Coal mine underground low-illumination image enhancement method and system based on imaging confidence

Abstract

The application relates to a coal mine underground low-illumination image enhancement method and system based on imaging confidence in the field of image processing and coal mine intelligent technology, and comprises the following steps: establishing a three-component imaging model, inputting an image, and extracting a brightness channel; filtering and denoising; calculating a pixel-level noise variance; obtaining an actual confidence; three-region division and differential recovery; performing color reconstruction and enhancement on the image; and outputting the enhanced image. The application effectively solves the problems that existing methods cannot quantitatively evaluate the reliability of pixel information in the glow area and the shadow area, and that applying a unified recovery operation to the whole image leads to noise and amplification of the truncation artifact, effectively solves the problem that a Retinex method misattributes scattering to light, leading to excessive brightening around the light source, and effectively solves the problem that a deep learning method depends on training data and has limited generalization in a coal mine scene.

Description

Technical Field

[0001] This invention belongs to the field of image processing and intelligent coal mining technology, specifically relating to a method and system for enhancing low-light images in underground coal mines based on imaging confidence. Background Technology

[0002] The development of artificial intelligence technology has made all-weather, all-space monitoring of underground coal mines possible. However, limited lighting in underground coal mines, sudden brightness changes caused by passing headlamps or vehicle lights, and glow effects caused by coal dust scattering severely affect the imaging quality of underground video monitoring, reducing the reliability of intelligent applications. Underground image degradation has three characteristics: (1) Low illumination: The underground environment relies on artificial lighting, and the overall illumination is much lower than that of the ground scene. The dynamic range of the image is narrow and the details in the dark area are severely lost. (2) Non-uniform illumination: The close-range illumination of LED mining lamps causes local overexposure and far-end underexposure, and the illumination changes drastically in space, making global enhancement methods difficult to be effective. (3) LED glow effect: LED light propagates in air containing high concentration of coal dust particles and produces multiple scattering, forming a hazy halo around the light source that is unrelated to the scene texture. This area has high brightness but low texture information, which not only covers the background details but also makes it difficult for the enhancement algorithm to distinguish between real texture and glow artifacts.

[0003] Existing image enhancement methods for underground coal mines mainly include: histogram equalization methods (HE, CLAHE), which are prone to overexposure or under-enhancement under non-uniform lighting conditions; Retinex theory methods (SSR, MSR, multi-scale gradient domain guided filtering, etc.), which, based on the premise that an image can be decomposed into the product of reflectance and illumination, misattribute the additional radiation from glow scattering as illumination, resulting in excessive brightening around the light source and the appearance of halo artifacts at abrupt changes in illumination; decomposition methods based on glow models, which iteratively solve for illuminance and reflectance components, but lack quantitative measurement of the reliability of information in the glow region, and apply the same restoration operation to damaged pixels as to normal pixels, which amplifies noise and truncation artifacts; and deep learning methods (Zero-DCE, Retinexformer, URetineXNet, etc.), which rely on training data and have limited generalization ability in coal mine scenes. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a low-light image enhancement method for coal mines based on imaging confidence. This method effectively solves the problems of existing methods being unable to quantitatively assess the reliability of pixel information in glow and shadow areas, and the amplification of noise and truncation artifacts caused by applying a uniform restoration operation to the entire image. It also effectively solves the problem of Retinex-type methods misclassifying glow scattering as illumination, leading to excessive brightening around the light source. Furthermore, it effectively solves the problem of deep learning methods relying on training data and having limited generalization ability in coal mine scenes.

[0005] This invention discloses a method and system for enhancing low-light images in coal mines based on imaging confidence. The core idea is "recovery first, enhancement later": first, information is recovered from damaged pixels using an imaging physical model and spatial priors; then, enhancement is performed on the reconstructed reliable signal. Specifically, this method uses formal analysis of the entire underground imaging chain to establish a three-component imaging model, decouples LED glow as an independent additive component from the multiplicative Retinex framework, defines pixel-level brightness confidence based on error propagation analysis, and performs differentiated recovery of high, medium, and low light intensity zones accordingly, ultimately achieving contrast enhancement and color reconstruction.

[0006] To address the aforementioned problems, this invention provides a method for enhancing low-light images in coal mines based on imaging confidence, comprising the following steps: S1 establishes a three-component imaging model, inputs a low-light RGB image from a coal mine, converts it to the luminance domain, and extracts the luminance channels of the input image. Y ; S2 for the brightness channel Y Perform guided filtering for denoising to obtain denoised brightness. ; S3 for noise reduction brightness Perform Gaussian difference calculation to estimate the glow component Ĝ; improve the denoised brightness. Perform guided filtering to smooth the estimation of illumination components ; Calculate pixel-level noise variance based on CMOS sensor model ; S4 will combine the glow component (Ĝ) and the illumination component. Noise variance Substitute the values ​​into the error propagation formula to calculate the variance of the reflectance estimate. Then calculate the luminance confidence level. The actual confidence level is obtained after scaling and cropping. S5 divides the region into high / medium / low confidence zones based on thresholds of 0.8 / 0.2: Gamma mapping is performed in the high confidence zone. ,in The reflectance estimate is for the high confidence region; the reflectance estimate is for the medium confidence region. With guided filter spatial prior Weighted fusion: ,in The reflectance estimate for the medium confidence region, with fusion weights. , The spatial priors obtained by applying a guided filter operator to the reflectivity estimate, The Bayesian posterior term is used; the low-confidence region is reconstructed using the guided filter posterior mean. GF is the guided filter operator; the results from the three regions are merged according to the confidence level indicator mask. ,in , , These are high, medium, and low confidence region indicator masks, respectively. The three masks are mutually exclusive and cover each pixel completely. The output is then merged to restore brightness. ; S6 When brightness is restored Applying CLAHE enhances brightness when the mean value is below the brightness threshold of 40. Otherwise, skip the CLAHE. S7 with The scaling factor is r, which multiplies each channel of the original color image by r to output the enhanced image.

[0007] To establish the three-component imaging model, in step S1, a formal analysis of the entire imaging chain in underground coal mines is performed, modeling the observed brightness as the sum of the multiplicative coupling of scene reflection and illumination, the independent additive components of glow scattering, and noise: ;in For the observed brightness, ρ is the scene reflectivity, L is the equivalent illumination, G is the equivalent glow independent of the scene texture, and ε is the noise.

[0008] To accurately extract the luminance channel Y In step S1, the red, green, and blue components of each pixel are analyzed according to the standard luminance formula. Brightness calculations are performed, which converts the image to the brightness domain, allowing the extraction of the brightness channels from the input image. Y .

[0009] To ensure the noise reduction effect, in step S2, the guided filtering uses the image itself as a guide map to remove sensor noise while maintaining the edge structure.

[0010] To enhance the discriminative power of the high / medium / low confidence regions, the actual confidence level is calculated in step S4 as follows: Composed of glow component Ĝ, illumination component Noise variance Calculate the variance of the reflectance estimate ,in This is an estimate of reflectivity. To estimate the variance of the glow, Estimating the variance of illumination; Then define pixel-level brightness confidence. Quantitatively characterize the reliability of the reflectance estimate for each pixel. A value close to 1 indicates that the reflectivity estimate is reliable, while a value close to 0 indicates that the signal is overwhelmed by noise. The confidence level for actual use is obtained by scaling and cropping. .

[0011] To achieve differentiated recovery, in step S5, the high confidence region... Reflectivity estimation is reliable, using gamma mapping. Preserve texture; medium confidence zone The reflectance estimate is fused with the spatial prior obtained from the guided filter using a confidence-weighted method; the higher the confidence level, the greater the weight of the reflectance estimate. Low-confidence regions... It relies entirely on spatial priors and uses the posterior mean of the guided filter as the reconstruction result.

[0012] The present invention also provides a low-light image enhancement system for underground coal mines based on imaging confidence, comprising: Image input module, used to acquire low-light RGB images of underground coal mines; The filtering and denoising module is used to filter and denoise low-light RGB images to obtain a preprocessed image; The confidence level three-region segmentation module is used to divide the preprocessed image into high confidence region, medium confidence region and low confidence region, and perform differential restoration; Enhancement and Color Reconstruction Module: Used for color reconstruction and image enhancement of the image after merging the three regions; Image output module, used to output enhanced RGB images.

[0013] In summary, compared with the prior art, the present invention has the following technical advantages and beneficial effects: (1) By establishing a three-component imaging model, the present invention decouples the LED glow as an independent additive component from the multiplicative Retinex framework, so that the glow does not enter the illumination estimation, eliminating the excessive brightening and halo artifacts caused by the glow entering the illumination around the light source, and maintaining the consistency of the image space. (2) This invention defines pixel-level brightness confidence through error propagation analysis, divides pixels into high, medium and low regions based on confidence and performs differentiated restoration, so that noise and truncation artifacts of low confidence pixels (glossy region and shadow region) can be effectively suppressed. (3) This invention significantly improves image quality indicators through confidence-driven three-region differential restoration: On a coal mine test set of 200 images, the standard NIQE is 4.90, BRISQUE is 18.09, entropy is 7.34, and gradient is 9.56 (the lower the NIQE and BRISQUE, the better; the higher the entropy and gradient, the better); the ablation experiment shows that after removing the three-region differential restoration (using uniform weight), the NIQE deteriorates by 21.8% and the BRISQUE deteriorates by 24.7%, proving that this mechanism has a substantial contribution to the improvement of the indicators; (4) All modules of this invention are constructed based on deterministic image operators, requiring no training data, having low computing power requirements, good reproducibility, and adaptable to real-time offline deployment of edge devices in coal mines. Attached Figure Description

[0014] Figure 1 This is a flowchart of the low-light image enhancement method for underground coal mines based on imaging confidence, according to the present invention. Figure 2 This is a schematic diagram of the low-light image enhancement method for underground coal mines based on imaging confidence, according to the present invention. Figure 3 This is a structural block diagram of the coal mine low-light image enhancement system based on imaging confidence according to the present invention. Figure 4 This is a confidence state visualization diagram in a specific implementation example of the present invention. Detailed Implementation

[0015] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0016] like Figure 1 , Figure 2 The method for enhancing low-light images in coal mines based on imaging confidence includes the following steps: S1. Establish a three-component imaging model. Input a low-light RGB image from a coal mine. For each pixel, calculate the red, green, and blue components (0-255) according to the standard brightness formula. Brightness calculations are performed, which converts the image to the brightness domain, allowing the extraction of the brightness channels from the input image. Y ; Specifically, a three-component imaging model is established: a formal analysis is performed on the entire imaging chain in underground coal mines, from "light source → scene reflection → coal dust scattering → optical lens → CMOS photoelectric conversion → ADC quantization". The observed brightness is modeled as the sum of the multiplicative coupling of scene reflection and illumination, the independent additive components of glow scattering, and noise. Where ρ is the scene reflectivity, L is the equivalent illumination, G is the equivalent glow independent of the scene texture, and ε is the noise; when G≈0, it degenerates into the classic Retinex model; experimental results show that the glow has no significant systematic coupling with the scene texture and is highly mixed with the illumination component in the frequency domain, making frequency domain separation infeasible. It must be separated in the spatial domain by means of a confidence mechanism.

[0017] S2 for the brightness channel Y Perform guided filtering for denoising to obtain denoised brightness. ; The guided filtering uses the image itself as the guide map (filter radius r=5, regularization parameter eps=0.02) to remove sensor noise while preserving edge structure.

[0018] S3 for noise reduction brightness Perform Gaussian difference operations (two Gaussian kernel scales) to estimate the glow component Ĝ; improve the denoised brightness. Perform guided filtering to smooth the estimation of illumination components Based on the CMOS sensor model (Where K=60 is the system gain,) Calculate pixel-level noise variance to obtain the standard deviation of readout noise. ; S4 will combine the glow component (Ĝ) and the illumination component. Noise variance Substitute the values ​​into the error propagation formula to calculate the variance of the reflectance estimate. Then calculate the luminance confidence level. The actual confidence level is obtained after scaling and cropping. Based on error propagation analysis, the noise variance Glow curve estimation variance With light estimation variance Calculate the variance of the reflectance estimate ,in This is the reflectance estimate; therefore, a pixel-level brightness confidence level is defined. Quantitatively characterize the reliability of the reflectance estimate for each pixel. A confidence level approaching 1 indicates a reliable reflectivity estimate, while a confidence level approaching 0 indicates the signal is overwhelmed by noise. To broaden the distinction between high and low confidence regions, a scaling and cropping process is used to obtain the actual usable confidence level. 5.0 is a scaling factor that amplifies the confidence distribution to a scale that matches the three-zone division threshold (0.8 / 0.2); the upper limit of 1.0 is the physical upper bound of the confidence (corresponding to a completely reliable reflectance estimate), and the lower limit of 0.7 prevents the confidence of extreme low-light pixels (with extremely large error variance) from being excessively suppressed.

[0019] S5 divides the region into high / medium / low confidence zones based on thresholds of 0.8 / 0.2: Gamma mapping is performed in the high confidence zone. ,in The reflectance estimate is for the high confidence region; the reflectance estimate is for the medium confidence region. With guided filter spatial prior Weighted fusion: ,in The reflectance estimate for the medium confidence region, with fusion weights. (follow) (Increases linearly from 0.2 to 0.8, then from 0 to 1). The spatial priors obtained by applying a guided filter operator to the reflectivity estimate, The Bayesian posterior term is used; the low-confidence region is reconstructed using the guided filter posterior mean. , where GF is the guided filter operator, and the results from the three regions are merged according to the confidence level indicator mask: ,in , , These are high / medium / low confidence region indicator masks, which are mutually exclusive and provide full pixel-by-pixel coverage. Brightness is restored after merging the output. ; Pixels are divided into high-confidence, medium-confidence, and low-confidence regions based on the stated brightness confidence level. This division requires no training data and is calculated directly from the error propagation formula. High-confidence region ( The reflectance estimation (approximately 44% of pixels) is reliable, using gamma mapping. (γ=0.9) Preserve texture; Medium confidence region ( (Approximately 27% of pixels) The reflectance estimate is fused with the spatial prior obtained from guided filtering in a confidence-weighted manner (Bayesian weighted fusion). The higher the confidence level, the greater the weight of the reflectance estimate; in low-confidence regions ( (Approximately 29% of the pixels) rely entirely on the spatial prior, using the posterior mean of the guided filter as the reconstruction result.

[0020] S6 When brightness is restored When the mean value is below the brightness threshold of 40, the image can be determined to be in an extremely low-light condition. Applying CLAHE can enhance the brightness. Otherwise, skip the CLAHE. Regarding the restored brightness Apply contrast-limited adaptive histogram equalization (CLAHE). Brightness is enhanced when Enabled when the mean is below 40 (8-bit), otherwise skipped to prevent over-sharpening; enhanced brightness is mapped back to RGB space using the color ratio method to obtain an enhanced image.

[0021] S7 with The scaling factor is r, which multiplies each channel of the original color image by r to output the enhanced image.

[0022] like Figure 3 The system shown is a low-light image enhancement system for underground coal mines based on imaging confidence, used to implement the above method, including: Image input module, used to acquire low-light RGB images of underground coal mines; The filtering and denoising module is used to filter and denoise low-light RGB images to obtain a preprocessed image; The confidence level three-region segmentation module is used to divide the preprocessed image into high confidence region, medium confidence region and low confidence region, and perform differential restoration; Enhancement and Color Reconstruction Module: Used for color reconstruction and image enhancement of the image after merging the three regions; Image output module, used to output enhanced RGB images.

[0023] To facilitate understanding of the technical solution of this invention, the following describes in detail the low-light image enhancement method and system for coal mines based on imaging confidence, using a specific underground construction scenario.

[0024] Scene description: as attached Figure 4 The drilling site had a high dust concentration during construction, so explosion-proof LED lights for mining were used for illumination. Intrinsically safe network cameras were installed on the side and rear of the drilling rig, capturing monitoring video at a resolution of 1920×1080 and 25 frames per second. The captured images exhibited three typical types of degradation: localized overexposure in areas directly illuminated by the headlamp (bright but with compressed textures), extremely low brightness in the deep tunnels and equipment shadow areas (low signal-to-noise ratio, details obscured), and glow halos formed around the light source due to dust scattering (masking coal wall textures and reducing image spatial consistency).

[0025] Application process: (1) Acquisition and transmission: The intrinsically safe network camera acquires video frame by frame, and transmits it to the underground explosion-proof edge computing device via the mine flame-retardant optical cable and the underground industrial ring network; (2) Luminance domain processing: The computing device executes steps S1 to S7 sequentially for each frame, first extracting the luminance channel. Y(S1), Guided filtering denoising (S2), Dual-core Gaussian difference estimation of glow Ĝ, guided filtering estimation of illumination L̂, calculation of noise variance according to CMOS model (S3), error propagation calculation of confidence and amplitude limiting (S4); (3) Three-zone differential restoration (S5): The reflectivity estimation of the direct headlamp area is reliable (high confidence zone), and gamma mapping is performed to preserve texture; the glow halo edge and shadow transition zone (medium confidence zone), the reflectivity estimation and guided filtering spatial prior are fused according to confidence; the glow core and deep shadow area (low confidence zone), the guided filtering spatial prior reconstruction is completely relied upon to suppress noise amplification; the R, G, and B channels are processed one by one and then merged according to confidence mask to restore brightness. (4) Enhancement and Output (S6, S7): When When the mean is less than 40, CLAHE is applied to expand the dynamic range, and RGB is reconstructed by the color ratio method to output the enhanced video stream; (5) Display and storage: The enhanced video stream is displayed in real time on the monitoring screen of the ground dispatch center, and the original frame and the enhanced frame are stored synchronously in the video server for future reference.

[0026] Implementation Results: Glow halos were effectively suppressed and no longer obscured coal face texture; the outline of the roadway and equipment details in the deep shadow areas were restored to visibility; textures in the direct headlamp illumination area remained unexposed. (See attached image) Figure 4 As shown, the confidence levels of pixels in the glow area and the deep shadow area are significantly lower. After the three-region differential restoration, the details in the above areas are restored and visible. The entire enhancement process consists of deterministic image operators and does not require training data.

[0027] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape, principle and application direction of this application should be covered within the scope of protection of this application.

Claims

1. A method for enhancing low-light images in coal mines based on imaging confidence, characterized in that, Includes the following steps: S1 establishes a three-component imaging model, inputs a low-light RGB image from a coal mine, converts it to the luminance domain, and extracts the luminance channels of the input image. Y ; S2 for the brightness channel Y Perform guided filtering for denoising to obtain denoised brightness. ; S3 for noise reduction brightness Perform Gaussian difference calculation to estimate the glow component Ĝ; improve the denoised brightness. Guided filtering is used to smooth the estimation of illumination components. ; Calculate pixel-level noise variance based on CMOS sensor model ; S4 will combine the glow component (Ĝ) and the illumination component. Noise variance Substitute the values ​​into the error propagation formula to calculate the variance of the reflectance estimate. Then calculate the luminance confidence level. The actual confidence level is obtained after scaling and cropping. S5 divides the region into high / medium / low confidence zones based on thresholds of 0.8 / 0.2: Gamma mapping is performed in the high confidence zone. ,in The reflectance estimate is for the high confidence region; the reflectance estimate is for the medium confidence region. With guided filter spatial prior Weighted fusion: ,in The reflectance estimate for the medium confidence region, with fusion weights. , The spatial priors obtained by applying a guided filter operator to the reflectivity estimate, The Bayesian posterior term is used; the low-confidence region is reconstructed using the guided filter posterior mean. GF is the guided filter operator; the results from the three regions are merged according to the confidence level indicator mask. ,in , , These are high, medium, and low confidence region indicator masks, respectively. The three masks are mutually exclusive and cover each pixel completely. The output is then merged to restore brightness. ; S6 When brightness is restored Applying CLAHE enhances brightness when the mean value is below the brightness threshold of 40. Otherwise, skip the CLAHE. S7 with The scaling factor is r, which multiplies each channel of the original color image by r to output the enhanced image.

2. The method for enhancing low-light images in coal mines based on imaging confidence as described in claim 1, characterized in that, In step S1, the method for establishing the three-component imaging model is as follows: Formal analysis of the entire imaging chain in underground coal mines is performed, modeling the observed brightness as the sum of the multiplicative coupling of scene reflection and illumination, the independent additive components of glow scattering, and noise: ;in For the observed brightness, ρ is the scene reflectivity, L is the equivalent illumination, G is the equivalent glow independent of the scene texture, and ε is the noise.

3. The method for enhancing low-light images in coal mines based on imaging confidence as described in claim 2, characterized in that, In step S1, the red, green, and blue components of each pixel are analyzed according to the standard luminance formula. Brightness calculations are performed, which converts the image to the brightness domain, allowing the extraction of the brightness channels from the input image. Y .

4. The method for enhancing low-light images in coal mines based on imaging confidence according to claim 3, characterized in that, In step S2, the guided filtering uses the image itself as a guide map to remove sensor noise while preserving the edge structure.

5. The method for enhancing low-light images in coal mines based on imaging confidence according to claim 4, characterized in that, In step S4, the actual confidence level is calculated as follows: Composed of glow component Ĝ, illumination component Noise variance Calculate the variance of the reflectance estimate ,in This is an estimate of reflectivity. To estimate the variance of the glow, Estimating the variance of illumination; Then define pixel-level brightness confidence. Quantitatively characterize the reliability of the reflectance estimate for each pixel. A value close to 1 indicates that the reflectivity estimate is reliable, while a value close to 0 indicates that the signal is overwhelmed by noise. The confidence level for actual use is obtained by scaling and cropping. .

6. The method for enhancing low-light images in coal mines based on imaging confidence according to claim 5, characterized in that, In step S5, the high confidence region Reflectivity estimation is reliable, using gamma mapping. Preserve texture; medium confidence zone The reflectance estimate is fused with the spatial prior obtained from the guided filter using a confidence-weighted method; the higher the confidence level, the greater the weight of the reflectance estimate. Low-confidence regions... It relies entirely on spatial priors and uses the posterior mean of the guided filter as the reconstruction result.

7. A coal mine low-light image enhancement system based on imaging confidence, used in the coal mine low-light image enhancement method as described in any one of claims 1-6, characterized in that, include: Image input module, used to acquire low-light RGB images of underground coal mines; The filtering and denoising module is used to filter and denoise low-light RGB images to obtain a preprocessed image; The confidence level three-region segmentation module is used to divide the preprocessed image into high confidence region, medium confidence region and low confidence region, and perform differential restoration; Enhancement and Color Reconstruction Module: Used for color reconstruction and image enhancement of the image after merging the three regions; Image output module, used to output enhanced RGB images.

Images