CN-121998859-A - Mine dust fog image defogging method and system based on real haze diffusion

CN121998859ACN 121998859 ACN121998859 ACN 121998859ACN-121998859-A

Abstract

Aiming at the problems of single training data, low reasoning speed of a diffusion model, rapid sampling texture deletion and poor non-uniform fog treatment effect of the existing defogging model, the invention adopts a four-stage treatment process, namely, firstly generating realism paired training data based on a lightweight degradation encoder and a characteristic modulation mechanism; the inference efficiency is improved through a single-step prediction strategy, texture details are migrated through slice statistics alignment operation to generate an initial defogging estimation diagram, and finally a density sensing fidelity guiding mechanism based on a transmissivity diagram is introduced to adaptively adjust the guiding strength of different fog concentration areas. The invention reduces the number of model parameters, improves the reasoning efficiency and the image restoration quality, realizes the balance of defogging thoroughness and structural fidelity, is suitable for the deployment of mine edge computing equipment, and effectively supports the underground target detection and intelligent perception tasks.

Inventors

KOU QIQI
ZHU YU
ZHAO SHIRAN
CHENG DEQIANG
SI LEI
DAI JIANBO

Assignees

中国矿业大学

Dates

Publication Date: 20260508
Application Date: 20260206

Claims (9)

1. The mine dust and fog image defogging method based on real haze diffusion is characterized by comprising the following steps of: S1, generating realism paired training data based on lightweight characteristic modulation, inputting a clear mine background image, a real reference haze image and a text control vector, extracting degradation characteristics by a lightweight degradation encoder, and generating a network through characteristic modulation injection, wherein a single step generation of a realism haze image is realized; s2, adopting a single-step prediction strategy, directly predicting noise in an early time step by using a trained prediction noise network, and estimating an early prediction mean value in a single step; s3, generating an initial defogging estimation graph based on slice statistics alignment, extracting corresponding overlapped slices of a real haze image and an early prediction mean value, and splicing and fusing after channel-level statistics alignment; S4, guiding the thinned image based on the density sensing fidelity, calculating a transmissivity image, and generating a clear image by combining MS-SSIM loss gradient weighting updating.
2. The method for defogging a mine dust and fog image based on real haze diffusion according to claim 1, wherein in step S1, the text control vector is obtained by inputting a text prompt describing the mine environment into a CLIP text encoder, and mapping the text prompt into a semantic embedded vector of a fixed length.
3. The method for defogging a mine dust and fog image based on real haze diffusion according to claim 1, wherein in step S1, a lightweight degradation encoder is composed of 1 initial convolution layer followed by 4 cascaded lightweight residual blocks, each residual block contains a convolution-ReLU-convolution structure inside, a channel attention mechanism is embedded, and finally spatial features of a reference haze image are compressed into one-dimensional degradation feature vectors through a global average pooling layer.
4. The method for defogging mine dust and fog images based on real haze diffusion according to claim 1, wherein in step S1, the feature modulation injection is to convert the extracted degradation features into style scaling coefficients through a mapping layer, and the weight modulation-demodulation mechanism is used to dynamically scale and normalize convolution kernel weights of the generated network U-Net.
5. The mine dust and fog image defogging method based on real haze diffusion according to claim 1, wherein in step S1, a generating network adopts an SD-Turbo architecture, a clear image is encoded as a potential feature, a single-step reasoning is combined with a text control vector and a degradation feature to generate a haze image potential feature, and the haze image potential feature is restored into a real haze image through a decoder.
6. The method for defogging a dust and fog image of a mine based on real haze diffusion according to claim 1, wherein in the step S3, the statistical alignment of the slices is to calculate a channel-level mean value and a standard deviation of a real haze image slice and an early prediction mean slice, subtract the mean value of the real haze image slice and the early prediction mean slice, multiply the mean value of the real haze image slice and the early prediction mean slice by the ratio of the standard deviation of the real haze image slice and the early prediction mean slice, and add the mean value of the prediction slices to complete the alignment operation.
7. The method for defogging a mine dust and fog image based on real haze diffusion according to claim 1, wherein in step S4, the transmittance map is calculated by a dark channel prior algorithm, and the transmittance map has a large value in a mist region and a small value in a dense mist region.
8. The method for defogging a mine dust and fog image based on real haze diffusion according to claim 1, wherein in step S4, the weighting update is to multiply the MS-SSIM loss gradient by elements with the transmittance map, the gradient weight of the haze region is high to preserve the original structure, and the gradient weight of the dense haze region is low to restore the occlusion content.
9. A mine dust and mist image defogging system based on true haze diffusion for implementing the method of claims 1 to 8, comprising: the realism data generation module is used for inputting clear mine background images, real reference haze images and text control vectors, and generating realism paired training data through lightweight degradation coding and feature modulation; The single-step prediction module is used for directly predicting noise in an earlier time step by using a prediction noise network and estimating an early prediction mean value in a single step; The slice statistics alignment module is used for extracting corresponding overlapped slices and completing statistics alignment to generate an initial defogging estimation graph; and the density perception guiding refinement module is used for calculating a transmissivity image, combining MS-SSIM loss gradient weighting updating and outputting a clear image.

Description

Mine dust fog image defogging method and system based on real haze diffusion Technical Field The invention relates to the technical field of computer vision and image processing, in particular to a mine dust and fog image defogging method and system based on real haze diffusion. Background In a coal mine underground operation scene, high-concentration coal dust and water mist particles suspend in the air due to dust fall measures such as long-term excavation, transportation, spraying and sprinkling, and the like, so that the underground environment is in a complex dust and mist mixing state. Meanwhile, underground illumination mainly depends on artificial light sources such as miner lamps, fixed LED lamps and the like, and the roadway wall surface has strong light reflection, so that a non-uniform illumination environment is formed. In such an environment, the images collected by the monitoring device generally have the problems of low contrast, loss of detail, local overexposure, halation effect and the like. Particularly, fog drops in a spraying area are large and unevenly distributed, so that thick fog and a thin fog area coexist in an image, and subsequent target detection and intelligent perception tasks are seriously influenced. At present, for the restoration processing of such images, a traditional defogging algorithm based on an atmospheric scattering model or an end-to-end defogging network based on deep learning is mainly adopted. In recent years, diffusion models have been increasingly used in the field of image defogging due to their excellent generation capability, and attempts have been made to fill in lost image details by the generation method. However, existing deep learning methods rely on synthetic data sets, which are typically generated by physical models, with a single texture, and cannot simulate the complex distribution of real mine dust and mist. Although partial methods introduce control Net guided generation, they directly multiplex huge U-Net encoder structures, resulting in large parameters, low computational efficiency, and difficulty in accurately extracting degradation features at the instance level. The diffusion model reasoning speed is slow, and the standard diffusion model defogging process needs to start from Gaussian noise and perform tens or even hundreds of iterative denoising samples. This cumbersome iterative process consumes a large amount of computational resources, and the inference delay is high, greatly limiting its deployment and application on mine edge computing devices. In addition, in the defogging process, the existing method generally adopts a globally unified guiding strategy, and a dense fog area and a thin fog area are not distinguished. However, in mine images, the haze region needs to retain the original structure, and the dense haze region needs to recover the lost content. The uniform treatment can cause distortion of the structure of the haze area or incomplete defogging of the dense haze area, and the balance between the image fidelity and the restoration quality can not be achieved. Disclosure of Invention Aiming at the technical defects, the invention aims to provide a mine dust fog image defogging method and system based on real fog diffusion, which are used for efficiently extracting fog characteristics of a reference image to generate high-quality paired real mine fog training data, avoiding a complete iterative sampling process from pure noise, improving reasoning efficiency, quickly migrating texture details of a real fog image into a prediction result to construct a high-quality initial defogging estimation image, adaptively adjusting guide intensity according to fog concentration, and effectively recovering details of a dense fog area while maintaining a thin fog area structure. In order to achieve the above purpose, the invention adopts the following technical scheme: a mine dust fog image defogging method based on real haze diffusion comprises the following steps: S1, generating realism paired training data based on lightweight characteristic modulation, inputting a clear mine background image, a real reference haze image and a text control vector, extracting degradation characteristics by a lightweight degradation encoder, and generating a network through characteristic modulation injection, wherein a single step generation of a realism haze image is realized; s2, adopting a single-step prediction strategy, directly predicting noise in an early time step by using a trained prediction noise network, and estimating an early prediction mean value in a single step; s3, generating an initial defogging estimation graph based on slice statistics alignment, extracting corresponding overlapped slices of a real haze image and an early prediction mean value, and splicing and fusing after channel-level statistics alignment; S4, guiding the thinned image based on the density sensing fidelity, calculating a transmissivity image, and generating a