A mine pollution image decontamination restoration method and system based on a decontamination diffusion probability model

Abstract

This invention discloses a method and system for decontamination and restoration of mine pollution images based on a decontamination diffusion probability model. The system includes a pollution degree regression model, a decontamination diffusion probability model, and a restoration calculation platform. The method first automatically assesses the pollution level of the input mine pollution image using the pollution degree regression model. Then, this level is used as a key condition input into the decontamination diffusion probability model to guide iterative decontamination, ultimately outputting the original clean image after decontamination and restoration. This invention, by combining quantitative assessment with conditional generation, achieves accurate reverse inference and restoration of mine pollution images, significantly improving the intelligence level of pollution monitoring and remediation assessment.

Description

Technical Field

[0001] This invention belongs to the field of mine pollution and remediation, and particularly relates to a method and system for decontamination and restoration of mine pollution images based on a decontamination diffusion probability model. Background Technology

[0002] Environmental pollution caused by dust, wastewater, and heavy metal diffusion during mining and production activities is becoming increasingly serious. Currently, monitoring of pollution in mining areas mainly relies on on-site sampling and remote sensing image analysis. Existing technologies mostly employ methods based on physical models (such as dark channel priors) or traditional image enhancement algorithms to attempt to recover clear scenes from polluted images. However, these methods have significant limitations: they can usually only perform visual "dehazing" or "enhancement," and cannot provide precise and continuous quantitative assessment of the degree of pollution; more importantly, they are essentially based on empirical formulas to repair images, and cannot reverse-engineer the dynamic process of pollution diffusion from its principles, thus making it difficult to achieve accurate restoration from any polluted state to the original clean state. This results in insufficient capabilities of existing technologies in pollution source tracing, precise quantification of remediation effects, and dynamic assessment.

[0003] In recent years, with the development of artificial intelligence technology, deep generative models such as diffusion models have demonstrated powerful capabilities in image synthesis and editing, providing new ideas for image restoration. However, directly applying such models to highly specialized and data-scarce mining pollution scenarios still faces significant challenges: on the one hand, mining pollution is complex and lacks a large amount of labeled "clean-polluted" pairs for model training; on the other hand, general diffusion models lack the ability to understand and control specific domain knowledge (such as pollution levels), and their generation process is somewhat arbitrary, making it difficult to guarantee the accuracy and reliability of the restoration results in terms of environmental monitoring. Therefore, how to construct an intelligent system that can accurately quantify the degree of pollution and perform controllable, high-fidelity image restoration has become a pressing technical problem in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for decontamination and restoration of mine contamination images based on a decontamination diffusion probability model, comprising: Input images of mine pollution; The pollution images of the mine are evaluated using a pollution degree regression model, and the pollution level k, representing the degree of pollution, is output. The mine pollution image and the pollution level k are used as data pairs and input into the decontamination diffusion probability model. The decontamination diffusion probability model performs iterative decontamination processing based on the contamination level k to generate a decontamination restoration image. The decontamination-restored image is post-processed to output the final original clean image.

[0005] Optionally, the training process of the decontamination diffusion probability model includes: Acquire a series of pollution images with progressively increasing levels of contamination; For each contaminated image, calculate the difference between it and the original clean image, which is taken as the actual contamination increment corresponding to that contaminated image; Construct a training dataset with contaminated images and corresponding real-world contamination increments as training data pairs; A neural network is trained using the training dataset. The input of the neural network is a contaminated image and its corresponding contamination level time step information, and the output is a predicted contamination increment. The parameters of the neural network are optimized by minimizing the difference between the predicted pollution increment and the actual pollution increment.

[0006] Optionally, the acquisition method for the series of pollution images with progressively increasing pollution levels includes: Based on a benchmark image of a mine and a sequence of cue words describing the gradual evolution of pollution, the Wensheng video big model is driven to generate a video of the gradual escalation of pollution, and keyframes are extracted from the video. Based on a sequence of text prompts with progressively enhanced pollution features, and under the constraint of fixed generation parameters, a series of static images with progressively increasing pollution levels and consistent visual elements are generated through a text-based image model. Keyframes are extracted by performing temporal consistency filtering and illumination change correction on real-scene monitoring videos that record gradual changes in pollution.

[0007] Optionally, the calculation of the actual pollution increment is achieved in the following way: Using a pre-trained image semantic encoder, high-dimensional semantic feature vectors are extracted from the original clean image and the contaminated image, respectively; The difference between the two high-dimensional semantic feature vectors is calculated using a distance function, and this difference is the true pollution increment.

[0008] Optionally, the neural network is a U-Net architecture with an encoder-decoder structure and skip connections, or a diffusion Transformer architecture built based on Transformer modules.

[0009] Optionally, during the use of the decontamination diffusion probability model, a deterministic sampling algorithm is used for iterative decontamination, wherein the deterministic sampling algorithm is DDIM, PLMS, or DPM-Solver.

[0010] Optionally, the training process of the pollution degree regression model includes: Acquire a series of pollution images labeled with continuous pollution levels; A regression neural network is trained using the polluted image and its corresponding pollution level label. The input of the regression neural network is the polluted image, and the output is the predicted pollution level value.

[0011] Optionally, the regression neural network is based on a deep feature extraction network pre-trained on a large-scale general image dataset and fine-tuned on mine pollution image data through transfer learning.

[0012] This invention provides a mine contamination image decontamination and restoration system based on a decontamination diffusion probability model, comprising: A pollution level regression model is used to receive input images of mine pollution and automatically estimate the pollution level k corresponding to the image. A decontamination diffusion probability model is used to receive a data pair consisting of the pollution level k output by the pollution degree regression model and the mine pollution image, and to generate a decontamination restoration image through iterative decontamination processing; The post-processing module is used to process the decontamination and restoration image to restore details and improve visual effects; The recovery calculation platform, on which the pollution degree regression model, the decontamination diffusion probability model, and the post-processing module are deployed, is used to execute the complete decontamination recovery process.

[0013] Optionally, the restoration computing platform is an edge computing device with limited computing resources, and the pollution degree regression model and the decontamination diffusion probability model deployed on it are compressed and accelerated by at least one of the following techniques: model quantization, pruning, or knowledge distillation.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention enables precise reverse engineering and intelligent restoration of mine pollution images. It automatically quantifies and assesses pollution levels using a pollution degree regression model, and uses this as a key condition to guide a decontamination diffusion probability model for targeted decontamination. This not only restores polluted images to their original clean state with high fidelity, but also fundamentally constructs a "digital twin" decision-making system connecting the current state of pollution with remediation goals. This invention overcomes the limitations of traditional methods that can only perform visual enhancement, achieving quantitative tracing of the pollution process, dynamic evaluation of remediation effects, and precise allocation of remediation resources. This transforms mine environmental remediation from an experience-based, extensive model to a new adaptive and precise remediation model that is quantifiable, predictable, and optimizable. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for decontamination and restoration of mine pollution images based on a decontamination diffusion probability model, including the following steps: Input images of mine pollution; The pollution images of the mine are evaluated using a pollution degree regression model, and the pollution level k, representing the degree of pollution, is output. The mine pollution image and the pollution level k are used as data pairs and input into the decontamination diffusion probability model. The decontamination diffusion probability model performs iterative decontamination processing based on the contamination level k to generate a decontamination restoration image. The decontamination-restored image is post-processed to output the final original clean image.

[0019] Furthermore, the training process for the decontamination diffusion probability model includes: Acquire a series of pollution images with progressively increasing levels of contamination; For each contaminated image, the difference between it and the original clean image is calculated as the true contamination increment corresponding to the contaminated image. The calculation of the true contamination increment is achieved by: using a pre-trained image semantic encoder to extract high-dimensional semantic feature vectors of the original clean image and the contaminated image respectively; and calculating the difference between the two high-dimensional semantic feature vectors through a distance function, which is the true contamination increment. Construct a training dataset with contaminated images and corresponding real-world contamination increments as training data pairs; The neural network is trained using the training dataset. The input of the neural network is a contaminated image and its corresponding contamination level time step information, and the output is a predicted contamination increment. The parameters of the neural network are optimized by minimizing the difference between the predicted pollution increment and the actual pollution increment.

[0020] Specifically, it can be divided into the following steps: (1) Data preparation and preprocessing: Prepare a series of contaminated images with progressively increasing levels of contamination and preprocess them. Preprocessing includes image alignment, resolution normalization, and pixel value normalization. Image alignment ensures that all images (baseline image and contaminated image) are strictly aligned in space to guarantee the effectiveness of changes in contamination level; resolution normalization scales or crops all images to the fixed input size required by the model; pixel value normalization normalizes the image pixel values ​​from [0, 255] to [-1, 1] or [0, 1], which helps with the numerical stability of the model during training and accelerates convergence.

[0021] (2) Constructing training data pairs: For each contaminated image, the difference between it and the original clean image is calculated as the "real contamination added in step k", thus constructing a training data pair (contaminated image, real contamination added in step k). Here, "contamination" is the target that the model needs to learn to predict.

[0022] (3) Training the network: Train a neural network. The network input is the contaminated image and the corresponding time step k. The network output is the predicted contamination. Its learning objective is to accurately predict the contamination applied at the kth step.

[0023] (4) Decontamination and restoration: An efficient sampling strategy is adopted. Starting from the input contaminated image, the decontamination diffusion probability model obtained by training is used to perform iterative decontamination and reconstruct the original clean image after decontamination within a small number of sampling steps.

[0024] (5) Restoration evaluation: Post-process the decontamination restored image to repair details and textures that may be lost during the decontamination restoration process or improve the overall visual naturalness and effect, and use peak signal-to-noise ratio and structural similarity index to automatically evaluate and output the decontamination effect.

[0025] Among them, the methods for acquiring a series of pollution images with progressively increasing levels of pollution include: Based on a benchmark image of a mine and a sequence of cue words describing the gradual evolution of pollution, the Wensheng video big model is driven to generate a video of the gradual escalation of pollution, and keyframes are extracted from the video. Specifically, the process involves providing a high-resolution, pollution-free baseline image of the mine (the source must be a genuine photograph, such as drone aerial photography or satellite imagery). Then, using the Wensheng video model, combined with precise prompts and the baseline image, a video is generated that dynamically presents the gradual evolution of pollution levels, starting from the baseline image. This video must ensure that each frame is based on the baseline image, using the model to simulate the temporal superposition and diffusion of pollution effects, ultimately forming a visual video of progressive pollution. Based on this, keyframes are extracted from the generated video to systematically obtain a series of static pollution images showing progressively increasing pollution levels.

[0026] The process of generating a video of the gradual evolution of pollution based on the Wensheng video big data model further includes: (1) Multimodal pollution alert engine: A structured pollution alert lexicon is constructed, which maps pollution types (such as heavy metals in water, suspended particulate matter, vegetation alteration, etc.) to quantifiable pollution intensity parameters. Based on the preset pollution evolution curve, the system automatically selects and combines a sequence of alert words with temporal logic from the lexicon to describe the entire dynamic evolution of pollution, forming precise temporal control instructions, which are then input into the Wensheng video big data model to drive it to generate a video showing the gradual evolution of pollution levels.

[0027] (2) Baseline Image Embedding and Consistency Constraint: The mine baseline image is embedded into the latent space of the Wensheng video big model through an encoder as the spatial layout and content prior for video generation; and during the generation process, consistency constraints are applied through a cross-attention mechanism to ensure that pollution diffusion is strictly based on the geographical elements of the baseline image and to prevent scene structure from drifting.

[0028] (3) Parametric definition of pollution evolution curve: The evolution process of the pollution degree gradually intensifies is controlled by a definable, nonlinear pollution evolution curve. The curve takes the time step as the independent variable and outputs the comprehensive pollution intensity value to dynamically modulate the noise injection level in the model.

[0029] Based on a sequence of text prompts with progressively enhanced pollution features, and under the constraint of fixed generation parameters, a series of static images with progressively increasing pollution levels and consistent visual elements are generated through a text-based image model. Specifically, this method systematically constructs a set of static mine images with progressively increasing pollution levels through structured text descriptions and fixed generation parameters. First, a sequence of text prompts with progressively stronger pollution features serves as content guidance, gradually transitioning from an unpolluted state to heavily polluted scenes such as water discoloration and vegetation degradation, thus achieving a semantically hierarchical expression of pollution levels. During generation, all prompts use the same random seed in the text-generated image model, ensuring consistency in visual elements such as scene layout, perspective, and lighting, making pollution level the only systematically controllable primary variable in the image sequence. This method, combining semantic guidance and generation techniques, can efficiently produce a series of images with a consistent scene foundation and progressively increasing pollution features.

[0030] The process of systematically constructing static mine images with progressively increasing levels of contamination further includes: (1) Design of structured pollution prompt word sequence: Construct a text prompt word sequence with progressively enhanced pollution features. The sequence is designed based on pollution type and pollution intensity parameters. Starting from the original state without pollution, it gradually transitions to pollution scenarios such as mild, moderate and severe through semantic hierarchical expression. Each prompt word corresponds to a specific pollution intensity level. The semantic consistency and pollution feature distinguishability of the prompt words are optimized through natural language processing technology, thereby forming a prompt word sequence with continuous and progressive pollution intensity.

[0031] (2) Fixed generation parameters and visual consistency maintenance: During the image generation process, all text prompts use the same random seed in the large model of the generated image. Furthermore, key generation parameters such as sampler, sampling steps and guiding scale are fixed to lock the macro and micro features of the scene layout, viewpoint, lighting and background elements from the source of noise distribution. This systematically eliminates random variations in the generation process and ensures that the degree of contamination becomes the only systematically controllable dominant variable in the entire image sequence.

[0032] (3) Control mechanism for hierarchical expression of pollution level: Define a quantifiable semantic pollution scale with progressive pollution intensity levels. The scale systematically maps specific pollution intensity parameters to the structured pollution prompt word sequence. Through a parameterized control process, based on the preset pollution evolution law, drive the prompt word sequence to execute in order of its pollution intensity level, thereby achieving precise and smooth progressive control of pollution level in semantic space.

[0033] Keyframes are extracted by performing temporal consistency filtering and illumination change correction on real-scene monitoring videos that record gradual changes in pollution.

[0034] Specifically, this involves collecting real-time monitoring videos showing the gradual changes in pollution levels at the mine site, or gathering complete real-world video footage documenting the gradual evolution of pollution levels, to construct a raw video library with a clear temporal sequence logic. Subsequently, the video frame sequences in the library undergo temporal consistency filtering and illumination change correction preprocessing to eliminate interference from changes in shooting conditions affecting pollution level assessment. Then, the videos are further processed by extracting frames at equal intervals or key nodes, decomposing the continuous time series into discrete static image frames, thereby constructing a series of structured static image sets showing a progressively increasing degree of pollution.

[0035] Among them, temporal consistency filtering is used to suppress high-frequency abrupt changes between frames in the video frame sequence caused by transient interference, sensor noise or compression artifacts that are unrelated to the long-term trend of pollution spread, and to ensure that the image obtained by frame extraction has temporal smoothness in terms of scene structure and content.

[0036] Lighting variation correction: This is used to compensate for differences in overall brightness, contrast, and color distribution between frames caused by variations in sunlight, weather, or automatic camera parameter adjustments due to different shooting times. It normalizes each frame to a stable reference lighting condition to eliminate the interference of lighting variations on the judgment of pollution level.

[0037] The above processing can eliminate the interference caused by changes in shooting conditions (such as weather and lighting angle) on the judgment of pollution level, and ensure that the visual elements of the extracted still image, other than pollution features, remain as stable as possible.

[0038] Furthermore, the calculation process for the actual pollution added in step k is as follows: Feature Extraction: A pre-trained CLIP model image encoder is selected to extract high-dimensional semantic feature vectors F(x_0) and F(x_k) for the original clean image x_0 and the contaminated image x_k at step k, respectively. This model is highly sensitive to semantic content and can efficiently extract high-level, abstract semantic features from images, while being insensitive to non-semantic changes such as lighting and shadows. It utilizes the pre-trained CLIP model image encoder to extract high-dimensional semantic feature vectors for the original clean image x_0 and the contaminated image x_k at step k; the feature space obtained by training the CLIP model on large-scale image-text pairs maximizes sensitivity to image semantic content while maintaining robustness to non-semantic, irrelevant visual changes.

[0039] Difference Calculation: The actual contamination t_k added in the k-th step is obtained by calculating the difference between the original clean image x_0 and the contaminated image x_k in the k-th step; the difference can be calculated directly on the image pixels, or on the semantic features extracted by the pre-trained model, as shown in the specific formula: t_k = D(x_0, x_k) or t_k = D(F(x_0), F(x_k)); in: x_0 is the original healthy image; x_k is the contaminated image at step k; F(x_0) is the feature vector or feature map of the original healthy image x_0.

[0040] F(x_k) is the feature vector or feature map of the contaminated image x_k at step k.

[0041] D(·) is a distance function used to calculate the difference between two images or features.

[0042] Constructing training data pairs: The training data pairs are pre-computed before training, which is different from the real-time generation of training data pairs during DDPM training. However, this does not affect the trainability of the model. As long as there are enough training data pairs and they cover all levels of contamination, the resulting model will also have good robustness and generalization.

[0043] For any contaminated image x_k, its corresponding t_k is automatically calculated using the steps described above. This results in many training data pairs (x_k, t_k).

[0044] The distance function D(·) is chosen as the tensor difference function, i.e.: t_k = x_k - x_0; Where x_k and x_0 are the contaminated image and the original clean image, respectively, and the resulting t_k is the difference tensor of the same dimension as the input image, representing the incremental change in pixel space from the original clean image to the contaminated image at step k.

[0045] The distance function D(·) is chosen as the vector difference function, i.e.: t_k = F(x_k) - F(x_0); Where F(x_k) and F(x_0) are the feature vectors or feature maps of the contaminated image and the original clean image, respectively, and the obtained t_k is the difference vector with the same dimension as the features, representing the incremental change in the feature space from the original clean image to the contaminated image at step k.

[0046] The distance function D(·) is selected based on a scalar distance function of the feature vector, including but not limited to cosine distance, Euclidean distance, and Mahalanobis distance. The resulting t_k is a scalar that represents the overall degree of difference in the feature metric space from the original clean image to the contaminated image at step k.

[0047] The distance function D(·) is used to calculate the difference between two CLIP feature vectors in a high-dimensional semantic space. This calculation method focuses on the difference in feature semantic content and can effectively capture the overall semantic change of the scene caused by pollution, rather than pixel-level local noise. This makes the actual pollution t_k added in the k-th step a more semantically meaningful quantitative indicator of pollution degree.

[0048] Through the above process, the measurement of the degree of image contamination is transformed into the semantic deviation from the original clean image in the CLIP semantic space. This deviation has a higher consistency with human subjective perception of the severity of contamination, providing a more discriminative supervisory target for the subsequent learning of the decontamination diffusion probability model.

[0049] Furthermore, the training network is configured as follows: (1) Network architecture: The neural network is a deep neural network capable of image-to-image prediction.

[0050] (2) Network input: contaminated image x_k and corresponding time step k. Time step k is converted into an embedding vector through sinusoidal positional encoding and then input into various parts of the network so that the network knows which level of contamination it is currently processing.

[0051] (3) Network output: The network output is the predicted pollution t_k.

[0052] (4) Network loss function: a loss function based on the difference between the network output and the actual pollution target; (5) Network optimization: The gradient descent algorithm and its variants are used for model optimization.

[0053] Among them, the neural network is either a U-Net architecture with an encoder-decoder structure and skip connections, or a diffusion Transformer architecture built based on the Transformer module. It uses a self-attention mechanism to globally model the semantics of the image, which can effectively capture and fuse global semantic context information related to pollution in the image.

[0054] The loss function uses mean squared error loss (MSE Loss), which continuously optimizes the neural network model parameters and prediction performance by minimizing the gap between predicted pollution and actual pollution.

[0055] The gradient descent algorithm is the Adam or AdamW optimizer. The optimization process incorporates a learning rate scheduling strategy, specifically cosine annealing. Gradient pruning is used to stabilize training during optimization. In the decontamination and restoration step, a deterministic sampling algorithm, such as DDIM, PLMS, or DPM-Solver, is employed for iterative decontamination during the use of the decontamination diffusion probability model.

[0056] A brief introduction to sampling algorithms and an efficiency comparison (Table 1) are as follows: DDIM (Denoising Diffusion Implicit Models); PLMS (Pseudo Linear Multi-step method); DPM-Solver (Diffusion Probabilistic Model Solver); Table 1 Optionally, during the iterative decontamination process, a series of intermediate process images can be output from the input contaminated image to the output original clean image.

[0057] Furthermore, the training process for the pollution level regression model includes: Acquire a series of pollution images labeled with continuous pollution levels; The polluted image and its corresponding pollution level label are used to train a regression neural network. The input of the regression neural network is the polluted image, and the output is the predicted pollution level value. The architecture of the regression neural network is a pre-trained ResNet, EfficientNet, or Vision Transformer. The regression neural network is based on a deep feature extraction network pre-trained on a large-scale general image dataset and fine-tuned on the mine pollution image data through transfer learning.

[0058] Specifically, it includes: (1) Prepare a series of pollution images with progressively increasing pollution levels, with each image corresponding to a continuous pollution level label; (2) Construct training data pairs, including contaminated images and corresponding contamination level labels; (3) Train a regression neural network. The network input is a polluted image and the output is the predicted pollution level. The loss function is based on the difference between the predicted value and the true label.

[0059] This embodiment employs a course-based learning approach to optimize the training process. Specifically, in the initial training phase, image data with low contamination levels (smaller k-values) is primarily used for training. As training progresses, the proportion of training data with higher contamination levels (larger k-values) is gradually increased until all contamination levels are covered. This training strategy, progressing from easy to difficult, helps the model converge more stably and quickly, ultimately resulting in a final model with stronger generalization ability and better decontamination performance.

[0060] This embodiment also provides a mine contamination image decontamination and restoration system based on a decontamination diffusion probability model, including: A pollution level regression model is used to receive input images of mine pollution and automatically estimate the pollution level k corresponding to the image. A decontamination diffusion probability model is used to receive a data pair consisting of the pollution level k output by the pollution degree regression model and the mine pollution image, and to generate a decontamination restoration image through iterative decontamination processing; The post-processing module is used to process the decontamination and restoration image to restore details and improve visual effects; The recovery calculation platform, on which the pollution degree regression model, the decontamination diffusion probability model, and the post-processing module are deployed, is used to execute the complete decontamination recovery process.

[0061] If the restoration computing platform is an edge computing device or other platform with very limited computing resources or capabilities, then when deploying the decontamination diffusion probability model and the contamination degree regression model, at least one of the following techniques should be used to compress and accelerate the model in order to adapt to the limited computing resources of the edge device and meet the near real-time or real-time requirements.

[0062] In this embodiment, the restoration computing platform is an edge computing device with limited computing resources. The pollution degree regression model and the decontamination diffusion probability model deployed on it are compressed and accelerated by at least one of the following techniques: model quantization, pruning, or knowledge distillation.

[0063] The beneficial effects of this invention are: This invention breaks through the limitations of traditional monitoring, which can only provide the current state of pollution. Through its unique decontamination and restoration capabilities, it achieves "reverse deduction" and "zero-based assessment" of the pollution process. It can not only accurately quantify the absolute gap between the current pollution state and the original healthy environment, but also reveal the diffusion patterns and potential sources of pollution, providing environmental supervision with revolutionary insights from tracing the causes to assessing the effectiveness of remediation.

[0064] Traditional remediation solutions rely on experience and abstract data, making it difficult to predict the final outcome. This invention, by decontaminating and restoring any polluted image to its original clean state, essentially constructs a visualized "target effect map," thereby allowing for an intuitive and clear understanding and definition of the ultimate goal of remediation and restoration.

[0065] Traditional remediation is like a "black box operation," making it difficult to assess the effectiveness of each step in real time. This invention precisely quantifies pollution, enabling refined monitoring and guidance of the remediation process. By inputting on-site images from the mid-remediation phase into the model, the "remaining pollution distance" between the current state and the original target can be calculated in real time. This allows for an objective assessment of the effectiveness of current remediation methods and dynamic adjustment of strategies, achieving a shift from "extensive governance" to "precise, adaptive remediation."

[0066] Mine ecological restoration is not a one-time solution; the stability of the restored ecosystem requires long-term monitoring. After restoration, images of the mining area are collected regularly, and "decontamination and restoration" calculations are performed. If the resulting "pollution distance" remains stable at a low level close to zero, it indicates that the ecosystem is healthy and stable. If this value rises, it may be an early warning signal of ecological degradation or pollution recurrence, thus achieving a transition from "endpoint assessment" to "long-term health monitoring."

[0067] For large-scale polluted areas, the types and degrees of pollution vary in different regions. By dividing the polluted areas into zones for decontamination and restoration and quantifying pollution, a "pollution degree heat map" can be generated. This map can intuitively indicate the core areas with the most severe pollution and the peripheral areas with mild impact, thereby guiding the differentiated allocation of treatment resources (such as manpower, equipment, and reagents), achieving precise treatment with a "one-site-one-policy" approach, and avoiding resource waste.

[0068] In summary, the innovation of this invention lies in its creative deep integration of cutting-edge decontamination diffusion probability models and pollution degree regression models, constructing a complete technical closed loop of "pollution assessment-targeted decontamination," resulting in significant cross-domain synergistic enhancement effects. Specifically, the system uses an independent pollution degree regression model to accurately quantify complex visual pollution into continuous pollution level parameters, providing key and clear reverse guidance conditions for the subsequent decontamination process. This transforms diffusion model-based decontamination from general image processing into targeted, high-fidelity reverse inference of the physical process of mine pollution diffusion. Simultaneously, this invention systematically integrates a progressive pollution data construction method for mine scenarios, a precise pollution degree assessment module, and a conditional decontamination diffusion module, forming a full-cycle solution covering data preparation, state diagnosis, and precise execution. This overall concept not only solves the two core problems of insufficient pollution quantification and blind restoration processes in existing technologies, but also simulates and reverses the pollution process through a data-driven approach, achieving quantifiable and traceable precise restoration from any polluted state to the original clean state. This provides a new intelligent paradigm for mine environmental monitoring and remediation, demonstrating outstanding non-obviousness and significant technological progress.

[0069] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A mine pollution image decontamination restoration method based on a decontamination diffusion probability model, characterized in that, Includes the following steps: Input images of mine pollution; The pollution images of the mine are evaluated using a pollution degree regression model, and the pollution level k, representing the degree of pollution, is output. The mine pollution image and the pollution level k are used as data pairs and input into the decontamination diffusion probability model. The decontamination diffusion probability model performs iterative decontamination processing based on the contamination level k to generate a decontamination restoration image. The decontamination-restored image is post-processed to output the final original clean image.

2. The method of claim 1, wherein, The training process of the decontamination diffusion probability model includes: Acquire a series of pollution images with progressively increasing levels of contamination; For each contaminated image, calculate the difference between it and the original clean image, which is taken as the actual contamination increment corresponding to that contaminated image; Construct a training dataset with contaminated images and corresponding real-world contamination increments as training data pairs; The neural network is trained using the training dataset. The input of the neural network is a contaminated image and its corresponding contamination level time step information, and the output is a predicted contamination increment. The parameters of the neural network are optimized by minimizing the difference between the predicted pollution increment and the actual pollution increment.

3. The method of claim 2, wherein, The methods for acquiring the series of progressively more polluted images include: Based on a benchmark image of a mine and a sequence of cue words describing the gradual evolution of pollution, the Wensheng video big model is driven to generate a video of the gradual escalation of pollution, and keyframes are extracted from the video. Based on a sequence of text prompts with progressively enhanced pollution features, and under the constraint of fixed generation parameters, a series of static images with progressively increasing pollution levels and consistent visual elements are generated through a text-based image model. Keyframes are extracted by performing temporal consistency filtering and illumination change correction on real-scene monitoring videos that record gradual changes in pollution.

4. The method of claim 2, wherein, The actual increase in pollution is calculated in the following way: Using a pre-trained image semantic encoder, high-dimensional semantic feature vectors are extracted from the original clean image and the contaminated image, respectively; The difference between two high-dimensional semantic feature vectors is calculated using a distance function, and this difference is the true pollution increment.

5. The method according to claim 2, characterized in that, The neural network is either a U-Net architecture with an encoder-decoder structure and skip connections, or a diffusion Transformer architecture built on the Transformer module.

6. The method according to claim 2, characterized in that, In the process of using the decontamination diffusion probability model, a deterministic sampling algorithm is used for iterative decontamination, which includes DDIM, PLMS and DPM-Solver.

7. The method according to claim 1, characterized in that, The training process of the pollution level regression model includes: Acquire a series of pollution images labeled with continuous pollution levels; The contaminated image and its corresponding contamination level label are used to train a regression neural network, where the input to the regression neural network is the contaminated image and the output is the predicted contamination level value.

8. The method according to claim 7, characterized in that, The regression neural network is based on a deep feature extraction network pre-trained on a large-scale general image dataset and fine-tuned on mine pollution image data through transfer learning.

9. A system for decontamination and restoration of mine contamination images based on a decontamination diffusion probability model, characterized in that, include: A pollution level regression model is used to receive input images of mine pollution and automatically estimate the pollution level k corresponding to the image. A decontamination diffusion probability model is used to receive a data pair consisting of the pollution level k output by the pollution degree regression model and the mine pollution image, and to generate a decontamination restoration image through iterative decontamination processing; The post-processing module is used to process the decontamination and restoration image to restore details and improve visual effects; The recovery calculation platform, on which the pollution degree regression model, the decontamination diffusion probability model, and the post-processing module are deployed, is used to execute the complete decontamination recovery process.

10. The system according to claim 9, characterized in that, The restoration computing platform is an edge computing device with limited computing resources. The pollution degree regression model and the decontamination diffusion probability model deployed on it are compressed and accelerated by at least one of the following techniques: model quantization, pruning, or knowledge distillation.

Images