Coal conveying image sharpening method and device, medium and program product

Abstract

This invention provides a method, apparatus, medium, and program product for enhancing coal conveying images. First, an initial transmittance map is obtained by estimating the initial transmittance based on a dark channel prior. Then, a CNN residual correction network is used to correct the initial transmittance map. Finally, based on an atmospheric scattering model, the restored image is reconstructed using the corrected transmittance map. This coal conveying image enhancement method, by fusing dark channel priors and a lightweight residual correction network, and introducing adaptive atmospheric light estimation and failure region guidance mechanisms, can significantly improve image quality in complex industrial dust environments, providing high-contrast, high-detail-fidelity input for subsequent tasks such as foreign object detection and tear recognition.

Description

Technical Field

[0001] This invention belongs to the field of machine vision enhancement technology, specifically relating to a method, device, medium, and program product for clarifying coal conveying images. Background Technology

[0002] More and more industrial sites are adopting machine vision technology for real-time monitoring of coal conveying systems, such as automatically identifying abnormal conditions like conveyor belt tears and coal blockages. However, during coal conveying in thermal power plants, a large amount of dust particles are generated at the coal chute, transfer points, and conveyor belt operation. These dust particles form a high-concentration suspended particulate environment in the air, causing the image acquisition to exhibit the following characteristics:

[0003] (1) Coal dust and particulate matter cause fog-like scattering;

[0004] (2) Uneven lighting or strong backlighting;

[0005] (3) When the dust density is not uniformly distributed in space, the transmittance estimation will be biased, resulting in unstable noise reduction effect.

[0006] Due to the high dust concentration and complex ambient lighting in the aforementioned industrial sites, the acquired images often suffer from significant image degradation, which severely affects the accuracy and stability of subsequent visual recognition algorithms.

[0007] Currently, common image dehazing methods can be broadly categorized into two types: physics-based methods and deep learning-based methods. Existing dehazing algorithms are mostly based on dark channel prior models, but these methods assume constant transmittance within local regions, making them unsuitable for the clumpy distribution of dust in industrial environments. Deep learning methods, such as AOD-Net, while capable of end-to-end reconstruction, lack physical constraints, exhibit poor generalization ability in unnatural scenes, and require significant computation. Furthermore, current technologies do not incorporate conveyor belt speed information for motion blur kernel physics modeling. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method, apparatus, medium, and program product for clarifying coal conveying images, thereby solving the problem of image dehazing in high-dust industrial environments.

[0009] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0010] A method for enhancing coal conveying images:

[0011] Step 1, for the original observation map The initial transmittance map is obtained by estimating the initial transmittance based on the prior of the dark channel. ;

[0012] Step 2: The following CNN residual correction network is used to process the initial transmittance map. Make corrections to obtain the corrected transmittance map. ,in:

[0013] The input of the network for:

[0014]

[0015] In the formula, This is a guide mask for the failed region. This is the original observation image. This is the initial transmittance diagram;

[0016] It consists of the following three channels:

[0017]

[0018] In the formula, For pixel variables, , , Each pixel Detection masks for white areas, highlight areas, and areas with uneven dust distribution;

[0019] The network consists of 4 convolutional layers, and its processing procedure is as follows:

[0020] S1, The intermediate feature map is obtained after the first three convolutional layers. ;

[0021] S2, utilizing attention weights right Recalibrate the channel dimensions:

[0022]

[0023] S3, Feature Map The residual transmittance map is then generated through a fourth convolution layer. :

[0024]

[0025] The corrected transmittance of each pixel for:

[0026]

[0027] Step 3: Based on the atmospheric scattering model, use the corrected transmittance map. Reconstructed restoration image .

[0028] Further, in step 1, the initial transmittance is calculated as follows:

[0029]

[0030] In the formula, To adjust the parameters, For color channel variables, when used as a superscript, it indicates that the component of the corresponding channel is being retrieved. Atmospheric light value, Represented by pixels A local window centered on the user.

[0031] Further, the atmospheric light value is calculated as follows:

[0032]

[0033]

[0034] In the formula, It is the set of pixels with the highest brightness in the image. and Each pixel saturation and brightness For image The maximum brightness in the range, To control for the standard deviation parameter of the effect of saturation, The standard deviation parameter is used to control for the effect of brightness.

[0035] Furthermore, the detection masks for white areas, highlight areas, and dust unevenness areas are calculated as follows:

[0036]

[0037]

[0038]

[0039] In the formula, For color channel variables, when used as a superscript, it indicates that the component of the corresponding channel is being retrieved. Scale factor; This is an indicator function; its value is 1 when the condition within its scope is true, and 0 otherwise. The brightness threshold; Indicates at pixel point The gradient of the initial transmittance map. It is the stability constant.

[0040] Furthermore, the attention weights Adjust the attention module via the following channels:

[0041]

[0042] In the formula, Indicates global average pooling. It is the ReLU activation function. It is the Sigmoid activation function. It is a linear mapping function. and For weights, and For bias.

[0043] Furthermore, the objective function of the network is:

[0044]

[0045]

[0046]

[0047]

[0048] In the formula, and These are the weighting coefficients. To reconstruct the loss, To smooth out the loss, To preserve loss a priori, This is a diagram showing the actual transmittance. This is the gradient plot of the corrected transmittance map. For pixels Confidence weights This represents the total number of pixels.

[0049] Furthermore, step 3 includes:

[0050] Step 3.1, proceed with the restoration and reconstruction as follows:

[0051]

[0052] In the formula, This is the lower limit of transmittance.

[0053] Step 3.2, restore the image Perform post-processing filtering:

[0054]

[0055] In the formula, For guided filtering, the original observation map is used. For guiding purposes, The radius of the filtering window. For regularization parameters, This is the filtered image.

[0056] A computer device, including a memory and a processor;

[0057] The memory is used to store computer programs;

[0058] The processor is used to execute the computer program and, in executing the computer program, implement the above-described coal conveying image sharpening method.

[0059] A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the above-described coal conveying image sharpening method.

[0060] A computer program product includes a computer program that, when executed by a processor, implements the above-described coal conveying image sharpening method.

[0061] The beneficial effects of this invention are as follows:

[0062] (1) This invention provides a method for improving the clarity of coal conveying images. By integrating dark channel priors and lightweight residual correction networks, and introducing adaptive atmospheric light estimation and failure region guidance mechanisms, the image quality can be significantly improved in complex industrial dust environments, providing high contrast and high detail fidelity input for subsequent tasks such as foreign object detection and tear recognition.

[0063] (2) In estimating the initial transmittance, the present invention retains a small amount of natural fog by adding adjustment parameters to avoid excessive defogging.

[0064] (3) This invention provides an adaptive estimation method based on the scattering characteristics of dust particles. By analyzing the relationship between dust concentration and image saturation and brightness, an adaptive estimation formula for atmospheric light value is constructed to solve the problem of inaccurate estimation of atmospheric light value by directly using the brightest value in the image in dust scenes. This method is more in line with the atmospheric photophysical characteristics caused by dust scattering.

[0065] (4) This invention adjusts the attention weights through the channel attention module to ensure that the network prioritizes correcting regions where the physical model may be incorrect in subsequent residual prediction. Through this design, the attention module seamlessly embeds the physical prior (failure region mask) into the data-driven residual learning process, enabling the network to focus more efficiently on key correction regions with limited parameters, ultimately improving the accuracy and robustness of transmittance correction. Attached Figure Description

[0066] Figure 1 This is a flowchart of the coal conveying image sharpening method of the present invention. Detailed Implementation

[0067] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0068] I. Technical Solution

[0069] Reference Figure 1 As shown, the coal conveying image sharpening method of the present invention includes the following:

[0070] 1. Atmospheric scattering model

[0071] Image degradation under weather conditions such as dust and haze can be described using atmospheric scattering models:

[0072]

[0073] In the formula, This is the original observation image (i.e., one with dust, haze, etc.). To restore the image (i.e., to remove dust and haze from the image); This is a transmittance diagram, representing the proportion of light reaching the camera, with values ​​ranging from 0 to 1; This is the atmospheric light value, usually the value of the brightest pixel in the image; For pixel variables, , , They represent , , medium pixel The value at that location.

[0074] 2. Atmospheric light value estimation

[0075] In dusty scenes, the brightest value in the image is directly used as the atmospheric light value. There is a potential problem of inaccurate estimation. To address this, this invention proposes an adaptive estimation method based on the scattering characteristics of dust particles: by analyzing the relationship between dust concentration and image saturation and brightness, an adaptive estimation formula for atmospheric light values ​​is constructed.

[0076]

[0077] In the formula, This is the set of the brightest pixels in the image, typically the top 0.1%. The weight function is expressed as follows:

[0078]

[0079] In the formula, and Each pixel saturation and brightness For image The maximum brightness in the range, To control for the standard deviation parameter of the effect of saturation, To control the standard deviation parameter of brightness, the above weighting function tends to select pixels with high brightness and low saturation, which is consistent with the physical characteristics of atmospheric light caused by dust scattering - atmospheric light areas are usually grayish-white (low saturation) and have high brightness.

[0080] 3. Prior estimation of transmittance in dark channels

[0081] For most dust-free (fog-free) natural images, within a local window, at least one color channel has a very low pixel value (close to 0), that is:

[0082]

[0083] In the formula, Represents pixels The dark channel value, For color channel variables, Indicates the red channel. Indicates a green channel. Indicates the blue channel. Represented by pixels A local window centered on the center. Representing an image exist Components on the channel.

[0084] Substituting the above atmospheric scattering model, we apply local minimum filtering to both sides, and assume that within the local window... If constant, it can be rewritten as:

[0085]

[0086] In the formula, Representing an image exist Components on the channel, for exist Components on the channel.

[0087] Due to the prior of the dark channel, the first term on the right-hand side of the above equation approaches 0, therefore:

[0088]

[0089] To retain a small amount of natural fog, adjust the parameters. The transmittance estimation expression is obtained as follows:

[0090]

[0091] In this embodiment The value is 0.95.

[0092] 4. CNN Residual Correction Network

[0093] 4.1 Because the dark passage prior may fail in certain situations, such as large white areas (e.g., sky, snow), areas of strong light, or areas with uneven dust density, it can lead to... Inaccurate estimates can lead to either "over-defogging" or "residual fog." Therefore, the transmittance calculated using the above transmittance estimation expression should only be used as the initial transmittance. :

[0094]

[0095] To address the issue of "over-dehazing" in dark channels, this invention proposes a physically guided residual learning architecture. This architecture primarily targets the aforementioned failure regions by introducing a lightweight CNN (composed of 4 convolutional layers) to learn residuals. Specifically:

[0096] 4.2 Introducing a failure region guide mask in the CNN input layer The mask consists of three channels:

[0097]

[0098] In the formula, Detect the mask for the white area. Use a mask to detect highlight areas. A mask for detecting uneven dust distribution areas;

[0099] 1) White area detection mask, calculated based on the variance of RGB channels within a local window:

[0100]

[0101] In the formula, It is an exponential function. It is the variance function. Indicates taking , , The variance among the three This is a scaling factor used to adjust the decay rate of the exponential function.

[0102] 2) Highlight area detection mask, calculated based on brightness threshold:

[0103]

[0104] In the formula, This is an indicator function; its value is 1 when the condition within its scope is true, and 0 otherwise. This is the brightness threshold.

[0105] 3) Detection mask for uneven dust areas, calculated based on local transmittance gradient:

[0106]

[0107] In the formula, Indicates at pixel point The gradient of the initial transmittance map is a two-dimensional vector, and its direction points in the direction of the fastest change in transmittance. Its modulus reflects the degree of drastic change in the transmittance of that pixel; It is a very small stability constant.

[0108] 4.3, Guide the mask image for the failed area. Compared with the original observation map and initial transmittance map The three components are concatenated along the channel dimension and used as input to the CNN residual correction network. :

[0109]

[0110] In the formula The scale of the stitched input image is where... and This indicates the height and width of the image.

[0111] Input image After processing by the first three convolutional layers of the CNN residual correction network, the intermediate feature map is obtained:

[0112]

[0113] In the formula This represents the number of channels in the image.

[0114] Then use attention weights right Recalibrate the channel dimensions:

[0115]

[0116] In the formula, " indicates element-wise multiplication.

[0117] Weighted feature map The residual transmittance map is then generated by the final fourth convolution layer. :

[0118]

[0119] Finally, the residual transmittance map was used. Initial transmittance map After correction, the corrected transmittance map is obtained. :

[0120]

[0121] 4.4, Channel Attention Module

[0122] To enhance network attention to failed regions, this invention designs a physically guided channel attention module. This module utilizes a failed region guidance mask map. The statistical characteristics are analyzed, and the weights of the feature channels are dynamically adjusted by fusing statistical information from traditional channel attention and failure region masks. :

[0123]

[0124] In the formula, Indicates global average pooling. It is the ReLU activation function. It is the Sigmoid activation function. It is a linear mapping function. and For weights, and For bias.

[0125] final The weights of each channel simultaneously reflect the importance of data-driven features and the guidance of physical priors, ensuring that the network prioritizes correcting regions where the physical model may err in subsequent residual predictions. Through this design, the attention module seamlessly embeds physical priors (failure region masks) into the data-driven residual learning process, enabling the network to focus more efficiently on key correction regions with limited parameters, ultimately improving the accuracy and robustness of transmittance correction.

[0126] 4.5, The objective function of the above CNN residual correction network It consists of three parts:

[0127]

[0128] In the formula, and These are the weighting coefficients. To reconstruct the loss, To smooth out the loss, The prior loss is preserved; their respective expressions are:

[0129]

[0130] In the formula, This is a graph showing the actual transmittance.

[0131]

[0132] In the formula, This loss is used to ensure the local smoothness of the transmittance map, which is the gradient map of the corrected transmittance map.

[0133]

[0134] In the formula, For pixels The confidence weight is used to ensure that the constraint correction amount in the non-failure region should not be too large. This represents the total number of pixels.

[0135] 5. Image restoration

[0136] 5.1 Based on the atmospheric scattering model and combined with the corrected transmittance map Reconstructing the restored image :

[0137]

[0138] 5.2 To avoid noise amplification caused by excessively low transmittance, a lower limit for transmittance can be set. (Generally taken as 0.1), then the final restoration and reconstruction will be carried out according to the following formula:

[0139]

[0140] 5.3 To address the potential noise issues remaining in the restored image, a post-processing filter based on the physical properties of dust is introduced, and the filtered image is used as the basis for further processing. As the final output image:

[0141]

[0142] In the formula, For guided filtering, the original observation map is used. As a guide map, it preserves edge structure while smoothing noise; The radius of the filtering window determines the size of the neighborhood range that participates in the filtering. This is a regularization parameter (smoothing factor) used to control the balance between the degree of edge preservation and the smoothing intensity of the filter.

[0143] II. Devices, storage media, and software products

[0144] 1. Based on the same inventive concept as the above-mentioned coal conveying image sharpening method, this application also provides an electronic device, which includes a processor and a memory, wherein computer-readable code is stored in the memory, and when the computer-readable code is executed by the processor, the coal conveying image sharpening method of the present invention is implemented.

[0145] The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium can store the operating system and computer-readable code. The computer-readable code includes program instructions that, when executed, cause the processor to perform a coal conveying image sharpening method. The processor provides computational and control capabilities to support the operation of the entire electronic device. The memory provides an environment for the execution of the computer-readable code in the non-volatile storage medium, which, when executed by the processor, causes the processor to perform the coal conveying image sharpening method.

[0146] It should be understood that a processor can be a central processing unit, other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, transistor logic devices, discrete hardware components, etc. Among them, a general-purpose processor can be a microprocessor or any conventional processor.

[0147] 2. This application also provides a readable storage medium, which may be an internal storage unit of the electronic device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, smart memory card, or secure digital card equipped on the electronic device.

[0148] 3. This application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the coal conveying image sharpening method of the present invention.

[0149] This invention is not limited to the above-described embodiments. Any obvious improvements, substitutions, or modifications that can be made by those skilled in the art without departing from the essence of this invention are within the scope of protection of this invention.

Claims

1. A method for enhancing coal conveying images, characterized in that: Step 1, for the original observation map The initial transmittance map is obtained by estimating the initial transmittance based on the prior of the dark channel. ; Step 2: The initial transmittance map is processed using the following CNN residual correction network. Make corrections to obtain the corrected transmittance map. ,in: The input of the network for: In the formula, A mask image for the failed region. This is the original observation image. This is the initial transmittance diagram; It consists of the following three channels: In the formula, For pixel variables, , , Each pixel Detection masks for white areas, highlight areas, and areas with uneven dust distribution; The network consists of four convolutional layers, and its processing procedure is as follows: S1, The intermediate feature map is obtained after the first three convolutional layers. ; S2, utilizing attention weights right Recalibrate the channel dimensions: S3, Feature Map The residual transmittance map is then generated through a fourth convolution layer. : The corrected transmittance of each pixel for: Step 3: Based on the atmospheric scattering model, use the corrected transmittance map. Reconstructed restoration image .

2. The method for enhancing coal conveying images according to claim 1, characterized in that: In step 1, the initial transmittance is calculated as follows: In the formula, To adjust the parameters, For color channel variables, when used as a superscript, it indicates that the component of the corresponding channel is being retrieved. Atmospheric light value, Represented by pixels A local window centered on the user.

3. The method for enhancing coal conveying images according to claim 2, characterized in that: Calculate atmospheric light values ​​as follows: In the formula, It is the set of pixels with the highest brightness in the image. and Each pixel saturation and brightness For image The maximum brightness in the range, To control for the standard deviation parameter of the effect of saturation, The standard deviation parameter is used to control for the effect of brightness.

4. The method for enhancing coal conveying images according to claim 1, characterized in that: Calculate the detection mask for white areas, highlight areas, and dust unevenness areas as follows: In the formula, For color channel variables, when used as a superscript, it indicates that the component of the corresponding channel is being retrieved. Scale factor; This is an indicator function; its value is 1 when the condition within its scope is true, and 0 otherwise. The brightness threshold; Indicates at pixel point The gradient of the initial transmittance map. It is the stability constant.

5. The method for enhancing coal conveying images according to claim 1, characterized in that: The attention weight Adjust the attention module via the following channels: In the formula, Indicates global average pooling. It is the ReLU activation function. It is the Sigmoid activation function. It is a linear mapping function. and For weights, and For bias.

6. The method for enhancing coal conveying images according to claim 1, characterized in that: The objective function of the network is: In the formula, and These are the weighting coefficients. To reconstruct the loss, To smooth out the loss, To preserve loss a priori, This is a diagram showing the actual transmittance. This is the gradient plot of the corrected transmittance map. For pixels Confidence weights This represents the total number of pixels.

7. The method for enhancing coal conveying images according to claim 1, characterized in that: Step 3 includes: Step 3.1, proceed with the restoration and reconstruction as follows: In the formula, This is the lower limit of transmittance; Step 3.2, restore the image Perform post-processing filtering: In the formula, For guided filtering, the original observation map is used. For guiding purposes, The radius of the filtering window. For regularization parameters, This is the filtered image.

8. A computer device, characterized in that: Including memory and processor; The memory is used to store computer programs; The processor is used to execute the computer program and, in executing the computer program, implement the coal conveying image sharpening method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The device contains a computer program that, when executed by a processor, causes the processor to perform the coal conveying image sharpening method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that: The method includes a computer program that, when executed by a processor, implements the coal conveying image sharpening method as described in any one of claims 1 to 7.

Images