A coal gangue dual-energy X-ray transmission image target distribution state recognition method

Abstract

This invention relates to the field of dry coal sorting technology, and more particularly to a method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images. The technical solution includes the following steps: acquiring transmission images of the material to be sorted as it passes through a dual-energy X-ray transmission imaging device, selecting a low-energy image as the input image; labeling the targets in the input image to construct a dataset; constructing a target detection model based on deep learning, training it using the dataset, enabling the model to identify independent targets and different numbers of overlapping and adhered targets in the image. This invention significantly improves the detection accuracy of overlapping and adhered coal gangue targets through refined dataset construction and multi-strategy model optimization; furthermore, it quantitatively characterizes the material distribution state based on multi-dimensional parameters and combines machine learning to accurately identify six distribution forms; finally, the identification results are used for adaptive control of the sorting system, effectively improving sorting efficiency and operational stability.

Description

Technical Field

[0001] This invention relates to the field of dry coal sorting technology, and in particular to a method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images. Background Technology

[0002] Coal sorting is a crucial step in the clean utilization of coal, and the effective separation of coal gangue is of great significance for improving coal quality, reducing transportation costs, and minimizing environmental emissions. Traditional coal gangue sorting methods mainly rely on manual picking or simple mechanical screening, which is not only labor-intensive and inefficient, but also difficult to adapt to complex and changing material conditions, making it difficult to guarantee sorting accuracy.

[0003] With the application of dual-energy X-ray transmission imaging (DE-XRT) technology in mineral identification, this technology, thanks to its sensitive attenuation characteristics for materials with different atomic numbers and its strong penetrating power, can effectively acquire information about the internal structure of coal and gangue. It overcomes the limitations of visible light vision, which is susceptible to interference from surface dust and changes in lighting, providing a new technical approach for dry separation of coal and gangue. However, in practical industrial applications, the processing of DE-XRT images still faces many challenges.

[0004] On the one hand, coal gangue materials on conveyor belts typically exhibit a wide range of particle sizes, highly irregular shapes, and disordered stacking, leading to widespread occlusion and edge adhesion between targets after imaging. Traditional image recognition methods are mostly designed for independent, regular targets, and their ability to detect adhered targets is insufficient, easily resulting in missed detections and false detections. This makes it difficult to accurately obtain the true number and location of targets, thus affecting the subsequent actions of actuators.

[0005] On the other hand, existing sorting systems only focus on the material properties of individual targets, neglecting the impact of the overall material distribution on the sorting effect. For example, uneven loading, excessively dense local accumulation, or overly sparse distribution of materials on the conveyor belt directly affect the X-ray imaging quality, the hit rate of the injection device, and the system's processing capacity. However, there is currently no effective means to quantitatively describe and automatically identify these distribution states, and on-site control still relies on manual experience and observation, making it difficult to achieve closed-loop adaptive control. Therefore, this application proposes a method for identifying the target distribution state in dual-energy X-ray transmission images of coal gangue. Summary of the Invention

[0006] The purpose of this invention is to address the problem of how to accurately identify overlapping and adhered targets based on DE-XRT images, and further to quantitatively characterize and classify the overall distribution state of materials. This invention proposes a method for identifying the target distribution state of coal gangue using dual-energy X-ray transmission images.

[0007] The technical solution of this invention: A method for identifying the target distribution state in dual-energy X-ray transmission images of coal gangue, comprising the following steps:

[0008] S1. Collect transmission images of the material to be sorted as it passes through a dual-energy X-ray transmission imaging device, and select the low-energy image as the input image.

[0009] S2. Label the targets in the input image to construct a dataset; wherein the labeling categories include independent target categories and multiple overlapping and connected target categories containing different numbers of targets.

[0010] S3. Construct a deep learning-based target detection model, train it using the dataset, and enable the model to identify independent targets and different numbers of overlapping and connected targets in the image, and output the category, location and quantity information of each target.

[0011] S4. Based on the trained target detection model, detect the input image to be identified, distinguish independent targets from overlapping and connected targets according to the category of the detection results, and extract the spatial distribution information of all targets.

[0012] S5. Based on the detection results and spatial distribution information, calculate multiple distribution state parameters used to quantitatively describe the overall distribution state of the material;

[0013] S6. Take the calculated multiple distribution state parameters as input features and input them into the pre-trained distribution pattern recognition model. Output the distribution pattern recognition result of the material in the current image.

[0014] Optionally, in step S2, the overlapping and sticky target categories are specifically divided into categories Num2 to NumY, where the NumY category represents overlapping and sticky targets containing more than or equal to Y targets, and Y is a preset integer greater than or equal to 2.

[0015] Optionally, the value of Y is 6, that is, the overlapping and sticky target categories include Num2, Num3, Num4, Num5 and Num6, wherein the Num6 category represents overlapping and sticky targets containing 6 or more targets.

[0016] Optionally, in step S2, before annotating the image, a step of preprocessing the input image to assist in annotation is also included: binarizing the image and filtering out connected components with an area smaller than a preset threshold through connected component analysis to reduce debris and noise interference.

[0017] Optionally, when training the object detection model in step S3, at least one of the following optimization strategies may also be included:

[0018] Data augmentation of training images, including horizontal and / or vertical flipping;

[0019] Adjust the size of the input image for the model;

[0020] To account for the difference in the number of samples in different categories, different weight coefficients are set for different categories in the loss function;

[0021] Embed a receptive field enhancement module into the model.

[0022] Optionally, the receptive field enhancement module includes multiple parallel convolutional branches, each using convolutional kernels of different sizes and / or dilated convolutions with different dilation rates to extract spatial features at different scales, and the outputs of each branch are fused.

[0023] Optionally, in step S5, the distribution state parameters include distribution density for describing the overall density of the material, adhesion ratio for describing the overlap between targets, dispersion for describing the lateral dispersion of the targets, and distribution deviation for describing the overall positional offset of the targets.

[0024] Optionally, the distribution state parameters may further include one or more of the following parameters: left and right offset rate, target centroid coordinates, density variance, maximum region density, horizontal distribution entropy, vertical distribution entropy, minimum bounding rectangle ratio, effective occupancy rate, number of segmentation gains, and segmentation gain ratio.

[0025] Optionally, in step S6,

[0026] The distribution pattern recognition model is a machine learning-based classification model, including any one of support vector machines, fully connected neural networks, or decision trees;

[0027] The output distribution pattern recognition results include one of six preset categories: left-leaning, right-leaning, centered, sparse, moderate, and dense.

[0028] Optionally, after outputting the distribution pattern identification result, the method further includes using the identification result to guide at least one control operation among feeding adjustment, cloth adjustment, conveyor belt operating parameter adjustment, blowing rhythm adjustment, and abnormal working condition early warning of the sorting system.

[0029] Compared with the prior art, this application includes at least one of the following beneficial technical effects:

[0030] This invention effectively reduces debris and noise interference by performing refined category labeling on independent targets and different numbers of overlapping and adhered targets, and by using binarization processing and connected component filtering to assist in labeling. This provides high-quality labeled data for subsequent model training and improves the model's adaptability to complex material morphologies.

[0031] This invention enhances the feature extraction capability for multi-scale targets and overlapping targets by coordinating and optimizing multiple strategies such as data expansion, input size adjustment, category weight reconstruction, and receptive field enhancement module embedding. It also improves the detection accuracy of independent targets and overlapping targets, and provides accurate target location and quantity information for subsequent distribution state parameter calculation.

[0032] This invention constructs a multi-dimensional parameter system based on target detection results, realizing a quantitative description of the material distribution state and avoiding the subjectivity and instability of traditional methods that rely on human experience for qualitative judgment.

[0033] This invention uses the calculated distribution state parameters as input features and identifies six distribution forms through a machine learning classification model. Compared with the method of directly classifying the original image, it improves the accuracy and generalization ability of the identification and provides clear distribution state discrimination results for the sorting system.

[0034] In summary, this invention significantly improves the detection accuracy of overlapping and adhering coal gangue targets through refined dataset construction and multi-strategy model optimization; furthermore, it quantitatively characterizes the material distribution state based on multi-dimensional parameters and combines machine learning to accurately identify six distribution forms; finally, the identification results are used for adaptive control of the sorting system, effectively improving sorting efficiency and operational stability. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the specific execution steps of the present invention.

[0036] Figure 2 This is a schematic diagram of the distribution state parameterization characterization and distribution form identification process of the present invention.

[0037] Figure 3 The example image for the dataset annotation of this invention shows the annotation methods for independent targets and different numbers of overlapping and connected targets.

[0038] Figure 4 This is a schematic diagram of the target detection model of the present invention, wherein MB_RFEM is the receptive field enhancement module.

[0039] Figure 5 This is an example diagram of the distribution state detection results of the present invention, demonstrating the model's detection performance on independent targets and overlapping and adhered targets. Detailed Implementation

[0040] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0041] Example: The present invention proposes a method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images. The method is described in detail below.

[0042] Step 1: Image Acquisition and Target Labeling

[0043] First, transmission images of the materials to be sorted are acquired as they pass through a dual-energy X-ray transmission imaging device. In practical applications, before the materials are conveyed to the imaging device via a conveyor belt, the system acquires each frame of image data and selects the low-energy image as the input image for subsequent processing. Each image frame corresponds to a set of image data showing the material distribution.

[0044] like Figure 3 As shown, targets in the input image are labeled, including independent targets and overlapping / adhesive targets. To improve recognition accuracy, the labeling process categorizes targets based on their number. The specific labeling rules are as follows: Num1 represents independent targets, and Num2 to Num6 represent overlapping / adhesive targets containing 2 to 6 targets, respectively. When an overlapping / adhesive target contains more than 6 targets, it is uniformly classified into the Num6 category.

[0045] To aid in annotation, the input image is binarized and connected component removal is performed to assist in labeling overlapping and bonded targets. Specifically, a preset threshold is used to segment the foreground of the image, and connected component regions with an area smaller than the set threshold are deleted to reduce the impact of debris and noise on annotation accuracy. Through the above annotation, a dataset containing independent targets and multiple bonded targets is established.

[0046] Step 2: Construction and Optimization of the Object Detection Model

[0047] Using the labeled image dataset described above, an object detection model is constructed. For example... Figure 4 As shown, this model employs an object detection algorithm based on the YOLOv10s architecture, embedding a receptive field enhancement module (labeled MB_RFEM in the figure). The YOLO algorithm can simultaneously identify independent and overlapping objects in an image, outputting information such as the category, location coordinates, and number of objects in the detection box. The construction of the object detection model includes network initialization, loss function design, and input of training data.

[0048] To improve the accuracy of target detection, the following optimization strategy is adopted:

[0049] (1) Data augmentation: The training images are flipped horizontally and vertically to increase the diversity of data and improve the model’s adaptability to different image conditions.

[0050] (2) Input size adjustment: The size of the model input image is adjusted from the first size to the second size to enhance the ability to express multi-scale target features, so that the model can adapt to material targets of different sizes.

[0051] (3) Class weight reconstruction: To address the class imbalance problem caused by the difference in the number of samples of different classes, different weight coefficients are set for different classes in the loss function to ensure the accuracy of the target detection results.

[0052] (4) Receptive Field Enhancement: A receptive field enhancement module is introduced into the object detection model to enhance the ability to extract local details and global contextual features. For example... Figure 4 As shown, this receptive field enhancement module includes multiple parallel convolutional branches. Each branch employs different kernel sizes and / or dilated convolutions to extract spatial features at different scales. The outputs of each branch are then concatenated and fused to obtain an enhanced feature map. This is particularly effective in identifying target boundaries when dealing with small or overlapping targets.

[0053] Step 3: Target identification and distribution state parameter calculation

[0054] The trained object detection model is used to detect objects in the input image to be identified, and the detection results are obtained. The detection results are as follows: Figure 5 As shown, the model outputs information on the category, location, and quantity of each target.

[0055] The system distinguishes between independent targets and overlapping / adhesive targets based on the category information in the target detection results. Specifically, if the detection result category is independent target (Num1), it is determined to be an independent target; if the detection result category is overlapping / adhesive (Num2 to Num6), it is determined to be an overlapping / adhesive target, and the number of targets contained in the overlapping / adhesive target is determined according to the corresponding category.

[0056] Based on the location and quantity information in the detection results, extract the spatial distribution information of all targets, including the number of targets, target coordinates, target centroid location, and target spatial distribution information.

[0057] Based on the extracted spatial distribution information, multiple distribution state parameters are calculated to quantitatively describe the overall distribution state of the material. The distribution state parameters calculated in this embodiment are shown in Table 1.

[0058] Table 1 Distribution State Parameter System

[0059]

[0060] Step 4: Identification and Classification of Six Distribution Patterns

[0061] The multiple distribution state parameters obtained from the above calculation are used as input features and input into the pre-trained distribution pattern recognition model to output the distribution pattern recognition result of the material in the current image.

[0062] The distribution pattern recognition model is a machine learning classification model based on the input of distribution state parameters, and can employ any of the following: Support Vector Machine (SVM), Fully Connected Neural Network, or Decision Tree. In this embodiment, Support Vector Machine is used as the classification model to classify the distribution state of the materials.

[0063] like Figure 2 As shown, the output distribution pattern recognition results include one of the following six preset categories:

[0064] Leftward: Material distribution is biased towards the left side of the conveyor belt;

[0065] Slightly to the right: The material distribution is biased towards the right side of the conveyor belt;

[0066] Centered: The material is evenly distributed in the center of the conveyor belt;

[0067] Sparse: The material distribution is relatively sparse, and the distance between targets is relatively large;

[0068] Moderate: The material distribution is moderate, the target spacing is uniform, and it is suitable for sorting;

[0069] Dense: The materials are distributed very densely, with almost no gaps between the targets.

[0070] Step 5: Real-time Application and Feedback Mechanism

[0071] After the output distribution pattern identification results are obtained, the identification results are used to guide at least one of the control operations of the sorting system, including feeding adjustment, cloth adjustment, conveyor belt running parameter adjustment, blowing rhythm adjustment, and abnormal working condition early warning.

[0072] Specifically, the control system can automatically adjust the conveyor belt speed, the feeding device rhythm, and the blowing device rhythm according to the material distribution pattern, thereby optimizing the sorting effect and avoiding material blockage or reduced sorting efficiency. By feeding back the distribution pattern recognition results to the sorting control system, the equipment can adaptively adjust parameters according to different feeding states, thereby reducing the risk of mis-sorting and missed sorting, and improving the stability and efficiency of the sorting system.

[0073] It is worth noting that, in the dataset construction phase, this invention performs refined category labeling on independent targets and overlapping targets containing 2 to 6 targets, and unifies targets with more than 6 targets into the same category, enabling the model to learn the morphological features of targets under different degrees of adhesion. In the model optimization phase, four strategies are employed: data augmentation, input size adjustment, category weight reconstruction, and receptive field enhancement module embedding, which improve the detection accuracy of multi-scale targets and overlapping targets. In the parameter calculation phase, the number of targets, coordinates, and centroid positions are extracted based on the detection results, and the distribution density and overlap / adhesion ratio are defined. Fourteen distribution state parameters, including rate, dispersion, distribution deviation, left and right offset rate, density variance, maximum area density, horizontal distribution entropy, vertical distribution entropy, minimum bounding rectangle ratio, effective occupancy rate, segmentation gain number, and segmentation gain ratio, are used to quantitatively describe the density, offset trend, uniformity, and compactness of materials. In the distribution identification stage, the above parameters are used as input features to the machine learning classification model, which outputs six distribution forms: left-biased, right-biased, centered, sparse, moderate, and dense. This enables the sorting system to adjust feeding, regulate conveyor belt speed, and control the blowing rhythm according to the incoming material status.

[0074] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for identifying the target distribution state in dual-energy X-ray transmission images of coal gangue, characterized in that, Includes the following steps: S1. Collect the transmission image formed when the material to be sorted passes through the dual-energy X-ray transmission imaging device, and select the low-energy image as the input image. S2. Label the targets in the input image and construct the dataset; S3. Construct a deep learning-based target detection model, train it using a dataset, and enable the model to identify independent targets and different numbers of overlapping and connected targets in the image, and output the category, location and quantity information of each target. S4. Based on the trained target detection model, detect the input image to be identified, distinguish independent targets from overlapping and connected targets according to the category of the detection results, and extract the spatial distribution information of all targets. S5. Based on the detection results and spatial distribution information, calculate multiple distribution state parameters used to quantitatively describe the overall distribution state of the material; S6. Take the calculated multiple distribution state parameters as input features and input them into the pre-trained distribution pattern recognition model. Output the distribution pattern recognition result of the material in the current image.

2. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 1, characterized in that, In step S2, the labeling categories include independent target categories and multiple overlapping and sticky target categories containing different numbers of targets. The overlapping and sticky target categories are specifically divided into categories Num2 to NumY, where the NumY category represents overlapping and sticky targets containing more than or equal to Y targets, and Y is a preset integer greater than or equal to 2.

3. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 2, characterized in that, The value of Y is 6, meaning that the overlapping and sticky target categories include Num2, Num3, Num4, Num5 and Num6, where the Num6 category represents overlapping and sticky targets containing 6 or more targets.

4. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 1, characterized in that, In step S2, before annotating the image, the step of preprocessing the input image to assist in annotation is also included: binarizing the image and filtering out connected components with an area smaller than a preset threshold through connected component analysis to reduce debris and noise interference.

5. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 1, characterized in that, When training the object detection model in step S3, at least one of the following optimization strategies is also included: Data augmentation of training images, including horizontal and / or vertical flipping; Adjust the size of the input image for the model; To account for the difference in the number of samples in different categories, different weight coefficients are set for different categories in the loss function; Embed a receptive field enhancement module into the model.

6. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 5, characterized in that, The receptive field enhancement module includes multiple parallel convolutional branches, each using convolutional kernels of different sizes and / or dilated convolutions with different dilation rates to extract spatial features at different scales, and then fuses the outputs of each branch.

7. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 1, characterized in that, In step S5, the distribution state parameters include distribution density for describing the overall density of the material, adhesion ratio for describing the overlap between targets, dispersion for describing the lateral dispersion of the targets, and distribution deviation for describing the overall positional offset of the targets.

8. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 7, characterized in that, The distribution state parameters further include one or more of the following parameters: left and right offset rate, target centroid coordinates, density variance, maximum region density, horizontal distribution entropy, vertical distribution entropy, minimum bounding rectangle ratio, effective occupancy rate, segmentation gain number, and segmentation gain ratio.

9. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 1, characterized in that, In step S6, The distribution pattern recognition model is a machine learning-based classification model, including any one of support vector machines, fully connected neural networks, or decision trees; The output distribution pattern recognition results include one of six preset categories: left-leaning, right-leaning, centered, sparse, moderate, and dense.

10. The method for identifying the target distribution state of coal gangue in dual-energy X-ray transmission images according to claim 1, characterized in that, After outputting the distribution pattern identification result, the method also includes using the identification result to guide at least one control operation in the sorting system, including feeding adjustment, cloth adjustment, conveyor belt running parameter adjustment, blowing rhythm adjustment, and abnormal working condition early warning.

Images