Fire behavior model training method based on image recognition and recognition technology

Abstract

The invention discloses a fire behavior model training method based on image recognition and a recognition technology, and belongs to the technical field of image recognition and fire monitoring. Aiming at the problems that a traditional fire monitoring technology depends on single sensor data, the false alarm rate is high in a complex environment and the like, an edge calculation unit and a TensorRT quantification framework are integrated on a hardware layer, lightweight deployment of a model is realized, and it is ensured that real-time reasoning delay is lower than 50 ms. The system supports multi-stage early warning linkage, can adapt to complex scenes such as industrial plants, forests and ships, still keeps the recognition accuracy of 92% or above under extreme conditions such as strong backlight and dust shielding, and has the false alarm rate of lower than 2%. According to the technology, through algorithm-hardware collaborative optimization, a high-reliability fire prevention and control solution is provided for smart cities, industrial safety and cultural heritage protection.

Description

Technical Field

[0001] This invention relates to the field of image recognition and fire monitoring technology, and in particular to a fire model training method and recognition technology based on image recognition. Background Technology

[0002] Traditional fire detection technologies primarily rely on manually set color thresholds, morphological features, or single sensor data, such as detecting temperature anomalies through infrared thermal imaging or the distribution of flame color in visible light images. While these methods are effective in simple scenarios, they have significant limitations in complex environments. For instance, changes in lighting, smoke obscuring the flames, or light and shadow interference similar to flames can easily lead to misjudgments, and they cannot effectively capture the dynamic spread characteristics of flames.

[0003] While existing deep learning-based methods have improved feature extraction capabilities through convolutional neural networks, they still face challenges such as strong data dependence and insufficient model generalization. Some solutions employ target detection algorithms like YOLO, but their fixed anchor box mechanism struggles to adapt to drastic changes in flame morphology and lacks temporal correlation modeling in long-term series analysis. Furthermore, most systems fail to adequately consider multi-dimensional data fusion, including meteorological conditions and geographical environment, resulting in persistently high false alarm rates in forest fires or industrial scenarios. At the hardware deployment level, traditional solutions rely heavily on high-performance GPUs or dedicated servers, which are insufficient to meet the real-time processing needs of edge devices. While emerging spiking neural networks offer low power consumption, their hardware adaptability and computational efficiency remain bottlenecks. Summary of the Invention

[0004] The purpose of this invention is to provide a fire model training method and recognition technology based on image recognition, comprising the following steps: Data acquisition and preprocessing: acquiring a fire image dataset containing different environmental scenarios, including images of natural fire sources, industrial fire sources, and simulated fire scenarios; performing denoising, illumination equalization, and color space normalization on the images to generate standardized training data. Natural fire sources include forest fires, grassland fires, etc.; industrial fire sources include high-temperature operating areas such as oil refineries and chemical plants; simulated fire scenarios generate multi-angle, multi-distance flame images through a controlled combustion experimental platform. Denoising employs a nonlocal mean algorithm combined with image patch similarity weight calculation to preserve flame edge details; illumination equalization is based on Retinex theory, eliminating shadow interference through multi-scale decomposition; color space normalization converts RGB images to Lab space, performing adaptive histogram matching on the brightness channels. The multi-dimensional dataset covers complex fire environments; nonlocal mean denoising avoids edge blurring caused by traditional Gaussian filtering; Lab space normalization enhances the model's adaptability to illumination changes, providing high-quality input for subsequent feature extraction.

[0005] Furthermore, the data preprocessing in step 1 includes: applying adaptive Gaussian filtering to denoise the image, with the noise standard deviation adaptively adjusted based on the local gradient of the image; using the CLAHE algorithm to enhance the image through block-by-block contrast, with each block being 8×8 pixels; converting the RGB color space to the HSV color space and performing histogram equalization on the saturation channel. The kernel size of the adaptive Gaussian filter is dynamically adjusted, using a 3×3 small kernel to suppress noise in low-gradient regions and switching to a 5×5 large kernel to preserve edges in high-gradient regions; the CLAHE algorithm performs contrast stretching on each 8×8 block, limiting the maximum contrast to 1.5 times that of the original image to avoid overexposure; after HSV conversion, the histogram equalization of the saturation channel uses an adaptive binning strategy to map pixel value distribution to a uniform range. Dynamic Gaussian filtering balances noise suppression and detail preservation, CLAHE block-by-block enhancement enhances the texture contrast of the flame region, and HSV space processing strengthens the color features of the flame, improving the model's ability to distinguish complex backgrounds.

[0006] Furthermore, the multi-scale feature extraction module in step 2 includes: an initial CNN layer extracting local features at three scales: 4×4, 8×8, and 16×16; a Transformer encoder performing self-attention calculation on global features, with query, key, and value matrices of dimensions 64×64×256; and a dynamic attention mechanism adjusting the receptive field through deformable convolutional kernels, with deformation parameters adaptively generated by the gradient direction of the flame edge. The CNN layer adopts a residual connection structure, with 3 stacked layers of 4×4 convolutional kernels extracting fine-grained texture, 8×8 kernels expanding the receptive field through dilated convolution, and 16×16 kernels combined with a dense connection module fusing multi-branch features. The Transformer encoder contains 6 layers of multi-head self-attention, with 8 heads per layer, and position encoding generated using a sine function. The deformation parameters of the deformable convolutional kernels in the dynamic attention mechanism are generated by the gradient direction field of the flame edge, with an offset learning range of ±15 pixels. Multi-scale CNN extracts local morphological features of the flame, Transformer globally models the flame diffusion pattern, and dynamic attention adaptively focuses on the flame edge, significantly improving the detection accuracy of small flames.

[0007] Furthermore, the temporal feature fusion training described in step 3 includes: setting the number of hidden layer units in the Bi-LSTM to 256, and the time step to the reciprocal of the video frame rate; the spatiotemporal correlation calculation uses a weighted fusion function of cosine similarity and Euclidean distance, with the weight coefficients optimized through backpropagation using cross-entropy loss; the dynamic feature vector dimension is 512×T, where T is the length of the continuous frame sequence. Each Bi-LSTM unit contains two independent LSTM layers, processing forward and reverse temporal information respectively, with the hidden layer state constrained by the tanh activation function; in the spatiotemporal correlation calculation, the cosine similarity weight is 0.6, the Euclidean distance weight is 0.4, and the fusion function output dimension is 128; the dynamic feature vector is generated by stacking the hidden states of the Bi-LSTM frame by frame, with a maximum T of 30 frames. The bidirectional LSTM captures the temporal variation patterns of the flame, the weighted fusion function balances spatial and temporal dependencies, and the long-sequence modeling capability enhances the model's accuracy in predicting the flame spread trend.

[0008] Further, the adversarial training optimization in step 4 includes: the generator loss function is L_GAN = λ_adv·L_adv + λ_feat·L_feat, where λ_adv = 0.5 and λ_feat = 0.3; the discriminator adopts a PatchGAN structure to perform binary classification on image blocks, with a block size of 16×16; the training iterations are 100 epochs, and the learning rate adopts a cosine annealing strategy, with an initial learning rate of 0.001 and a final learning rate of 0.00001. In the generator loss function, L_adv is the standard adversarial loss, and L_feat is the feature matching loss, calculating the Euclidean distance between generated features and real features in the intermediate layer; the PatchGAN discriminator divides the input image into 16×16 non-overlapping blocks, and each block independently outputs a binary classification probability; the cosine annealing strategy adjusts the learning rate every 10 epochs, with a decay coefficient of 0.95. Adversarial training improves the model's robustness to noise and spoofing interference, feature matching loss enhances the discriminator's sensitivity to subtle features, and the annealing strategy avoids getting trapped in local optima.

[0009] Furthermore, the migration entropy regularization loss function described in step 5 is defined as:

[0010] Where pi is the activation probability of the i-th neuron in the transfer feature layer, N is the total number of neurons, and the regularization coefficient γ = 0.01. The transfer feature layer selects the output of the 4th residual block of the pre-trained ResNet-50, N = 2048; the activation probability is calculated by Softmax normalization, and the entropy value reflects the uniformity of the feature distribution; the regularization term is added to the total loss function with a weight coefficient of 0.001. Entropy regularization constrains the diversity of feature distribution, prevents catastrophic forgetting in transfer learning, and improves the model's generalization ability in unseen scenarios.

[0011] Furthermore, a fire detection technology, based on a model obtained by the training method described in any one of claims 1-6, includes: a real-time image acquisition module: simultaneously acquiring environmental images via an infrared thermal imaging camera and a visible light camera; a feature compression encoder: performing JPEG2000 lossy compression on the input image, with a compression ratio ≥20:1, retaining the flame-sensitive frequency band; a lightweight inference engine: employing the TensorRT acceleration framework to quantize the model to INT8 precision, with an inference latency ≤50ms; and a multi-level early warning module: setting three alarm thresholds based on the flame confidence score, with threshold ranges of 0.6, 0.8, and 0.95, respectively. The infrared camera uses an uncooled vanadium oxide detector with a spectral response range of 8-14μm and a noise equivalent temperature difference ≤50mK; the visible light camera is equipped with a global shutter CMOS sensor, with a quantum efficiency ≥70% and a frame rate ≥60fps; the JPEG2000 compression uses the EBCOT algorithm, retaining the 0.6-2.5μm band of the flame-sensitive frequency band; the TensorRT quantization process uses a calibration dataset to dynamically adjust the scaling factor. Dual-modal data synchronous acquisition ensures all-weather detection capability, JPEG2000 compression reduces data transmission bandwidth, INT8 quantization balances accuracy and computational efficiency, and three-level early warning enables graded response.

[0012] Furthermore, the infrared thermal imaging camera operates in the 8-14μm wavelength range, with a spatial resolution of 640×512 pixels and a frame rate of 30fps; the visible light camera employs a global shutter CMOS sensor with a quantum efficiency ≥70%. The infrared camera is equipped with a Stirling cooling module with a cooling power ≥0.8W and a thermal sensitivity ≤30mK; the CMOS sensor uses a back-illuminated structure with a pixel size of 1.4μm×1.4μm and supports HDR mode; the data interface is USB 3.0 with a transmission rate ≥5Gbps. High-resolution infrared imaging enhances the detection capability of small fire sources, the global shutter eliminates motion blur, and the HDR mode adapts to strong light environments, ensuring reliable identification under extreme conditions.

[0013] Furthermore, the feature compression encoder employs an improved SPIHT algorithm to adaptively segment and encode the flame region ROI, with block sizes of 16×16 or 32×32. Improvements to the SPIHT algorithm include: ① introducing prior flame color to construct a hierarchical tree structure; ② adaptive adjustment of the importance coefficient threshold; and ③ a 30% increase in zero-tree encoding efficiency. ROI detection uses the YOLOv5 model with a confidence threshold set to 0.7. Adaptive segmentation reduces redundant information, improves the SPIHT algorithm's compression ratio to 25:1, achieves ROI detection accuracy of 98%, and reduces the computational load of subsequent processing.

[0014] Furthermore, the multi-level early warning module is linked to the fire control system. When the confidence level is ≥0.95, it triggers an audible and visual alarm and activates the automatic sprinkler system, with an alarm response time ≤200ms. The audible and visual alarm module includes a 105-decibel buzzer and an LED strobe light, with a coverage radius ≥50 meters. The sprinkler system uses a high-pressure fine water mist system with atomized particle diameter ≤200μm and a flow rate ≥30L / min. The response time from confidence level determination to action execution delay is ≤150ms. Multi-level early warning enables rapid response, high-pressure fine water mist provides efficient fire suppression, and system-level linkage ensures timely triggering of emergency measures, reducing fire losses.

[0015] This invention provides a fire model training method and recognition technology based on image recognition, which has the following beneficial effects:

[0016] This scheme effectively extracts local flame texture and global semantic features through a multi-scale feature fusion mechanism, combining convolutional neural networks and the Transformer architecture, significantly improving recognition accuracy in complex backgrounds. A temporal feature fusion module is introduced, utilizing a bidirectional long short-term memory network to associate the flame diffusion patterns of consecutive frames, addressing the lag in response to dynamic changes inherent in traditional methods. An adversarial training strategy, optimized through a game between the generator and discriminator, enhances the model's robustness against noise and camouflage interference. Experiments show a 42% reduction in false alarm rate under conditions of strong light and smoke obstruction.

[0017] Transfer learning combined with entropy regularization technology enables cross-scene generalization with limited labeled data. The model achieves recognition accuracy exceeding 91% in various scenarios, including forests, industrial areas, and urban buildings. In terms of hardware adaptation, the lightweight inference engine employs INT8 quantization and TensorRT acceleration, maintaining a 50ms real-time response even after deployment to edge devices, while reducing power consumption by 67% compared to traditional solutions. System-level innovation lies in multimodal data collaborative processing. By fusing temperature-sensitive frequency bands from infrared thermal imaging and visible light images, and combining climate data and geographic information, a comprehensive judgment model is constructed, effectively distinguishing between real fires and environmental interference, providing timely support for fire safety decisions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the fire model training process of the present invention;

[0019] Figure 2 This is a flowchart of the real-time fire identification and early warning process of this invention;

[0020] Figure 3 This is the system maintenance and upgrade process of the present invention. Detailed Implementation

[0021] How to use

[0022] This technical solution achieves accurate fire identification and rapid response through multi-dimensional data fusion and intelligent algorithm collaboration. System deployment requires the configuration of a dual-modal sensor module, an edge computing unit, and a central control terminal, forming a closed-loop link from data acquisition to decision execution.

[0023] Data Acquisition and Preprocessing Stage: The system simultaneously captures environmental images using infrared thermal imaging and visible light cameras, covering natural fire sources, high-temperature industrial areas, and simulated fire scenarios. The acquired raw data undergoes denoising and illumination equalization to eliminate environmental interference while preserving flame morphology and temperature-sensitive characteristics. The preprocessing workflow employs an adaptive filtering algorithm to dynamically adjust noise suppression intensity, combined with color space conversion to enhance flame color characteristics, providing high-quality input for subsequent analysis.

[0024] Model training phase: A hybrid feature extraction network is constructed to simultaneously process local texture and global semantic information. A dynamic attention mechanism is used to focus on the flame edges and diffusion regions, and a temporal modeling module is used to analyze the morphological changes between consecutive frames. An adversarial mechanism is introduced during training, using a generator to simulate interference features and a discriminator to optimize feature discrimination capabilities, enhancing the model's adaptability to complex backgrounds. A transfer learning strategy freezes the lower-level feature extraction layers, fine-tuning only the higher-level semantic parameters, and combines entropy regularization constraints to improve cross-scene generalization performance.

[0025] Real-time detection phase: A lightweight inference engine is deployed to compress the trained model to edge devices. Dual-modal data, after feature compression and encoding, is input into the model to extract flame-sensitive frequency band information in real time. Dynamic feature vectors are calculated based on spatiotemporal correlation to generate a comprehensive feature description including diffusion velocity and morphological changes. The system determines the fire status based on multi-level confidence thresholds; low confidence triggers a primary warning, while high confidence triggers the fire control system to execute an emergency response.

[0026] Early warning and response phase: Multi-level early warning modules activate audible and visual alarms, emergency broadcasts, and sprinkler systems based on confidence levels. The fire alarm linkage interface transmits real-time fire location and intensity data, controlling the sprinkler system to precisely cover the fire source area. The system supports manual verification; operators can retrieve detection results and historical records through a visual interface and remotely initiate fire suppression procedures after confirming the fire.

[0027] System maintenance and upgrades: The training dataset is updated regularly to cover new types of fire sources and environmental interference scenarios. The model optimization module supports online incremental training, adapting to environmental changes through incremental adversarial learning. Hardware maintenance processes include sensor calibration and heat dissipation system testing to ensure data acquisition stability. This solution automates the entire fire identification process through algorithm-hardware co-design, maintaining high reliability and low latency response in complex scenarios, providing intelligent support for fire emergency response.

[0028] Example

[0029] Example 1: Fire Monitoring in Complex Industrial Plant Environments

[0030] This system is deployed in a metal processing workshop, which contains high-temperature furnaces, metal dust, and high-frequency mechanical vibration interference. The system uses an infrared thermal imaging camera to capture abnormal surface temperatures on equipment, while a visible light camera simultaneously acquires flame morphology features. To address dust obscuring issues, the preprocessing module employs adaptive Gaussian filtering and the CLAHE enhancement algorithm to eliminate image blurring caused by dust particles while preserving the dynamic texture of molten metal splashes. In the multi-scale feature extraction module, the CNN layer expands the receptive field through dilated convolutions, enabling the identification of features as small as 0.5cm. 2 Metal sparks are generated; the Transformer encoder analyzes the flame propagation path, and combined with Bi-LSTM temporal modeling, the risk of oil stain ignition is predicted. During the adversarial training phase, the generator simulates smoke interference features, and the discriminator uses a PatchGAN structure to distinguish real flames from dust reflections, enabling the model to operate at dust concentrations >150 mg / m³. 3 The system maintains a 92% recognition accuracy rate. When it detects localized overheating (>800℃) at the equipment base accompanied by smoke diffusion, the system triggers a three-level warning: first, it alerts the operator with an audible and visual alarm; second, it activates the workshop smoke exhaust system; and finally, it controls the automatic fire extinguishing device to spray fine water mist. The entire response process takes 180ms.

[0031] Example 2: Early Warning of Forest Fires at the Forest Edge

[0032] Smart poles equipped with dual-modal infrared and visible light cameras were deployed along the forest boundary, covering an area with a radius of 5 kilometers. The system uses an improved SPIHT algorithm to compress and encode infrared images, retaining surface temperature information in the 6-8μm band. Simultaneously, an improved YOLO model is used to identify human activity hotspots (such as unextinguished campfires). The training dataset contains 2000 hours of forest monitoring video, covering different vegetation types and meteorological conditions. A dynamic attention mechanism focuses on anomalous light spots in the canopy layer, combined with Bi-LSTM analysis of flame spread rate. When a continuous rise in surface temperature and a flame height growth rate > 0.3 m / s are detected, a potential fire is identified. The early warning module sends coordinate information to the patroller's terminal via a LoRa network, while simultaneously utilizing a drone equipped with a hyperspectral sensor for aerial verification. Under strong backlight conditions, the system uses the HSV color space to separate the flame from the sky background, avoiding misinterpreting sunset reflections as fires. In actual testing, the false alarm rate was less than 3% when visibility was < 50 meters.

[0033] Example 3: Rapid Response to Fires in Urban Underground Parking Garages

[0034] To address the challenges of rapid smoke diffusion and low visibility in enclosed underground parking garages, the system integrates multispectral sensors and edge computing units. The infrared camera employs an uncooled detector, using a multi-frame differential method to eliminate vehicle exhaust interference. The visible light camera is equipped with a global shutter sensor to capture the electrical sparks generated during vehicle startup. In the preprocessing stage, an improved Retinex algorithm enhances the flame outline within the smoke, while a dynamic attention mechanism adaptively adjusts the region of interest using deformable convolutional kernels. A temporal feature fusion module analyzes the vehicle's battery pack temperature curve, triggering a primary warning when a single battery temperature exceeds 85°C and the temperature difference between adjacent frames is greater than 15°C. During adversarial training, the generator simulates vehicle headlights and dashboard reflections, while the discriminator enhances flame flicker feature recognition through feature matching loss. Upon confirmation of a fire, the system activates fire sprinklers and exhaust fans, dynamically adjusting the sprinkler coverage based on the fire source location to ensure uniform water mist coverage within a 2-meter radius of the burning vehicle. The entire response process is completed within 120 seconds.

[0035] Example 4: Fire Prevention and Control in the Enclosed Space of Ship Engine Room

[0036] An electromagnetic interference-resistant monitoring system was deployed in the engine room of an ocean-going cargo ship. Shielded cables were used to connect infrared and visible light cameras to prevent electromagnetic noise from the frequency converter from affecting image quality. The preprocessing module used a nonlocal means algorithm to eliminate reflections from the metal bulkheads, while a dynamic attention mechanism focused on the high-temperature area of ​​the diesel engine exhaust pipe. Training data included 20 types of fault scenarios, such as hydraulic oil leaks and cable short circuits in the engine room. The model learned the environmental characteristics of different compartments through transfer entropy regularization. When an abnormally high exhaust pipe temperature (>450℃) accompanied by lubricating oil splashing was detected, the system initiated a three-level response: first, a vibration motor alerted the crew; second, the fuel supply valve was shut off; and finally, the high-pressure fine water mist system in the engine room was activated. The warning information was simultaneously transmitted to the captain's cabin and the shore-based monitoring center, supporting remote video verification and emergency command. Experiments showed that the system maintained stable operation even in sea state 8, with a false alarm rate controlled below 1.5%.

[0037] Example 5: Collaborative Fire Prevention and Control in Ancient Building Complexes

[0038] Distributed monitoring nodes were deployed across a cluster of ancient wooden buildings. Each node included an infrared thermal imager and a visible light camera, transmitting data back to an edge computing center via 5G. The system employed a federated learning framework to enable collaborative training of multi-node models while protecting privacy. The feature extraction module shared cross-building texture features. For complex structures such as eaves and brackets, a dynamic attention mechanism adjusted the receptive field through deformation convolution kernels to accurately identify incense burner embers and overheated electrical wiring. A temporal analysis module established a flame propagation dynamics model to predict the fire's spread path between wooden beams and columns. When an abnormal temperature area was detected, the system prioritized activating fire-resistant curtains inside the building while simultaneously sending 3D thermal maps and building structural data to the fire department. In a 2025 fire at an ancient temple in Shanxi Province, this system provided an 11-minute advance warning, buying crucial time for the rescue of cultural relics.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fire model training method and recognition technology based on image recognition, characterized in that: Includes the following steps: Data acquisition and preprocessing: Acquire a dataset of fire images containing different environmental scenarios, including images of natural fire sources, industrial fire sources and simulated fire scenarios. Perform noise reduction, illumination equalization and color space normalization on the images to generate standardized training data. Multi-scale feature extraction module construction: A hybrid feature extraction network including convolutional neural network (CNN) and Transformer architecture is designed. Local texture features and global semantic features of the image are extracted through multi-scale feature pyramid, and a dynamic attention mechanism is introduced to enhance the flame shape features. Temporal feature fusion training: Input continuous frame images into the feature extraction module, and use a bidirectional long short-term memory network (Bi-LSTM) to model the temporal features in relation to each other. Combine the spatiotemporal correlation between flame spread speed and morphological changes to generate dynamic feature vectors. Adversarial training optimization: Generative Adversarial Network (GAN) is used to train the feature extraction module adversarially. The discriminator distinguishes between real fire features and interference features, and the feature discrimination ability of the generator is iteratively optimized. Model generalization enhancement: During training, a transfer learning strategy is introduced to freeze the low-level feature extraction layer of the pre-trained model and only fine-tune the high-level semantic feature layer. Combined with the transfer entropy regularization loss function, the robustness of the model to complex backgrounds is improved.

2. The fire model training method and recognition technology based on image recognition according to claim 1, characterized in that: The data preprocessing described in step 1 includes: The image is denoised using adaptive Gaussian filtering, with the noise standard deviation adaptively adjusted based on the local gradient of the image. The CLAHE algorithm is used to enhance the contrast of the image by segmentation, with each segment being 8×8 pixels. Convert the RGB color space to the HSV color space and perform histogram equalization on the saturation channel.

3. The fire model training method and recognition technology based on image recognition according to claim 1, characterized in that: The multi-scale feature extraction module in step 2 includes: The initial CNN layer extracts local features at three scales: 4×4, 8×8, and 16×16. The Transformer encoder performs self-attention calculation on global features, and the dimensions of the query, key, and value matrices are 64×64×256 respectively. The dynamic attention mechanism adjusts the receptive field through deformable convolution kernels, and the deformation parameters are adaptively generated by the gradient direction of the flame edge.

4. The fire model training method and recognition technology based on image recognition according to claim 1, characterized in that: Step 3, the temporal feature fusion training, includes: Set the number of hidden layer units of Bi-LSTM to 256, and the time step to the reciprocal of the video frame rate; The spatiotemporal correlation calculation adopts a weighted fusion function of cosine similarity and Euclidean distance, and the weight coefficients are optimized through backpropagation of cross-entropy loss; The dynamic feature vector has a dimension of 512×T, where T is the length of the continuous frame sequence.

5. The fire model training method and recognition technology based on image recognition according to claim 1, characterized in that: The adversarial training optimization described in step 4 includes: The generator loss function is L_GAN = λ_adv·L_adv + λ_feat·L_feat, where λ_adv = 0.5 and λ_feat = 0.3; The discriminator uses a PatchGAN structure to perform binary classification on image blocks, with a block size of 16×16. The training iterations consisted of 100 epochs, with a cosine annealing learning rate of 0.001 and a final learning rate of 0.00001.

6. The fire model training method and recognition technology based on image recognition according to claim 1, characterized in that: The migration entropy regularization loss function described in step 5 is defined as follows: Where pi is the activation probability of the i-th neuron in the transfer feature layer, N is the total number of neurons, and the regularization coefficient γ = 0.

01.

7. The fire model training method and recognition technology based on image recognition according to claims 1-6, characterized in that: Fire detection technology, based on models obtained through training methods, includes: Real-time image acquisition module: Simultaneously acquires environmental images via an infrared thermal imaging camera and a visible light camera; Feature compression encoder: Performs JPEG2000 lossy compression on the input image, with a compression ratio ≥20:1, while preserving the flame-sensitive frequency band; Lightweight inference engine: It adopts the TensorRT acceleration framework to quantize the model to INT8 precision, with inference latency ≤50ms; Multi-level early warning module: Three alarm thresholds are set according to the flame confidence score, with threshold ranges of 0.6, 0.8 and 0.95 respectively.

8. The fire model training method and recognition technology based on image recognition according to claim 7, characterized in that: The infrared thermal imaging camera operates in the 8-14μm band, has a spatial resolution of 640×512 pixels, and a frame rate of 30fps; the visible light camera uses a global shutter CMOS sensor with a quantum efficiency of ≥70%.

9. The fire model training method and recognition technology based on image recognition according to claim 7, characterized in that: The feature compression encoder uses an improved SPIHT algorithm to adaptively encode the flame region ROI in blocks, with a block size of 16×16 or 32×32.

10. The fire model training method and recognition technology based on image recognition according to claim 7, characterized in that: The multi-level early warning module is linked to the fire control system. When the confidence level is ≥0.95, it triggers an audible and visual alarm and starts the automatic sprinkler system. The alarm response time is ≤200ms.

Images