Constrained information flow identification method and system for early small target fire source of fire

Abstract

This invention relates to the field of fire detection technology, specifically a constrained information flow identification method and system for small-scale fire sources in the early stages of a fire. It includes the following steps: S1, constrained multi-scale feature extraction for small-scale fire sources; S2, selective lightweight convolutional feature processing based on fire source correlation; S3, fire source feature enhancement mechanism based on channel response concentration; S4, residual fusion strategy triggered by feature response conditions; S5, modular feature reuse and information flow control for stable perception of small targets; S6, multi-scale prediction output based on the fire source scale distribution characteristics. Compared with existing technologies, by constraining the design of feature extraction, lightweight computation, feature enhancement, and fusion strategies, the accuracy and stability of detecting small-sized flames and initial smoke targets are improved while maintaining model lightweightness and real-time inference capabilities.

Description

Technical Field

[0001] This invention relates to the field of fire identification technology, specifically to a constrained information flow identification method and system for small target fire sources in the early stages of a fire. Background Technology

[0002] In applications such as video surveillance, industrial production, and public safety, early-stage fires often manifest only as small-scale, short-duration, and unstable flames or initial smoke targets. These targets occupy a very small percentage of pixels in the monitored image, and their discriminative information is mainly reflected in local brightness, color energy, and short-term dynamic changes. They lack obvious structural outlines and are therefore easily obscured by background information in the image feature space.

[0003] In real-world engineering environments, monitoring footage is often accompanied by complex lighting changes, specular reflections, welding sparks, high-temperature equipment light emission, and other high-energy interference sources. This makes it difficult for fire detection methods based on thresholds, rules, or single visual features to reliably distinguish between real fire sources and background interference, resulting in long-standing problems of high false alarm rates and poor generalization ability.

[0004] In recent years, deep learning-based target detection methods have achieved good results in the field of general target recognition. However, their network structure design is usually geared towards medium- to large-scale targets, with low dependence on feature resolution and information integrity. In the process of lightweighting models and accelerating inference for real-time monitoring scenarios, networks often reduce computational complexity by continuously downsampling, channel compression, or simplifying feature fusion paths. This leads to small-scale fire source targets being structurally compressed or weakened during the feature extraction stage, resulting in a significant increase in the false negative rate of lightweight models in early fire scenarios.

[0005] Among existing early fire warning technologies, there are various solutions based on image recognition or multi-source information fusion. For example, CN121438483A discloses an early fire warning system based on AI image recognition. This system analyzes monitoring video streams in real time and uses a pre-trained model to extract visual features related to flames or smoke, thus providing early warning in the initial stages of a fire. This type of solution achieves fire recognition through "video + AI" at the system functional level, but its technical focus is on whether the fire target is detected. It does not specifically design the feature extraction and information transmission mechanisms within the detection model to address the problems of small flame or smoke targets, extremely low pixel ratio, and susceptibility to compression by network structures in the early stages of a fire. Another example is CN120495836A, which proposes an automatic fire detection system based on image recognition. This system achieves fire recognition through image preprocessing, fire feature extraction, and machine learning classification. This type of technology mainly relies on manual design or a combination of traditional visual features and classification models. Its improvement focuses on feature types and classification strategies, making it difficult to stably identify small-scale fire source targets in the early stages of a fire under complex lighting conditions, high-brightness interference, and lightweight deployment. Furthermore, CN119314302A discloses a fire early warning method based on multi-source information fusion. This method integrates multi-source data such as vehicle images, location data, and temperature data, and combines this with BIM modeling to achieve fire risk assessment and visualization. While this approach improves the reliability of fire early warning by introducing external sensor information and system-level modeling, its enhanced detection capability relies on multi-source data fusion and macro-scene analysis. It does not address the issue of easily lost small-scale fire source features in lightweight networks from the perspective of the visual detection model's internal structure. In summary, existing technologies generally suffer from the following shortcomings: First, they primarily focus on system function implementation or external information fusion, lacking specific optimization for addressing the structural weakening of small-scale fire source targets within deep learning models during the early stages of a fire. Second, in pursuing real-time performance and lightweight deployment, they fail to constrain the impact of feature downsampling, channel compression, and feature fusion paths on the perception capability of small targets.

[0006] Furthermore, in scenarios involving concurrent processing of multiple video streams or deployment of edge computing devices, simply increasing the model size or introducing multi-source sensor information to improve detection accuracy often significantly increases system complexity and deployment costs, making it difficult to meet the requirements of engineering applications for real-time performance, stability, and long-term operation under resource-constrained conditions. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides a target recognition method optimized from the perspective of network structure and information flow transmission mechanism for small target fire source detection scenarios. By implementing constrained design of feature extraction, lightweight computation, feature enhancement, and fusion strategies, the method improves the detection accuracy and stability of small-sized flames and initial smoke targets while maintaining the model's lightweight nature and real-time inference capabilities.

[0008] To achieve the above objectives, a constrained information flow identification method for small target fire sources in the early stages of a fire is designed, including the following steps: S1, Constrained multi-scale feature extraction for small-scale fire sources: Layered feature extraction is performed on the input fire monitoring image, limiting the downsampling ratio of high-rise features during the feature extraction process; at the same time, high-resolution spatial information in shallow features is preserved. S2, Selective lightweight convolutional feature processing based on fire source correlation: In the feature extraction process, lightweight convolutional units are introduced to group channels according to the intensity of flame or smoke response in the feature channels; S3, a fire source feature enhancement mechanism based on channel response concentration: Perform channel dimension analysis on the feature map output in step S2. Calculate the concentration of channel responses by statistically analyzing the response distribution of each channel to the flame or smoke target, and generate corresponding channel attention weights accordingly. S4, Residual fusion strategy triggered by feature response conditions: Based on the channel attention analysis results and feature space distribution, determine whether the current feature meets the discrimination condition for small target fire source. When the condition is met, perform residual fusion between the input feature and the output feature after channel enhancement. S5, modular feature reuse and information flow control for stable perception of small targets: Multiple lightweight feature enhancement units are combined in a modular manner, and cross-layer and cross-branch feature reuse paths are built within the modules, so that small target features that have been identified as fire source related remain accessible in multiple feature levels. S6, Multi-scale prediction output based on fire source-scale distribution characteristics: The location and category information of the fire source target are output based on feature maps with different spatial resolutions, and higher spatial sensitivity is set for small-scale flames and initial smoke targets.

[0009] In step S2, the complete spatial feature extraction process is retained for channels with high fire correlation, while low correlation channels are processed using a lightweight mapping method with dimensionality reduction or shared parameters.

[0010] In step S3, when a feature response is detected to exhibit a pattern of high-intensity activation in a few channels and weaker response in the remaining channels, the high-response channels are weighted and enhanced.

[0011] The criteria for identifying small target fire sources in step S4 include: the response concentration of the fire source-related channels exceeds a preset threshold, and the proportion of the spatially activated region to the feature map is lower than a preset range.

[0012] To achieve the above objectives, a system for constrained information flow identification of small target fire sources in the early stages of a fire is designed, including: The constrained multi-scale feature extraction module is used for hierarchical feature extraction, which limits the downsampling ratio, preserves shallow high-resolution spatial details, prevents small-scale fire source features from being structurally compressed, and outputs multi-scale basic feature maps. A lightweight convolutional module driven by fire source correlation is used to analyze the channel response intensity, reduce redundant computation, retain discrimination information, and output selective lightweight feature maps. The feature enhancement module based on channel response concentration is used to statistically analyze the channel response distribution, calculate the response concentration, enhance the high-energy concentrated fire source-related channels, suppress background and interference responses, and output a fire source feature enhancement map. The condition-triggered residual feature fusion module is used to determine whether the fire source features meet the preset conditions. If the conditions are met, residual fusion is enabled; if the conditions are not met, invalid residuals are blocked, shallow details of small targets are preserved, and a stable fusion detection feature map is output. The modular feature reuse and information flow control module is used to modularly combine multiple feature enhancement units to construct cross-layer or cross-branch reuse paths and output multi-level reusable fire source features. The scale-sensitive multi-scale detection output module outputs the location and category information of flames or smoke.

[0013] Compared with the prior art, the present invention has the following advantages: 1. Significantly improve the detection reliability of small target fire sources under strictly controlled computational complexity: By constraining the feature downsampling ratio, lightweight computation channels and residual fusion conditions, small flames and initial smoke are not easily compressed or covered in the feature space. Thus, while maintaining low parameter quantity and low computational complexity, the detection accuracy of small targets in the early stage of fire is significantly improved, rather than relying on simply increasing the network size to achieve performance improvement. 2. Effectively reduce the risk of missed detection of lightweight target detection models in early fire scenarios: By using a feature enhancement mechanism based on channel response concentration and a condition-triggered residual fusion strategy, fire source-related features are given priority enhancement and multi-level retention in the network, avoiding them from being submerged by background features in the lightweight structure, thereby reducing the probability of missed detection of small target fire sources in the early stage from the mechanism level. 3. Maintain stable real-time performance in multi-channel video concurrency and edge computing environments: Since only the channels with high correlation to the fire source are subjected to relatively complete spatial feature calculations, while the other channels are processed using a lightweight mapping method, the model can still maintain a stable frame rate in multi-channel video concurrency processing and edge computing devices with limited computing power, thus meeting the real-time requirements of the fire monitoring system. 4. Improve detection robustness under complex lighting and background conditions: By analyzing and selectively enhancing the response distribution of feature channels, the impact of complex lighting changes, dynamic backgrounds or high-brightness interference on the detection results is reduced, making the model focus more on the discriminative features of flames and smoke themselves, thereby improving the overall robustness in complex monitoring environments. 5. Excellent structural reusability and engineering scalability: It adopts a modular feature enhancement and information flow control structure, which can be embedded as an independent functional module into different types of target detection networks. It can enhance the detection capability of small target fire sources without destroying the original network architecture, and is easy to reuse and expand in different application scenarios. Detailed Implementation

[0014] The present invention will now be further described.

[0015] The constrained information flow identification method for small target fire sources in the early stage of a fire in this embodiment includes the following steps: S1, Constrained multi-scale feature extraction for small-scale fire sources: Layered feature extraction is performed on the input fire monitoring image. During the feature extraction process, the downsampling ratio of high-level features is limited to prevent small-scale flames and initial smoke from being over-compressed during continuous downsampling. At the same time, high-resolution spatial information in shallow features is preserved so that small target fire sources still maintain a distinguishable spatial support region in the feature space, thus structurally ensuring the perceptibility of small targets, rather than relying solely on subsequent detection head compensation. S2, Selective lightweight convolutional feature processing based on fire source correlation: Lightweight convolutional units are introduced during feature extraction. Instead of performing uniform separate calculations on all feature channels, channels are grouped according to the intensity of the flame or smoke response in the feature channels. S3, a fire source feature enhancement mechanism based on channel response concentration: Perform channel dimension analysis on the feature map output in step S2. Calculate the concentration of channel responses by statistically analyzing the response distribution of each channel to the flame or smoke target, and generate corresponding channel attention weights accordingly. S4, Residual fusion strategy triggered by feature response conditions: Based on the channel attention analysis results and feature space distribution, it is determined whether the current feature meets the discrimination condition for small target fire source. When the condition is met, the input feature and the output feature after channel enhancement are subjected to residual fusion to selectively retain the details in the shallow layer that are crucial for the discrimination of small target fire source, and avoid introducing invalid residual interference in large target or background-dominated scenarios. S5, modular feature reuse and information flow control for stable perception of small targets: Multiple lightweight feature enhancement units are combined in a modular manner, and cross-layer and cross-branch feature reuse paths are built within the modules, so that small target features that have been identified as fire source related remain accessible in multiple feature levels. S6, Multi-scale prediction output based on fire source-scale distribution characteristics: The system outputs the location and category information of fire source targets based on feature maps with different spatial resolutions. It also sets higher spatial sensitivity for small-scale flames and initial smoke targets, enabling them to be detected and confirmed first in the high-resolution feature layer. This achieves simultaneous detection of fire source targets of different sizes and significantly improves the identification accuracy in the early stage of a fire.

[0016] In step S2, the complete spatial feature extraction process is retained for channels with high fire correlation, while low correlation channels are processed by dimensionality reduction or lightweight mapping with shared parameters, thereby reducing the overall computational complexity without significantly weakening the small target fire source discrimination information.

[0017] In step S3, when a feature response is detected to exhibit a feature pattern of high-intensity activation in a few channels and weak response in the remaining channels, the high-response channels are weighted and enhanced to highlight the energy concentration characteristics of the fire source target in the feature space, thereby compensating for the decrease in expressive power that may result from lightweight computation.

[0018] The criteria for identifying small target fire sources in step S4 include: the response concentration of the fire source-related channels exceeds a preset threshold, and the proportion of the spatially activated region to the feature map is lower than a preset range.

[0019] Step S5 extends the effective lifespan of small target fire source features in the network by constraining the feature transmission path, preventing them from being rapidly decayed or covered in a single scale or stage, thereby improving the stability of small target fire source detection from a structural level.

[0020] To achieve the above objectives, a system for constrained information flow identification of small target fire sources in the early stages of a fire is designed, including: The constrained multi-scale feature extraction module is used for hierarchical feature extraction, which limits the downsampling ratio, preserves shallow high-resolution spatial details, prevents small-scale fire source features from being structurally compressed, and outputs multi-scale basic feature maps. A lightweight convolutional module driven by fire source correlation is used to analyze the channel response intensity, reduce redundant computation, retain discrimination information, and output selective lightweight feature maps. The feature enhancement module based on channel response concentration is used to statistically analyze the channel response distribution, calculate the response concentration, enhance the high-energy concentrated fire source-related channels, suppress background and interference responses, and output a fire source feature enhancement map. The condition-triggered residual feature fusion module is used to determine whether the fire source features meet the preset conditions. If the conditions are met, residual fusion is enabled; if the conditions are not met, invalid residuals are blocked, shallow details of small targets are preserved, and a stable fusion detection feature map is output. The modular feature reuse and information flow control module is used to modularly combine multiple feature enhancement units to construct cross-layer or cross-branch reuse paths and output multi-level reusable fire source features. The scale-sensitive multi-scale detection output module outputs the location and category information of flames or smoke.

[0021] Small target flames and initial smoke are small in size, have limited spatial support areas, and are easily submerged by the background in images. The method of this invention is based on the characteristics of fire source targets in feature space, which are characterized by concentrated channel response, limited spatial support areas, and unstable morphology. It conducts a constrained design on the feature extraction and information flow transmission mechanism in the target detection network to improve the perceptibility and recognition stability of small target fire sources.

[0022] By preserving shallow, high-resolution features during multi-scale feature extraction, and differentiating channels based on their response intensity to flame or smoke targets, a selective computation strategy based on channel response intensity is employed. This strategy involves retaining relatively complete spatial feature calculations only for channels with high fire relevance, while using lightweight mapping methods such as sharing or dimensionality reduction for low-relevance channels. This reduces computational complexity while ensuring the integrity of fire source discrimination information, allowing the network to focus on enhancing its ability to discriminate small target flames and initial smoke while maintaining overall computational complexity. Channel weights are generated by analyzing the concentration of feature channel responses, and channels exhibiting high-energy concentration features are selectively enhanced to highlight the discrimination information of small target fire sources in the feature space, thereby compensating for the decrease in expressive power caused by lightweight feature processing.

[0023] Meanwhile, a conditional residual feature fusion mechanism is introduced, which allows shallow detail features to participate in deep feature fusion only when the detected fire source-related features have high channel concentration and small spatial activation area. This selectively preserves the spatial detail information of small target fire sources and avoids introducing invalid residual interference in background-dominated or large target scenes, thereby avoiding the masking of small target fire source information by large-scale target features.

[0024] By modularly combining multiple feature enhancement units and constructing cross-layer or cross-branch feature reuse paths within the modules, fire source-related features remain accessible across multiple feature levels. This enables the multiple utilization and stable transmission of small target fire source features within the network, extending the effective lifespan of small target fire source features in the network and thus significantly improving the detection reliability and real-time performance in the early stages of a fire.

[0025] This invention does not focus on system architecture design, fire function process improvement, or multi-source information fusion. Instead, it addresses the scenario of detecting small-scale fire sources in the early stages of a fire by starting from the structure and information flow transmission mechanism of the target detection network. It solves the problem that small-scale flames and initial smoke targets in the lightweight model are easily downsampled and compressed in calculation.

[0026] Compared with existing fire early warning technologies that rely on system implementation, multi-source information fusion, or traditional image feature recognition, this invention imposes structural constraints on the feature downsampling ratio, channel calculation method, and feature enhancement and fusion triggering conditions. This allows small-scale fire source targets to obtain stable and sustainable effective representation in the network while maintaining the model's lightweight design and real-time inference capabilities, thereby significantly improving the detection accuracy and reliability in the early stages of a fire.

Claims

1. A constrained information flow identification method for small target fire sources in the early stage of a fire, characterized by: Includes the following steps: S1, Constrained multi-scale feature extraction for small-scale fire sources: Layered feature extraction is performed on the input fire monitoring image, limiting the downsampling ratio of high-rise features during the feature extraction process; at the same time, high-resolution spatial information in shallow features is preserved. S2, Selective lightweight convolutional feature processing based on fire source correlation: In the feature extraction process, lightweight convolutional units are introduced to group channels according to the intensity of flame or smoke response in the feature channels; S3, a fire source feature enhancement mechanism based on channel response concentration: Perform channel dimension analysis on the feature map output in step S2. Calculate the concentration of channel responses by statistically analyzing the response distribution of each channel to the flame or smoke target, and generate corresponding channel attention weights accordingly. S4, Residual fusion strategy triggered by feature response conditions: Based on the channel attention analysis results and feature space distribution, determine whether the current feature meets the discrimination condition for small target fire source. When the condition is met, perform residual fusion between the input feature and the output feature after channel enhancement. S5, modular feature reuse and information flow control for stable perception of small targets: Multiple lightweight feature enhancement units are combined in a modular manner, and cross-layer and cross-branch feature reuse paths are built within the modules, so that small target features that have been identified as fire source related remain accessible in multiple feature levels. S6, Multi-scale prediction output based on fire source-scale distribution characteristics: The location and category information of the fire source target are output based on feature maps with different spatial resolutions, and higher spatial sensitivity is set for small-scale flames and initial smoke targets.

2. The constrained information flow identification method for small target fire sources in the early stage of a fire, as described in claim 1, is characterized in that: In step S2, the complete spatial feature extraction process is retained for channels with high fire correlation, while low correlation channels are processed using a lightweight mapping method with dimensionality reduction or shared parameters.

3. The constrained information flow identification method for small target fire sources in the early stage of a fire, as described in claim 1, is characterized in that: In step S3, when a feature response is detected to exhibit a pattern of high-intensity activation in a few channels and weaker response in the remaining channels, the high-response channels are weighted and enhanced.

4. The constrained information flow identification method for small target fire sources in the early stage of a fire, as described in claim 1, is characterized in that: The criteria for identifying small target fire sources in step S4 include: the response concentration of the fire source-related channels exceeds a preset threshold, and the proportion of the spatially activated region to the feature map is lower than a preset range.

5. The system for constrained information flow identification of small target fire sources in the early stage of a fire, according to any one of claims 1-4, is characterized in that: include: The constrained multi-scale feature extraction module is used for hierarchical feature extraction, which limits the downsampling ratio, preserves shallow high-resolution spatial details, prevents small-scale fire source features from being structurally compressed, and outputs multi-scale basic feature maps. A lightweight convolutional module driven by fire source correlation is used to analyze the channel response intensity, reduce redundant computation, retain discrimination information, and output selective lightweight feature maps. The feature enhancement module based on channel response concentration is used to statistically analyze the channel response distribution, calculate the response concentration, enhance the high-energy concentrated fire source-related channels, suppress background and interference responses, and output a fire source feature enhancement map. The condition-triggered residual feature fusion module is used to determine whether the fire source features meet the preset conditions. If the conditions are met, residual fusion is enabled; if the conditions are not met, invalid residuals are blocked, shallow details of small targets are preserved, and a stable fusion detection feature map is output. The modular feature reuse and information flow control module is used to modularly combine multiple feature enhancement units to construct cross-layer or cross-branch reuse paths and output multi-level reusable fire source features. The scale-sensitive multi-scale detection output module outputs the location and category information of flames or smoke.