Multi-degradation flame image recovery method and system

Abstract

The invention discloses a multi-degradation flame image recovery method and system, and the method comprises the steps: collecting a cable fire smoke sample and a flame real image in real time through simulating the real underground pipe gallery high-voltage cable fire condition, constructing an atmospheric scattering model, and combining the characteristics of the cable fire smoke sample, and obtaining a flame image of the cable fire. Generating smoke degradation images under various working conditions based on the flame real image; a smog image enhancement model based on an improved DCGAN architecture is constructed and trained, the smog image enhancement model comprises a generator and a discriminator, the generator is used for reconstructing a smog degradation image to generate a flame restoration image, and the discriminator is used for carrying out authenticity judgment on the flame restoration image and a real flame image. And the generator is ensured to continuously optimize the difference between the flame restoration image and the flame real image. According to the method, dense smoke interference is effectively removed, shielded image details are recovered, and clear and reliable visual information is provided for subsequent fire identification.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a method and system for restoring multi-degraded flame images. Background Technology

[0002] Underground cable tunnels pose a significant fire hazard, seriously threatening power grid security and the normal operation of cities. Therefore, accurate fire identification and real-time monitoring are crucial. However, traditional sensor monitoring methods are susceptible to interference in complex environments and lack accuracy.

[0003] Existing research primarily focuses on image-based fire identification methods, which are highly dependent on image quality. Furthermore, the dense smoke generated by cable tunnel fires tends to linger in the space, causing severe occlusion and loss of detail in the monitoring images, leading to a decline in image quality and interfering with subsequent image recognition and image-based decision-making. Therefore, there is currently a lack of technical methods to address image occlusion issues in cable tunnel fire smoke scenarios.

[0004] For example, Chinese patent CN118314675A discloses a method and system for accurate identification of multi-composite smoke detectors. The method includes the following steps: using sensors to sample smoke particle size data, temperature change rate, and carbon monoxide gas concentration data of the current environment; inputting the acquired smoke particle size data, temperature change rate, and carbon monoxide gas concentration data into a multi-dimensional data identification and judgment matrix for identification processing; obtaining the identification result, and performing corresponding processing according to the identification result. The temperature sensor in this invention is easily damaged in high-temperature environments, leading to false alarms or failure. Chinese patent CN113221715A discloses a fire detection and identification method that integrates a visual attention mechanism. It incorporates a saliency attention mechanism into fire detection and identification to remove a large amount of redundant data from the original fire image, quickly determine suspected fire areas from complex fire image information, and finally use an SVM support vector machine for fire identification to obtain feature values, thereby achieving fire identification and determination. This invention relies on Otsu segmentation and morphological operations to extract suspected areas from the saliency map. In environments filled with dense smoke, such as cable tunnels, image details are severely lost and contrast is reduced. The threshold automatically determined by the Otsu algorithm may not be able to accurately segment effective suspected areas, thus affecting the accuracy of subsequent feature extraction and recognition. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for restoring flame images with multiple degradations, effectively removing dense smoke interference, restoring obscured image details, providing clear and reliable visual information for subsequent fire identification, and ultimately improving the reliability of intelligent fire early warning systems.

[0006] The technical solution of the present invention is as follows: On the one hand, the present invention provides a method for restoring multi-degraded flame images, comprising the following steps: By simulating real-world high-voltage cable fires in underground utility tunnels, real-time smoke samples and flame images of cable fires are collected. An atmospheric scattering model is constructed and combined with the characteristics of smoke samples from cable fires, smoke degradation images under various working conditions are generated based on real flame images. A smoke image enhancement model based on an improved DCGAN architecture is constructed and trained. The smoke image enhancement model includes a generator and a discriminator. The generator is used to reconstruct the degraded smoke image to generate a flame restoration image. The discriminator is used to judge the authenticity of the flame restoration image and the real flame image, ensuring that the generator continuously optimizes the difference between the flame restoration image and the real flame image.

[0007] Preferably, the atmospheric scattering model is expressed as:

[0008] In the formula, Image degraded by smoke; This is a true image of the flame before it deteriorated; For global atmospheric light; The smoke concentration gradient, representing the proportion of light retained during propagation, is expressed as:

[0009] In the formula, The smoke density attenuation coefficient; This is the equivalent distance from the pixel to the imaging device.

[0010] Preferably, an atmospheric scattering model is constructed and combined with the characteristics of smoke samples from cable fires to generate smoke degradation images under various working conditions based on real flame images. Specifically: The operating conditions include: different smoke colors, including white smoke, gray smoke and black smoke, different fire light diffuse reflection conditions and different smoke concentration gradients; Global atmospheric light in atmospheric scattering models With smoke density attenuation coefficient The range of values ​​is limited by parameter combinations, where, under white smoke conditions, The value range is set to 0.75 to 1.00. The value range is 0.03 to 0.15; under gray smoke conditions, The value range is set to 0.55 to 0.85. The value range is 0.3 to 0.6; under black smoke conditions The value range is set to 0.10 to 0.40. The value range is 0.6 to 1.5; set up The value range is 0.40 to 0.60. When the value ranges from 0.8 to 1.2, it indicates a condition with weak flame intensity and significant smoke obstruction; [Setting] The value range is 0.60 to 0.80. A value ranging from 0.6 to 0.8 indicates a condition where the flame intensity is moderate and the smoke is partially illuminated; [Setting] The value range is 0.80 to 1.00. When the value ranges from 0.3 to 0.5, it represents the working condition where smoke particles are exposed to high brightness under strong fire light, resulting in obvious reflection and brightening phenomena. Through the combination of the parameters, the atmospheric scattering model is controlled to generate smoke degradation images under various working conditions based on real flame images.

[0011] Preferably, the generator adopts a progressive multi-stage generator structure, and the generator is internally divided into multi-level sub-generators according to function. Each sub-generator adopts a cascaded structure to form a complete generator. The sub-generator includes a symmetrical encoder and decoder. The encoder extracts multi-scale deep features of the smoke degradation image through continuous convolution and pooling operations. The decoder gradually restores the spatial resolution and generates a flame restoration image through deconvolution or upsampling. The encoder and decoder use a skip connection to directly transfer the shallow detail features of the encoder to the corresponding decoding layer. The discriminator employs a multi-layer convolutional downsampling method to map the input image into a feature space with progressively lower resolution, thereby enhancing the ability to discriminate local statistical features by increasing the number of convolutional kernels. The smoke image enhancement model employs a progressive cumulative training strategy during training, training multiple sub-generators in stages according to the principle of increasing difficulty: the first sub-generator is responsible for recovering the low-frequency structural information of the flame, and subsequent sub-generators continuously supplement texture, highlight details, and edge information based on the output of the previous sub-generator; during training, backpropagation not only optimizes the current sub-generator but also fine-tunes the preceding sub-generators; simultaneously, the discriminator applies adversarial constraints to the intermediate results of the flame reconstruction images generated by each sub-generator against the real flame image, ensuring that each sub-generator in the generator maintains consistency with the real flame image, thereby achieving multi-stage progressive reconstruction from coarse to fine and obtaining a clear flame reconstruction image.

[0012] Preferably, the smoke image enhancement model introduces attention modules in the deep layers of the encoder and the key layers of the decoder, the attention modules including channel attention modules and spatial attention modules; The channel attention module adaptively adjusts the weights of different feature channels, assigning different importance to each feature channel; The spatial attention module generates a spatial weight matrix through several Conv layers, and then multiplies the spatial weight matrix with the original feature map position by position using the Sigmoid activation function to obtain the spatial attention-weighted enhanced feature map.

[0013] Preferably, the smoke image enhancement model uses peak signal-to-noise ratio (PSNR) as an evaluation metric, whereby PSNR is expressed as:

[0014] In the formula, Peak signal-to-noise ratio (PSNR) is used to measure image quality or the fidelity of prediction results. This represents the maximum possible value of the image pixels. This represents the squared error of a single image sample.

[0015] On the other hand, the present invention provides a multi-degraded flame image restoration system, including a smoke-degraded image module and a smoke image enhancement module; The smoke degradation image generation module is used to simulate real-world underground utility tunnel high-voltage cable fires, collect smoke samples and real flame images of cable fires in real time, construct an atmospheric scattering model, and combine the characteristics of smoke samples from cable fires to generate smoke degradation images under various working conditions based on real flame images. A smoke image enhancement module is used to build and train a smoke image enhancement model based on an improved DCGAN architecture. The smoke image enhancement model includes a generator and a discriminator. The generator is used to reconstruct the smoke-degraded image to generate a flame restoration image. The discriminator is used to judge the authenticity of the flame restoration image and the real flame image, ensuring that the generator continuously optimizes the difference between the flame restoration image and the real flame image.

[0016] In another aspect, the present invention also provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any embodiment of the present invention.

[0017] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0018] Compared with the prior art, the present invention has the following technical effects: 1. This invention constructs a degradation model based on an atmospheric scattering model and incorporating the characteristics of real smoke samples. This model can accurately and controllably generate a variety of degradation images that highly match the real pipe gallery fire environment, covering different smoke colors (white / black / gray), different fire light diffuse reflection conditions, and continuous concentration gradients from low to high (20%-100%). This provides a high-quality and highly diverse paired dataset foundation for training a highly generalizable enhancement model, solving the data scarcity problem faced by deep learning in this field.

[0019] 2. This invention proposes an image enhancement model with the fundamental goal of "restoring recognition capability". It uses objective indicators such as loss function and PSNR to evaluate image quality, ensuring that the enhanced image can effectively support fire early warning.

[0020] 3. This invention establishes a complete technical route from image reconstruction to intelligent early warning, providing a stable and reliable visual perception foundation for accurately quantifying fire spread behavior and implementing precise fire fighting, and providing a systematic solution to the image occlusion problem in cable tunnel fire monitoring. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the multi-deterioration flame image restoration method described in this invention; Figure 2 This is the overall architecture diagram of the generator; Figure 3 This is the overall structure diagram of the sub-generator; Figure 4 This is an overall structural diagram of the spatial attention module; Figure 5 It is a graph of three core indicators: MSE, MAE, and PSNR; Figure 6 This is an SSIM trend curve chart; Figure 7 This is a diagram showing the effect of restoring the black smoke working condition. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0023] Example 1 This embodiment provides a method for restoring flame images with multiple degradations. (See also...) Figure 1 As shown, it includes the following steps: By simulating real-world high-voltage cable fires in underground utility tunnels, smoke samples and real flame images of cable fires are collected in real time. An atmospheric scattering model is constructed and combined with the characteristics of smoke samples from cable fires, smoke degradation images under various working conditions are generated based on real flame images.

[0024] A smoke image enhancement model based on an improved DCGAN architecture is constructed and trained. This model includes a generator and a discriminator. The generator reconstructs the smoke-degraded image to generate a flame restoration image, while the discriminator distinguishes between the restored image and the actual flame image, ensuring the generator continuously optimizes the difference between the restored and actual flame images. During training, this embodiment constrains the actual flame image and the smoke-degraded image to be on the same feature plane and have the same target, effectively reducing pixel deviation. Based on real fire scenarios, this model constructs a data system and enhancement model that closely resembles the actual environment, achieving closed-loop optimization from "physical degradation" to "intelligent restoration," significantly improving the restoration quality and reliability of flame images under dense smoke interference.

[0025] In a preferred embodiment of this invention, the atmospheric scattering model is expressed as follows:

[0026] In the formula, Image degraded by smoke; This is a true image of the flame before it deteriorated; For global atmospheric light; The smoke concentration gradient, representing the proportion of light retained during propagation, is expressed as:

[0027] In the formula, The smoke density attenuation coefficient; This is the equivalent distance from the pixel to the imaging device.

[0028] As a preferred embodiment of this practice, an atmospheric scattering model is constructed and combined with the characteristics of smoke samples from cable fires to generate smoke degradation images under various working conditions based on real flame images. Specifically: The operating conditions include: different smoke colors including white smoke, gray smoke and black smoke, different fire light diffuse reflection conditions (with / without), and different smoke concentration gradients (e.g. 20%, 40%, 60%, 80%, 100%). To simulate different smoke types and diffuse reflection conditions of firelight in an underground cable tunnel fire scenario, this embodiment uses a global atmospheric light scattering model. With smoke density attenuation coefficient The range of values ​​is limited by parameter combinations. White smoke exhibits strong multi-directional scattering characteristics under high-temperature vaporization, leading to a significant increase in overall image brightness. Therefore, under white smoke conditions... The value range is set to 0.75 to 1.00. The value range is 0.03 to 0.15; the scattering intensity of the gray smoke is moderate, and it is semi-transparent. Therefore, under gray smoke conditions, The value range is set to 0.55 to 0.85. The value range is 0.3 to 0.6; because black smoke contains a high concentration of carbon particles, its light absorption is significantly stronger than its scattering, therefore, under black smoke conditions... The value range is set to 0.10 to 0.40. The value range is 0.6 to 1.5. This can be achieved by adjusting the above... and The range of the interval corresponds to the generation of smoke degradation images for different smoke colors, namely white smoke, gray smoke, and black smoke.

[0029] Meanwhile, to reflect the illuminating effect of firelight on smoke particles, this embodiment will incorporate atmospheric scattering model... With smoke density attenuation coefficient The range of values ​​is limited by parameter combinations to characterize firelight diffuse reflection scenes of varying intensities. (Settings) The value range is 0.40 to 0.60. When the value ranges from 0.8 to 1.2, it represents a condition where the flame intensity is weak and there is significant smoke obstruction. This indicates that when the flame intensity is weak, the increase in ambient light is limited, and it simulates significant smoke obstruction. The value range is 0.60 to 0.80. When the value ranges from 0.6 to 0.8, it represents a condition where the flame intensity is moderate and the smoke is partially illuminated, simulating a state where the flame intensity is moderate and the smoke is partially illuminated; set The value range is 0.80 to 1.00. When the value ranges from 0.3 to 0.5, it represents the condition where smoke particles are exposed to high brightness under strong fire light, resulting in obvious reflection and brightening phenomena, reflecting the attenuation characteristics caused by fire light penetrating smoke. Through the aforementioned parameter combination, the atmospheric scattering model is controlled to generate smoke degradation images under various conditions based on real flame images. This constructs a paired dataset containing real flame images and corresponding smoke degradation images, providing a realistic and controllable sample foundation for the subsequent training of image enhancement models.

[0030] As a preferred embodiment of this example, Figure 2The image shown is a preferred generator structure in this embodiment. The generator adopts a progressive multi-stage generator structure, which is internally divided into multiple levels of sub-generators according to their functions. These sub-generators are cascaded together to form a complete generator.

[0031] like Figure 3 As shown, the sub-generator includes a symmetrical encoder and decoder. The encoder extracts multi-scale deep features of the smoke-degraded image through continuous convolution and pooling operations. The decoder gradually restores the spatial resolution and generates a flame reconstruction image through deconvolution or upsampling. A skip connection is used between the encoder and decoder to directly transfer the shallow detail features of the encoder to the corresponding decoding layer, enabling the model to possess both strong semantic information and high-resolution details. This structure ensures: 1. Low-level features are used to preserve the image's detailed texture (such as edges and brightness transitions); 2. Deep features are used to capture global information about smoke distribution and firelight reflection; 3. A multi-scale fusion mechanism ensures that the model can still achieve stable and reliable image enhancement in scenes with high smoke concentration, strong reflection, and severe local occlusion. This generator structure gives the model strong generalization ability, maintaining high reconstruction accuracy under different smoke concentrations, smoke colors, and firelight illumination conditions. Furthermore, the number of layers and convolutional kernel selection of the sub-generators (encoders and decoders), as well as the number of layers, convolutional kernels, and activation functions selection of the discriminator, can be adaptively set according to factors such as the actual resolution of the acquired flame images: For high-resolution images (such as 1920×1080), the number of generator and discriminator layers can be appropriately increased (6~8 layers for each generator and 5~6 layers for the discriminator), paired with 3×3 / 5×5 convolutional kernels (generator) and 4×4 / 6×6 convolutional kernels (discriminator), with LeakyReLU used for the initial activation to improve gradient stability; For low-resolution images (such as 640×480 and below), the number of layers can be reduced (4~5 layers for each generator and 3~4 layers for the discriminator), mainly using 3×3 (generator) and 4×4 (discriminator) convolutional kernels, with ReLU used for the initial activation to simplify calculation. The core is to maintain the symmetry of the generator encoder and decoder, and to increase the number of channels when the discriminator is downsampled. At the same time, the rationality of the adjustment is verified by PSNR and the convergence effect of the loss function, balancing feature extraction capability and computational efficiency.

[0032] The discriminator employs multi-layer convolutional downsampling to map the input image into a feature space with progressively decreasing resolution. Increasing the number of convolutional kernels enhances its ability to discriminate local statistical features. The discriminator's output is a single-channel probability map, representing the confidence level that the input image is a realistic and clear image. The multi-scale convolutional structure allows the discriminator to simultaneously focus on the overall texture changes and the degree of detail restoration in large-scale smoke regions, thus providing stable and effective gradient information during generator training. This structure ensures that the generator can continuously optimize the difference between the reconstructed flame image and the real flame image.

[0033] To improve the model's ability to restore images under varying smoke opacity intensities, the smoke image enhancement model employs a progressive cumulative training strategy during training. Multiple sub-generators are trained in stages, progressing from easy to difficult: the first sub-generator is responsible for restoring the low-frequency structural information of the flame; subsequent sub-generators continuously supplement texture, highlight details, and edge information based on the output of the previous sub-generator. Backpropagation during training not only optimizes the current sub-generator but also fine-tunes the preceding sub-generators, achieving a "gradual enhancement, weight accumulation" effect. Simultaneously, the discriminator applies adversarial constraints to the intermediate flame restoration images generated by each sub-generator against the real flame image, ensuring consistency between each sub-generator and the real flame image. This achieves a multi-stage progressive restoration from coarse to fine, resulting in a clear flame restoration image. The flame restoration image is compared by progressively blurring the original real flame image using Gaussian blur. The discriminator is trained using the Gaussian blurred image corresponding to the image generated by the generator, and the sub-generators are fine-tuned through backpropagation. Furthermore, the backpropagation process is characterized by "step-by-step backtracking and layered updating." Specifically, when the first... Child generator After receiving the output from the previous stage and generating a staged restoration result, this output will be input into the discriminator along with the real, clear image. The discriminator calculates adversarial and reconstruction losses, which are then propagated back along the network structure using a chain rule. The backpropagation gradient first updates the current sub-generator. The parameters; if the preceding sub-generator participates in the joint optimization, the gradient will continue to propagate forward to... , The generator is divided into sub-generators, whose parameters are fine-tuned at a small learning rate; if some preceding sub-generators have converged and been frozen, the gradient is truncated at this layer and no longer updated. Through this backpropagation method of "focusing on updating in this stage and selectively fine-tuning in previous stages", the generator can gradually enhance its restoration ability while maintaining the stability of the overall structure, achieving a progressive image restoration effect from coarse to fine.

[0034] In a preferred embodiment of this invention, to improve the reconstruction accuracy of the smoke region, the smoke image enhancement model introduces attention modules in the deep layers of the encoder and the key layers of the decoder. This allows the image enhancement model to more accurately focus on the flame regions obscured by smoke, such as blurred flame edges, light reflection areas, and texture loss areas, thereby improving reconstruction quality and learning to recover visual features crucial for fire identification tasks from the degraded smoke image. The attention modules include a channel attention module and a spatial attention module. The channel attention module adaptively adjusts the weights of different feature channels, assigning different importance to each feature channel, enabling the model to focus on channel features that are more valuable for distinguishing flames and smoke.

[0035] like Figure 4 As shown, the spatial attention module generates a spatial weight matrix through several Conv layers (the number of Conv layers is preferably 2-4 layers in this embodiment, with a minimum of no less than 2 layers to ensure that effective spatial information is extracted from the feature map and corresponding weights are generated; the maximum number of layers is no more than 4 layers, as too many convolutions will cause the weight matrix to be overly smoothed, losing local details (such as the fine edges of the flame). At the same time, the activation function in the Conv layer can be selected according to performance requirements) and multiplies the spatial weight matrix with the original feature map position by position through the Sigmoid activation function to obtain the enhanced feature map after spatial attention weighting, so as to highlight the sensitive area of ​​smoke occlusion and enhance the model's ability to restore key areas (such as the main body of the flame and the edge brightening area) during the decoding process.

[0036] The training objective of the image enhancement model is to take a degraded image as input and output a reconstructed flame image, which is then compared with a realistic flame image. For model performance evaluation, this embodiment uses objective metrics such as the loss function (MSE) and peak signal-to-noise ratio (PSNR) as the main evaluation criteria. Experimental results show that the smoke image model exhibits rapid convergence of the loss function and a significant improvement in PSNR during training, demonstrating that the method described in this embodiment can effectively restore image quality and provide a reliable information foundation for subsequent fire identification and early warning.

[0037] As a preferred embodiment of this practice, to further quantify the enhancement effect of the model, this embodiment introduces Peak Signal-to-Noise Ratio (PSNR) as a core evaluation metric. PSNR focuses on measuring the fidelity between the reconstructed image and the original sharp image at the pixel level; a higher value indicates less image distortion. The reconstructed image processed by the image enhancement model described in this embodiment has a significantly higher PSNR value than the unprocessed smoke-degraded image. The peak signal-to-noise ratio is expressed as:

[0038] In the formula, Peak signal-to-noise ratio (PSNR) is used to measure image quality or the fidelity of prediction results. The maximum possible value of the image pixels is given in the PSNR calculation formula. This represents the maximum possible value for an image pixel, used to characterize the upper limit of pixel brightness. This value is determined by the image's data type: when the input image is normalized and represented using floating-point values ​​in the range [0,1], The value is set to 1; when the image is represented in 8-bit unsigned integer (uint8) format, its pixel value ranges from 0 to 255, therefore... The value is set to 255. In this embodiment, it is preferable to normalize all images during training and evaluation, scaling the pixel values ​​to the [0,1] range. The fixed value is 1. This is determined by... The value of this value ensures the consistency and repeatability of the PSNR calculation process; The squared error of a single image sample is expressed as:

[0039] In the formula, The number of pixels in the image test set; For the first The true value of each pixel (or pixel channel); For the first The true value of each pixel (or pixel channel).

[0040] The image enhancement model described in this embodiment exhibits excellent convergence characteristics during training. The value of the loss function (MSE) used decreases rapidly and stabilizes in the early stages of training. Specifically, within the initial 25 training epochs, its value significantly decreases from 5.64 to 0.87 and remains stable in subsequent training. This trend indicates that the model training process is robust, effectively learning the mapping relationship from smoke-degraded images to clear images, and possesses strong generalization ability. The combined good convergence trend of the loss function and the significant improvement in the PSNR index confirm that the smoke image enhancement model proposed in this embodiment not only has a stable training process but also effectively improves the visual quality of fire images obscured by smoke, laying a solid foundation for subsequent accurate fire identification.

[0041] To verify the effectiveness and superiority of the method provided in this embodiment, some specific examples are provided below: 1. For scenarios involving high concentrations of smoke in cable fires, this system effectively removes smoke and restores image details. To comprehensively evaluate its performance, this embodiment uses three sets of core metrics for integrated verification of the image enhancement model.

[0042] like Figure 5 As shown in (a), the loss function (MSE) used by the image model during training exhibits significant convergence characteristics. Within the first 500 training epochs, the MSE value rapidly decreases from 0.032 to 0.0171 and then remains stable. This trend indicates that the model achieves effective feature learning and parameter optimization in a short time, effectively suppressing the error between the predicted results and the true annotations. Further analysis from the perspective of error magnitude, such as... Figure 5 As shown in (b), the mean absolute error (MAE) plateaued after approximately 43 training cycles, eventually stabilizing within the range of 0.067 ± 0.003. MAE reflects the model's comprehensive ability to characterize both local details and overall trends. Its low and stable level indicates that the model has reached a good learning equilibrium in capturing the nonlinear relationship between input and output, with minimal fluctuations in the deviation between the predicted results and the actual samples. Furthermore, as... Figure 5 As shown in (c), the Peak Signal-to-Noise Ratio (PSNR), a key indicator for measuring the quality of image reconstruction, also showed a continuous upward trend during model training, significantly increasing from an initial 11.12 to a stable 38.8. The higher PSNR value indicates that the model has a strong ability to suppress reconstruction noise, preserve image texture details, and recover structural information.

[0043] The combined performance of MSE, MAE, and PSNR indicators shows that the model achieved fast, stable, and high-quality convergence during training, demonstrating its superior image reconstruction performance and robust learning ability.

[0044] 2. To evaluate the adaptability of the image enhancement model in complex fire scenarios, a comparative analysis was conducted on the reconstruction performance under different smoke concentrations and diffuse reflection conditions. For example... Figure 6 As shown, the model exhibits high structural similarity (SSIM, ranging from 0 to 1, with higher values ​​indicating better reconstruction) in both white smoke and black smoke scenarios, with stable overall reconstruction results. As smoke concentration increases, the SSIM index gradually decreases, indicating that the masking effect of dense smoke on image structural features gradually intensifies. Specifically, in the white smoke scenario, SSIM decreases from 0.853 to 0.783, indicating that high-concentration smoke significantly weakens the model's ability to extract effective texture and structural information, resulting in a slight decrease in reconstruction performance. In contrast, the decrease in SSIM in the black smoke scenario is relatively smaller (from 0.779 to 0.731), indicating that the model still has some adaptability in black smoke environments. However, overall, the model's enhancement effect in white smoke scenarios is still better than that in black smoke conditions. Figure 7 As shown, even under severe occlusion by black smoke concentration of 1.0, the model is still able to effectively restore image details and structural information, demonstrating strong enhancement capabilities.

[0045] When the diffuse reflection condition of firelight is introduced, the SSIM (Signal Scale Mean Square) decreases to varying degrees for both white and black smoke. The decrease is more significant in the white smoke diffuse reflection scenario, with an average SSIM decrease of 0.0614 across all smoke concentrations. The SSIM in the black smoke diffuse reflection scenario also shows an average decrease of 0.0088. This is mainly because the diffuse reflection effect caused by firelight significantly alters the local brightness distribution, leading to reduced image contrast and blurred texture boundaries. This makes it difficult for the model to accurately distinguish between the real target and the area affected by illumination interference during feature extraction and structure reconstruction, thus causing an overall decrease in SSIM.

[0046] The comprehensive experimental results show that the model performs best in environments with low smoke concentration and no reflection interference, effectively maintaining the structural consistency and detail reproduction of the images. While the model's performance slightly decreases under high smoke concentration and strong reflection interference conditions, it still maintains good reconstruction stability and generalization ability, demonstrating its robustness and potential for improvement under extremely complex conditions. Overall, the model's SSIM value remains above 0.71 in all test scenarios, fully validating its stability and robustness in various complex fire scenarios, providing strong support for its engineering applications in practical fire monitoring and smoke visual recognition tasks.

[0047] Example 2 Accordingly, this embodiment provides a multi-degraded flame image restoration system to implement the method described in Embodiment 1, including a smoke degraded image module and a smoke image enhancement module.

[0048] The smoke degradation image generation module is used to simulate real-world high-voltage cable fires in underground utility tunnels, collect smoke samples and real flame images of cable fires in real time, construct an atmospheric scattering model, and combine the characteristics of smoke samples from cable fires to generate smoke degradation images under various working conditions based on real flame images.

[0049] A smoke image enhancement module is used to build and train a smoke image enhancement model based on an improved DCGAN architecture. The smoke image enhancement model includes a generator and a discriminator. The generator is used to reconstruct the smoke-degraded image to generate a flame restoration image. The discriminator is used to judge the authenticity of the flame restoration image and the real flame image, ensuring that the generator continuously optimizes the difference between the flame restoration image and the real flame image.

[0050] Example 3 This embodiment provides an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in Embodiment 1 of the present invention.

[0051] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in Embodiment 1 of the present invention.

[0052] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0053] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0054] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0055] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0056] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for restoring flame images with multiple degradations, characterized in that, Includes the following steps: By simulating real-world high-voltage cable fires in underground utility tunnels, real-time smoke samples and flame images of cable fires are collected. An atmospheric scattering model is constructed and combined with the characteristics of smoke samples from cable fires, smoke degradation images under various working conditions are generated based on real flame images. A smoke image enhancement model based on an improved DCGAN architecture is constructed and trained. The smoke image enhancement model includes a generator and a discriminator. The generator is used to reconstruct the degraded smoke image to generate a flame restoration image. The discriminator is used to judge the authenticity of the flame restoration image and the real flame image, ensuring that the generator continuously optimizes the difference between the flame restoration image and the real flame image.

2. The method for restoring multi-degraded flame images according to claim 1, characterized in that, The atmospheric scattering model is expressed as follows: In the formula, Image degraded by smoke; This is a true image of the flame before it deteriorated; For global atmospheric light; The smoke concentration gradient, representing the proportion of light retained during propagation, is expressed as: In the formula, The smoke density attenuation coefficient; This is the equivalent distance from the pixel to the imaging device.

3. The method for restoring multi-degraded flame images according to claim 2, characterized in that, An atmospheric scattering model is constructed and combined with the characteristics of smoke samples from cable fires. Based on real flame images, smoke degradation images under various working conditions are generated as follows: The operating conditions include: different smoke colors, including white smoke, gray smoke and black smoke, different fire light diffuse reflection conditions and different smoke concentration gradients; Global atmospheric light in atmospheric scattering models With smoke density attenuation coefficient The range of values ​​is limited by parameter combinations, where, under white smoke conditions, The value range is set to 0.75 to 1.

00. The value range is 0.03 to 0.15; under gray smoke conditions, The value range is set to 0.55 to 0.

85. The value range is 0.3 to 0.6; under black smoke conditions The value range is set to 0.10 to 0.

40. The value range is 0.6 to 1.5; set up The value range is 0.40 to 0.

60. When the value ranges from 0.8 to 1.2, it indicates a condition with weak flame intensity and significant smoke obstruction; [Setting] The value range is 0.60 to 0.

80. A value ranging from 0.6 to 0.8 indicates a condition where the flame intensity is moderate and the smoke is partially illuminated; [Setting] The value range is 0.80 to 1.

00. When the value ranges from 0.3 to 0.5, it represents the working condition where smoke particles are exposed to high brightness under strong fire light, resulting in obvious reflection and brightening phenomena. Through the combination of the parameters, the atmospheric scattering model is controlled to generate smoke degradation images under various working conditions based on real flame images.

4. The method for restoring multi-degraded flame images according to claim 1, characterized in that, The generator adopts a progressive multi-stage generator structure. The generator is internally divided into multi-level sub-generators according to their functions. Each sub-generator adopts a cascaded structure to form a complete generator. The sub-generator includes a symmetrical encoder and decoder. The encoder extracts multi-scale deep features of the smoke degradation image through continuous convolution and pooling operations. The decoder gradually restores the spatial resolution and generates a flame restoration image through deconvolution or upsampling. The encoder and decoder use a skip connection to directly transfer the shallow detail features of the encoder to the corresponding decoding layer. The discriminator employs a multi-layer convolutional downsampling method to map the input image into a feature space with progressively lower resolution, thereby enhancing the ability to discriminate local statistical features by increasing the number of convolutional kernels. The smoke image enhancement model employs a progressive cumulative training strategy during training, training multiple sub-generators in stages according to the principle of increasing difficulty: the first sub-generator is responsible for recovering the low-frequency structural information of the flame, and subsequent sub-generators continuously supplement texture, highlight details, and edge information based on the output of the previous sub-generator; during training, backpropagation not only optimizes the current sub-generator but also fine-tunes the preceding sub-generators; simultaneously, the discriminator applies adversarial constraints to the intermediate results of the flame reconstruction images generated by each sub-generator against the real flame image, ensuring that each sub-generator in the generator maintains consistency with the real flame image, thereby achieving multi-stage progressive reconstruction from coarse to fine and obtaining a clear flame reconstruction image.

5. The method for restoring multi-degraded flame images according to claim 4, characterized in that, The smoke image enhancement model introduces attention modules in the deep layers of the encoder and the key layers of the decoder. The attention modules include channel attention modules and spatial attention modules. The channel attention module adaptively adjusts the weights of different feature channels, assigning different importance to each feature channel; The spatial attention module generates a spatial weight matrix through several Conv layers, and then multiplies the spatial weight matrix with the original feature map position by position using the Sigmoid activation function to obtain the spatial attention-weighted enhanced feature map.

6. The method for restoring multi-degraded flame images according to claim 1, characterized in that, The smoke image enhancement model uses peak signal-to-noise ratio (PSNR) as an evaluation metric, where PSNR is expressed as: In the formula, Peak signal-to-noise ratio (PSNR) is used to measure image quality or the fidelity of prediction results. This represents the maximum possible value of the image pixels. This represents the squared error of a single image sample.

7. A multi-degraded flame image restoration system, characterized in that, The system is used to implement the method as described in any one of claims 1 to 6, comprising a smoke degradation image module and a smoke image enhancement module; The smoke degradation image generation module is used to simulate real-world underground utility tunnel high-voltage cable fires, collect smoke samples and real flame images of cable fires in real time, construct an atmospheric scattering model, and combine the characteristics of smoke samples from cable fires to generate smoke degradation images under various working conditions based on real flame images. A smoke image enhancement module is used to build and train a smoke image enhancement model based on an improved DCGAN architecture. The smoke image enhancement model includes a generator and a discriminator. The generator is used to reconstruct the smoke-degraded image to generate a flame restoration image. The discriminator is used to judge the authenticity of the flame restoration image and the real flame image, ensuring that the generator continuously optimizes the difference between the flame restoration image and the real flame image.

8. An electronic device, the electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images