Coal mine drilling center positioning method based on image processing and deep learning

Abstract

The invention discloses a coal mine drilling center positioning method based on image processing and deep learning. The method comprises the following steps: performing traditional image enhancement on an original pipeline image; generating a composite image based on an extended data synthesis method to highly approach a real borehole image; constructing a data set including an original acquisition image, a traditional enhanced image and a synthetic image, performing joint optimization on a drilling pipeline center point detection model based on YOLO, and outputting coordinates of a drilling center point; the model volume is compressed through a model pruning method, and the reasoning calculation amount is reduced; and the efficiency of the pipeline center point detection model is improved by applying a model dynamic quantification method. According to the invention, accurate positioning of the center point of the drilling image can be realized, and the positioning precision is obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal mine drilling imaging and center positioning, and relates to a coal mine drilling center positioning method based on image processing and deep learning. Background Art

[0002] The internal environment of a coal mine borehole is complex and constantly changing, often affected by factors such as insufficient or uneven lighting, equipment installation angle deviations, borehole deformation, and the harsh environment unique to coal mines, such as dust and humidity. These factors can cause significant distortion, offset, and noise in the acquired borehole images. This variability in image quality places extremely high demands on the accuracy and robustness of the borehole image center point positioning.

[0003] In coal mines, center positioning in borehole images is not only fundamental to pipeline image expansion but also directly impacts the accuracy of subsequent deformation detection, crack analysis, and repair decisions. Therefore, a novel center positioning algorithm is urgently needed that comprehensively considers the complexity of the coal mine drilling environment and the diversity of image characteristics, ensuring high accuracy while also possessing robustness and real-time performance to comprehensively address the challenges of existing technologies.

[0004] In the existing technology, commonly used drilling center positioning methods are mainly divided into two categories: manual judgment and algorithms based on traditional image processing. However, these methods have significant shortcomings in practical applications:

[0005] Subjectivity in manual positioning: Manual judgment relies on operator experience and is difficult to standardize. Due to subjective factors, positioning results can exhibit significant deviations. Especially when processing large numbers of images, manual positioning is not only inefficient but can also lead to geometric distortion of the subsequent unfolded image due to human error, compromising the accuracy of the unfolded image and the effectiveness of the analysis.

[0006] Limitations of Traditional Algorithms: While algorithms based on traditional image processing techniques such as edge detection and the Hough transform can provide a certain level of positioning accuracy under ideal conditions, their performance degrades significantly in complex coal mine environments. For example, in conditions of uneven lighting or high image noise, traditional algorithms struggle to reliably extract feature points or edge information. Furthermore, these algorithms often require complex parameter tuning and are computationally intensive, making them difficult to meet the real-time requirements of practical applications. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a coal mine drill hole center positioning method based on image processing and deep learning, which realizes the accurate positioning of the center point of the drill hole image through data enhancement and deep learning technology, combined with a target detection algorithm based on deep learning.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A method for locating the center of a coal mine drill hole based on image processing and deep learning includes the following steps:

[0010] S1, using image acquisition equipment to capture the internal structure of the coal mine borehole to obtain the original pipeline image, and using traditional image enhancement methods to enhance the original pipeline image;

[0011] S2, based on the traditional image enhancement operation, uses the augmented data synthesis method based on the generative adversarial network to generate realistic synthetic images that are highly close to the real drilling images in terms of appearance and texture;

[0012] S3, based on the original and enhanced pipeline image data, constructs a borehole image dataset covering multiple working conditions. This dataset includes original acquired images, traditional enhanced images, and images synthesized by the generative adversarial process. This dataset expands the number and diversity of training samples and jointly optimizes a borehole pipeline center point detection model based on YOLO. The model input is the original and enhanced pipeline images, and the output is the coordinates of the borehole center point.

[0013] S4, after completing the training of the borehole pipe center point detection model based on YOLO, the model pruning method is used to remove redundant channels, convolution kernels, or neurons according to the network structure of the pipe center point detection model, thereby compressing the model size and reducing the amount of inference calculations;

[0014] S5, based on model pruning, applies the model dynamic quantization method to convert the pipeline center point detection model weights originally represented by floating-point numbers into weights represented by low-bit integers, thereby improving the efficiency of the pipeline center point detection model.

[0015] The present invention also includes the following technical features:

[0016] Specifically, in S1, the enhancement of the original pipeline image includes rotation, translation, scaling, mirror flipping and noise addition, which can be expressed as:

[0017] I rot =R(θ)·I (1)

[0018] I trans (x,y)=I(x+Δx,y+Δy) (2)

[0019] I scale =S(α)·I (3)

[0020] I flip (x,y)=I(-x,y) or I(x,-y) (4)

[0021]

[0022] Where I is the original image, R(θ) is the rotation matrix, θ is the rotation angle, Δx and Δy are the translation amounts, S(α) is the scaling matrix, and α is the scaling factor. The mean is 0 and the variance is σ 2 Gaussian noise.

[0023] Specifically, in S2, the expanded data synthesis method includes a generator G and a discriminator D. The generator G is used to generate synthetic samples with drilling image characteristics from random noise or prior image distribution, and the discriminator D is used to determine whether the image originates from real collected data or is generated by the generator; through the adversarial optimization of the generator and the discriminator during the training process, the synthetic image finally generated can be highly close to the real drilling image in terms of appearance and texture, further enhancing the generalization and robustness of the model to images under complex working conditions.

[0024] Specifically, in S2, the generator model is:

[0025] G(z):z→I synthetic (6)

[0026] Where z is the random noise vector, I synthetic is the generated synthetic image;

[0027] The discriminator model is:

[0028] D(I):I→[0,1] (7)

[0029] Among them, D(I) is the discriminator output, which is used to determine whether the image I is a real image.

[0030] Specifically, in S2, the loss function of the expanded data synthesis method is:

[0031]

[0032] By training synthetic images generated by the augmented data synthesis method together with real images, the generalization ability of the model in complex environments can be improved.

[0033] Specifically, in S3, the input of the model is expressed as:

[0034]

[0035] Where H is the image height, W is the image width, and C is the number of channels;

[0036] The coordinates of the drilling center point output by the model are expressed as:

[0037]

[0038] Among them, (x center ,y center ) are the coordinates of the drilling center point.

[0039] Specifically, in S3, the overall loss function of the pipeline center point detection model is:

[0040] L=L coord +L conf (11)

[0041] Among them, l coord is the coordinate loss, l conf is the confidence loss.

[0042] Specifically, in S4, the specific operation of removing redundant channels, convolution kernels or neurons in the network structure of the pipeline center point detection model by the model pruning method is described as follows:

[0043] W ′ =Prune(W,∈) (12)

[0044] Where W is the weight matrix of the pipeline center point detection model, ∈ is the pruning threshold, which indicates the minimum absolute value limit of the deleted weight; W ′ The pruned weight matrix of the model makes the pipeline center point detection model more sparse and significantly reduces the computational burden.

[0045] Specifically, in S5, the application model dynamic quantization method is specifically expressed as follows:

[0046] W q =Quantize(W,n) (13)

[0047] Where n is the number of quantization bits, W q is the weight matrix of the pipeline center point detection model after quantization, and W is the weight matrix before quantization.

[0048] Compared with the prior art, the present invention has the following technical effects:

[0049] (1) By adopting diverse data augmentation strategies and an optimized deep learning target detection model, the present invention significantly improves the positioning accuracy of the borehole center point. Data augmentation expands the model's adaptability to diverse scenarios and complex environments, ensuring accurate positioning of the center point despite the influence of lighting conditions and borehole deformation, thereby providing a stable reference point for subsequent image expansion and analysis.

[0050] (2) This invention combines the advantages of automated data processing technology and deep learning feature extraction to demonstrate good robustness in extreme environments such as low light and uneven imaging. By optimizing the model architecture and loss function design, the algorithm can maintain stability and consistency when dealing with noise, high-contrast changes, and non-ideal drilling morphology.

[0051] (3) The lightweight YOLO target detection model enables the present invention to run efficiently on embedded devices. By optimizing the model structure and improving the inference speed, the real-time positioning of the drilling center point is achieved, meeting the real-time and efficiency requirements of actual coal mine production scenarios and facilitating large-scale deployment of embedded systems.

[0052] (4) The high-precision center positioning point provided by the present invention provides reliable technical support for subsequent tasks such as coal mine image expansion, crack analysis, and pipeline deformation detection. The accuracy of the center point directly affects the effectiveness of subsequent tasks. With the help of this invention, the overall performance of the entire coal mine image processing system can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a framework diagram of the coal mine drilling center positioning method based on image data enhancement.

[0054] Figure 2 Augment and extend images for pipeline data.

[0055] Figure 3 This is the rendering of coal mine drilling center positioning based on deep learning. DETAILED DESCRIPTION

[0056] This paper proposes an efficient and accurate center point positioning algorithm based on in-depth research into the diversity and complexity of coal mine drilling data, as well as the technical difficulties in image feature extraction and analysis. This algorithm fully combines the advantages of traditional image processing technology and deep learning models, and accurately locates the center point of coal mine drilling holes through multiple data enhancement methods and target detection optimization. This method not only improves the stability and reliability of imaging results, but also provides a precise center reference point for planar expansion of coal mine pipeline images and subsequent inner wall analysis, significantly improving the intelligent level of coal mine pipeline structure monitoring and maintenance. The present invention has the following characteristics:

[0057] Enhanced data diversity: Through traditional methods such as rotation, translation, and scaling, as well as data augmentation technology based on generative adversarial networks, the model's adaptability to complex scenarios is improved.

[0058] Efficient center positioning algorithm: Using a lightweight YOLO model, it meets the operating requirements of embedded devices while ensuring high accuracy, providing a key reference point for subsequent image expansion.

[0059] Robustness enhancement: Through automated data augmentation and optimized model architecture, the algorithm's applicability in low-light and uneven imaging conditions is improved.

[0060] Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent modifications made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0061] Example 1:

[0062] This embodiment provides a method for locating the center of a coal mine drill hole based on image processing and deep learning, in order to solve the problem of image diversity caused by factors such as equipment position, angle and lighting changes during the process of coal mine drilling photography, and to improve the robustness of the coal mine drilling center positioning process; Figure 1 As shown, the specific steps include:

[0063] S1, collects the internal structure image of the coal mine borehole through the camera module installed on the drilling detection equipment, such as Figure 2 As shown in the figure, in order to enhance the generalization ability of the coal mine drilling center positioning method, the data enhancement method is used to enhance the original pipeline image collected by the image acquisition device. The traditional image enhancement method is used to rotate, translate, scale, mirror flip and add noise to the original pipeline image, which can be expressed as:

[0064] I rot =R(θ)·I (1)

[0065] I trans (x,y)=I(x+Δx,y+Δy) (2)

[0066] I scale =S(α)·I (3)

[0067] I flip (x,y)=I(-x,y) or I(x,-y) (4)

[0068]

[0069] Where I is the original image, R(θ) is the rotation matrix, θ is the rotation angle, Δx and Δy are the translation amounts, S(α) is the scaling matrix, and α is the scaling factor. The mean is 0 and the variance is σ 2 Gaussian noise.

[0070] S2, based on traditional image enhancement operations, uses an augmented data synthesis method based on generative adversarial networks (GANs) to further expand the collected image data and generate realistic synthetic images. This method consists of two parts: a generator G and a discriminator D. The generator G is used to generate synthetic samples with drilling image characteristics from random noise or a prior image distribution, while the discriminator D is used to determine whether the image originates from real collected data or is generated by the generator. Through adversarial optimization between the generator and discriminator during training, the resulting synthetic images are highly similar to real drilling images in terms of appearance and texture, further enhancing the model's generalization and robustness to images under complex working conditions.

[0071] The generator model is:

[0072] G(z):z→I synthetic (6)

[0073] Where z is the random noise vector, I synthetic The generated synthetic image.

[0074] The discriminator model is:

[0075] D(I):I→[0,1] (7)

[0076] Among them, D(I) is the discriminator output, which is used to determine whether the image I is a real image.

[0077] The loss function of the augmented data synthesis method is:

[0078]

[0079] By training synthetic images generated by the augmented data synthesis method together with real images, the generalization ability of the model in complex environments can be improved.

[0080] S3, based on the original and enhanced pipeline image data, further constructed a borehole image dataset covering multiple working conditions. This dataset includes original acquired images, traditional enhanced images, and images synthesized by the generative adversarial process, significantly expanding the number and diversity of training samples. This is used to jointly optimize the borehole pipeline center point detection model based on YOLO. The model input is the original and enhanced pipeline images, and the output is the coordinates of the borehole center point, such as Figure 3 shown.

[0081] The input of the model can be expressed as:

[0082]

[0083] Among them, H is the image height, W is the image width, and C is the number of channels.

[0084] The coordinates of the drilling center point output by the model can be expressed as:

[0085]

[0086] Among them, (x center ,y center ) are the coordinates of the drilling center point.

[0087] The overall loss function of the pipeline center point detection model is:

[0088] L=L coord +L conf (11)

[0089] Among them, L coord is the coordinate loss, L conf is the confidence loss.

[0090] S4. After completing the YOLO-based drill hole center point detection model training, in order to deploy the pipeline center point detection model in resource-constrained embedded devices, it is necessary to reduce the model's computational and storage overhead while maintaining its performance indicators. To this end, this embodiment uses a model pruning method to remove redundant channels, convolution kernels, or neurons in the pipeline center point detection model according to its network structure, thereby compressing the model size and reducing the amount of inference calculations. The specific operations can be described as follows:

[0091] W ′ =Prune(W,∈) (12)

[0092] Where W is the weight matrix of the pipeline center point detection model, ∈ is the pruning threshold, which represents the minimum absolute value limit for deleting weights. ′ The pruned weight matrix of the model makes the pipeline center point detection model more sparse and significantly reduces the computational burden.

[0093] S5, in order to further improve the efficiency of the pipeline center point detection model, the model dynamic quantization method is applied on the basis of pruning to convert the pipeline center point detection model weights originally represented by floating-point numbers into weights represented by low-bit integers, which can be expressed as:

[0094] W q =Quantize(W,n) (13)

[0095] Where n is the number of quantization bits, W q is the weight matrix of the pipeline center point detection model after quantization, and W is the weight matrix before quantization.

[0096] Through the above optimization technology, the pipeline center point detection model can significantly reduce its complexity. The optimized model running in embedded devices can not only complete the detection of drilling center points in real time and accurately, but also save hardware costs, ensure the model has good scalability and industrial deployment capabilities, and provide real-time and accurate center positioning services for on-site drilling camera systems.

Claims

1. A method for locating the center of a coal mine drill hole based on image processing and deep learning, characterized in that: The following steps are involved: S1, using image acquisition equipment to capture the internal structure of the coal mine borehole to obtain the original pipeline image, and using traditional image enhancement methods to enhance the original pipeline image; S2, based on the traditional image enhancement operation, uses the augmented data synthesis method based on the generative adversarial network to generate realistic synthetic images that are highly close to the real drilling images in terms of appearance and texture; S3, based on the original and enhanced pipeline image data, constructs a borehole image dataset covering multiple working conditions. This dataset includes original acquired images, traditional enhanced images, and images synthesized by the generative adversarial process. This dataset expands the number and diversity of training samples and jointly optimizes a borehole pipeline center point detection model based on YOLO. The model input is the original and enhanced pipeline images, and the output is the coordinates of the borehole center point. S4, after completing the training of the borehole pipe center point detection model based on YOLO, the model pruning method is used to remove redundant channels, convolution kernels, or neurons according to the network structure of the pipe center point detection model, thereby compressing the model size and reducing the amount of inference calculations; S5, based on model pruning, applies the model dynamic quantization method to convert the pipeline center point detection model weights originally represented by floating-point numbers into weights represented by low-bit integers, thereby improving the efficiency of the pipeline center point detection model.

2. The method for coal mine drilling center positioning based on image processing and deep learning according to claim 1, characterized in that: In S1, the enhancement of the original pipeline image includes rotation, translation, scaling, mirror flipping and noise addition, which is expressed as: I rot =R(θ)·I (1) I trans (x,y)=I(x+Δx,y+Δy) (2) I scale =S(α)·I (3) I flip (x,y) = I(-x,y) or I(x,-y) (4) Where I is the original image, R(θ) is the rotation matrix, θ is the rotation angle, Δx and Δy are the translation amounts, S(α) is the scaling matrix, and α is the scaling factor. The mean is 0 and the variance is σ 2 Gaussian noise.

3. The method for coal mine drilling center positioning based on image processing and deep learning according to claim 1, characterized in that: In S2, the expanded data synthesis method includes a generator G and a discriminator D. The generator G is used to generate synthetic samples with drilling image characteristics from random noise or prior image distribution, and the discriminator D is used to determine whether the image originates from real collected data or is generated by the generator; through the adversarial optimization of the generator and the discriminator during the training process, the synthetic image finally generated can be highly close to the real drilling image in terms of appearance and texture, further enhancing the generalization and robustness of the model to images under complex working conditions.

4. The method for locating the center of a coal mine drill hole based on image processing and deep learning according to claim 3, wherein: In S2, the generator model is: G(z):z→I synthetic (6) Where z is the random noise vector, I synthetic is the generated synthetic image; The discriminator model is: D(I):I→[0,1] (7) Among them, D(I) is the discriminator output, which is used to determine whether the image I is a real image.

5. The method for locating the center of a coal mine drill hole based on image processing and deep learning according to claim 3, characterized in that: In S2, the loss function of the expanded data synthesis method is: By training synthetic images generated by the augmented data synthesis method together with real images, the generalization ability of the model in complex environments can be improved.

6. The method for coal mine drilling center positioning based on image processing and deep learning according to claim 1, characterized in that: In S3, the input of the model is represented as: Where H is the image height, W is the image width, and C is the number of channels; The coordinates of the drilling center point output by the model are expressed as: Among them, (x center ,y center ) are the coordinates of the drilling center point.

7. The method for locating the center of a coal mine drill hole based on image processing and deep learning according to claim 6, characterized in that: In S3, the overall loss function of the pipeline center point detection model is: L=L coord +L conf (11) Among them, L coord is the coordinate loss, L conf is the confidence loss.

8. The method for locating the center of a coal mine drill hole based on image processing and deep learning according to claim 1, wherein: In S4, the specific operation of removing redundant channels, convolution kernels or neurons in the network structure of the pipeline center point detection model by the model pruning method is described as follows: IN ′ =Prune(W,∈) (12) Where W is the weight matrix of the pipeline center point detection model, ∈ is the pruning threshold, which indicates the minimum absolute value limit of the deleted weight; W ′ The pruned weight matrix of the model makes the pipeline center point detection model more sparse and significantly reduces the computational burden.

9. The method for locating the center of a coal mine drill hole based on image processing and deep learning according to claim 1, characterized in that: In S5, the application model dynamic quantization method is specifically expressed as: W q =Quantize(W,n) (13) Where n is the number of quantization bits, W q is the weight matrix of the pipeline center point detection model after quantization, and W is the weight matrix before quantization.

Images