-student model cooperates in an alternating manner to complete the utilization of the unlabeled data; in each iteration, the teacher model first performs inference on the unlabeled forest fire image, performs weak enhancement operation on the input image, and takes the prediction box and the class output by the teacher model as a pseudo label; secondly, the student model receives a strong enhanced version of the same image and is trained under the guidance of the pseudo label; in the training process, the total loss function of the student model is: ​ wherein, represents the supervised loss of the labeled sample, including the classification loss and the bounding box regression loss, represents the pseudo-label loss of the unlabeled sample, represents the balance coefficient; To ensure the long-term stability of the teacher model output, the model parameters are not directly updated by backpropagation, but are iteratively updated by the exponential moving average of the student model parameters: ; wherein, and denote the parameters of the teacher and student, respectively, denotes the decay coefficient. 3.The forest fire image detection method based on semi-supervised learning according to claim 1, wherein, In S200, the mixed attention mechanism is introduced to the model, comprising: S201、obtain an input feature map normalize the input feature map; S202, generating a shared query matrix by linear transformation and key matrix , the specific formula is: , ; wherein, and denotes a shared linear transformation weight matrix; S203, build a spatial attention module, project and spatial value layer to low-dimensional space respectively, obtain and , then calculate the spatial attention map according to the formula ; wherein, , and denote the shared query, the projected key matrix and the spatial value layer, respectively, denotes the dimension of each vector, denotes the projection weight of the spatial value; S204, build a channel attention module, use the same and to learn complementary features, get the channel value layer , the dimension is , then the channel attention map is calculated according to the formula: ; wherein, , and represent a channel value layer, a shared query and a shared key matrix, respectively, represent a projection weight of a channel value; S205, add the outputs of the two attention modules, and further process them through a convolution layer to obtain an enhanced feature representation; the final output of the mixed attention module is: ; wherein, and denote spatial and channel attention maps, respectively, and denote and convolution operations, respectively. 4.The forest fire image detection method based on semi-supervised learning of claim 1, wherein, The pseudo-label reliability evaluation module in S300 comprises: The pseudo-label reliability evaluation module introduces a double-branch structure in the detection head, one classification branch is used to predict the target class, and the other positioning branch is used to evaluate the geometric consistency of the bounding box, and the results output by the two are jointly calculated as the final pseudo-label confidence: ; wherein, denotes a classification score, denotes a positioning quality score; The supervision of the classification branch in the double-branch structure adopts a unified classification target, and the specific formula is: ; wherein, represents a focal loss, a supervisory signal The definitions in the labeled samples and unlabeled samples are respectively: ; In the labeled sample, IoU represents the intersection over union of the predicted box and the true box, and in the unlabeled sample, the maximum response of the teacher model in each class is used as a supervision signal to reduce the interference of pseudo-label noise; At the same time, in order to make the positioning branch focus on predicting geometric consistency, an additional IoU regression loss is introduced: ; wherein, denotes the binary cross-entropy loss. 5.The forest fire image detection method based on semi-supervised learning according to claim 1, wherein, In S400, the student model is trained based on a separation strategy, comprising: S401, student model input and strong enhancement: applying strong enhancement to the same batch of unlabeled images and inputting the student model; S402, performing separated sample distribution based on the pseudo label and the joint confidence obtained in S300 , setting two thresholds and adaptive ; dividing the candidate samples into the following three categories: ; wherein The sample distribution adaptive determination is as follows: ; In the classification task, all candidate samples participate in consistency learning to better imitate the classification output of the teacher model; while in the positioning task, candidate samples need to meet more stringent conditions: the same class as the positive sample, IoU higher than 0.6 and the center point located within the positive sample box, to be identified as potential positive samples; Then for the qualified potential positive samples, their supervision boxes are obtained by the weighted average of the matched positive sample boxes: ; wherein, represents a positive sample frame predicted by the teacher, represents the number of positive samples matched therewith; Then the final unsupervised loss is composed of three parts: ； where, and denote focal loss and IoU regression loss, respectively, denotes generalized intersection over union loss, and the weight parameter is set to 0.

5. 6.The forest fire image detection method based on semi-supervised learning according to claim 1, wherein, The experimental evaluation verification in S500 includes: S501, select FCOS as the baseline detector; S502, build an experimental platform based on PyTorch; S503, adopt random gradient descent algorithm to optimize the model, and the specific parameter configuration is: momentum parameter , weight decay rate , initial learning rate ; S504, configure the learning rate scheduling strategy, adopt the stepwise decay strategy as the learning rate updating strategy, set the learning rate to 1 / 10 of the current value after completing 30,000 iterations each time, and the learning rate updating formula is: ; wherein, denotes the scheduling period; S505, divide the whole training process into two stages: preheating stage and semi-supervised self-training stage. The first 100 complete learning cycles are used to obtain stable initial weights, and the subsequent 900 training cycles are used for full optimization of model performance. The total training time is 1000 training cycles to ensure that the model is fully converged. S506, set the basic evaluation index, and use precision and recall as the main evaluation index; Among them, precision is used to measure the accuracy of the prediction result, and recall is used to measure the ability of the model to identify positive samples. The specific ones are as follows: ， ； wherein, represents the number of samples correctly identified as positive class by the model, represents the number of samples that are actually negative class but are incorrectly identified as positive class, represents the number of samples that are actually positive class but are not identified by the model. S507, set the comprehensive performance index, and introduce the average precision and average precision mean as the comprehensive evaluation index; Among them, the average precision evaluates the detection performance of a single class by calculating the area under the precision-recall curve, and the average precision mean is obtained by averaging the AP of all classes, which comprehensively reflects the overall detection performance of the model, and the specific formulas are as follows: , ; wherein, denotes the total number of target classes.

Images