Ore particle size measurement method using image completion technology

Abstract

The invention provides an ore particle size measurement method using an image completion technology. The method comprises the following steps: collecting original ore image data; generating an ore mask data set; preprocessing the ore mask data set; an image completion module is adopted for processing; an ore edge completion algorithm EDRE network model is adopted, and the ore edge completion algorithm EDRE network model adopts a network structure of a generator and a discriminator; and ore particle size calculation: after the ore image is completed, obtaining particle size information by using a geometric fitting method. According to the method, through a mode of combining edge completion and granularity detection, on the basis of effectively restoring an ore image structure, high-precision granularity analysis in a shielding scene is realized, and the method has good engineering practicability and generalization ability.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and ore particle size detection, and in particular to a method for detecting ore particle size by image completion of obscured irregular ore particles. Background Art

[0002] In the ore processing and beneficiation process, accurate measurement of ore particle size is an important indicator for evaluating crushing efficiency and screening effects. Traditional particle size detection methods mostly rely on manual screening, which has problems such as cumbersome operation and low detection efficiency. In recent years, with the significant breakthroughs made by deep learning technology in tasks such as image segmentation and target recognition, vision-based particle size detection methods have gradually become mainstream. However, due to the natural irregular morphology of ore particles, complex stacking methods and severe obscuration, existing methods still face many technical bottlenecks under real working conditions. Especially when ore particles are severely occluded and contour information is missing, traditional image segmentation and recognition models often have difficulty in effectively restoring the true boundaries of the obscured ore particles, resulting in large errors and poor stability in the particle size calculation results. Current research focuses on improving the accuracy of image segmentation and target recognition models. However, there is still relatively little research on using image completion technology to restore the missing edge information caused by occlusion to improve the accuracy of particle size detection. CN118864831A discloses a method for obtaining ore semantic segmentation information based on the ore semantic segmentation model SDE-Net. By building an ore body-edge feature fusion network, it effectively segments ore images and resolves the adhesion between ores. CN112419397A relates to an ore particle size classification method based on images and deep neural networks. This method uses an optimized U-Net network model to segment ore images by ore particle size, and then derives the ore particle size distribution through particle size statistics, improving segmentation accuracy under complex backgrounds. CN113570542A provides a method for identifying coal and gangue by completing images based on a Completion-GAN network under occlusion. This method generates a complete image using a Completion-GAN network and then uses existing target algorithms to identify coal and gangue, improving recognition accuracy and making it suitable for target classification tasks. These methods perform well under ideal conditions, but have obvious shortcomings under actual complex working conditions: it is difficult to truly simulate the ore occlusion pattern, the completed edges are not continuous enough, they cannot meet the measurement accuracy requirements, and the tasks are inconsistent. Furthermore, existing research datasets are mostly based on ideal scenarios with minimal or no occlusion, lacking the ability to model and complete specific occlusion scenarios. Therefore, it is particularly important to develop a particle size detection method for occluded irregular ore particles by combining image completion technology. Summary of the Invention

[0003] The purpose of the present invention is to provide an image completion method for ore particle size detection of obscured irregular ore particles, using an ore edge completion algorithm EDRE to restore the lost contour information of the ore, laying the foundation for subsequent high-precision ore particle size detection.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions: a method for measuring ore particle size using image completion technology, (1) Collecting original ore image data: Using a high-performance industrial camera as the core of image acquisition to shoot ore samples under actual working conditions; and using LED white light as a fixed light source for acquisition, mainly to ensure the consistency and uniformity of image acquisition; (2) Generate an ore mask dataset: Use a structural mask to perform occlusion processing on the original image to construct an ore mask dataset. The mask construction method is to first binarize the original ore image to extract the foreground area and generate a basic mask. Then, the ore contour mask is translated and directional overlay is added to simulate the stacking occlusion in real situations. (3) Preprocessing of ore mask dataset: The preprocessing of ore mask dataset mainly includes the processing of mask grayscale image and mask edge image; (4) Using image completion module processing: using a mineral edge completion algorithm EDRE network model, the mineral edge completion algorithm EDRE network model adopts a generator and discriminator network structure, including: convolution block, feature extraction network TE module and discriminator; the generator takes the mask image, edge image and grayscale image as input, and uses the convolution block and self-built feature extraction module TE to encode and extract features from the image; the decoder part is responsible for receiving the high-dimensional features processed by the intermediate module, and gradually maps the high-dimensional features back to the image space through the convolution operation, and outputs the mineral image after reconstructing the edge loss. Finally, the completed mineral image and the original edge image are sent to the discriminator to judge whether the output result is reasonable; (5) Ore particle size calculation: After completing the ore image, the particle size information is obtained using the geometric fitting method.

[0005] Furthermore, the industrial camera in step (1) has a high-speed shutter, a low-noise sensor and an automatic exposure control function, and can achieve clear image capture in complex mining environments such as dust and high vibration, providing a reliable raw data basis for subsequent image completion and particle size detection.

[0006] Furthermore, in step (3): a. Mask grayscale image processing: Input the original ore image, convert the original image into a mask grayscale image, then use the mask image to perform pixel-by-pixel multiplication on the grayscale image, and finally output the grayscale image after mask processing; the specific process is expressed as follows:

[0007] wherein B is an image mask, is a mask edge image, is a final output mask edge image.

[0008] Further, in the step (4): a. The convolution module is composed of convolution, spectral normalization (SN) and ReLU activation function, adopts a symmetrical structure in the model, and can realize good structural coordination between feature encoding and decoding, and between the main network and the branch module; it can suppress gradient explosion, accelerate the convergence process, and effectively improve the stability and generalization ability of the model; b. The self-built feature extraction module TE is divided into two sub-modules: local feature enhancement module RFE and global context modeling module TR; the local feature enhancement module RFE extracts the texture edge information in the occluded area through a depth separable convolution, improves the recovery ability of details, and uses a GELU activation function for feature transformation, and finally combines a residual connection to ensure the integrity of the features; the calculation process can be represented as:

[0009] wherein the given input feature map , is a set of real numbers, B is the batch size, C is the number of channels, H is the height of the feature map, W is the width of the feature map, and DWConv represents a depth separable convolution operation, is a 1x1 convolution, and GELU is an activation function, is a local encoding intermediate feature; The TR module obtains global information of the ore through a global self-attention mechanism, thereby improving the feature extraction ability for complex morphologies; the calculation process can be represented as:

[0010] wherein the given input feature map , is in the form of a sequence of feature images, , , is a learnable linear transformation matrix, Q, K, and V are query matrix, key matrix, and value matrix obtained by learnable linear transformation mapping, is a scaling factor, is a multi-head attention mechanism, is a residual connection and LayerNorm normalization added in the self-attention mechanism, and MLP is a feedforward network, is a global fusion feature; Finally, the information extracted by the two is fused to realize complete reconstruction of the edge structure; the self-built feature extraction module TE can be represented as: in, is the local encoding intermediate feature, is the global fusion feature, is the final fusion result; c. The discriminator uses the PatchGAN architecture, which consists of several stacked convolutional blocks to extract discriminative features at different scales. The final layer outputs the discriminant result. Each convolutional block consists of convolution, spectral normalization (SN), and LeakyReLU activation functions, effectively ensuring stability during training. The generated result and the original edge image are first input, and then an adversarial learning mechanism is introduced to improve the authenticity and consistency of the image completion results. d. In the model loss process, a joint loss function is used for optimization, mainly including adversarial loss, feature matching loss and reconstruction loss.

[0011] Furthermore, in step (5): For each particle connected domain, the major axis and minor axis of the minimum circumscribed rectangle are calculated as L and S respectively. The geometric mean particle size D is used to represent the equivalent particle size of the particles, and then the relative error is calculated. The specific calculation formula is:

[0012] Among them, RE is the relative error, is the average particle size measured in the present invention, is the actual measured average particle size.

[0013] Beneficial effects: The present invention combines edge completion with particle size detection to achieve high-precision particle size analysis in occluded scenes on the basis of effectively restoring the ore image structure, and has good engineering practicality and generalization ability.

[0014] The present invention proposes a particle size detection method for severely occluded ore images, introduces an ore image completion network and a semantic boundary recovery strategy, and realizes the geometric reconstruction of occluded ore particles. Specifically, the present invention constructs an irregular ore particle mask dataset for occluded scenes, and designs a new feature extraction network that integrates self-built local feature enhancement and global attention mechanisms. Different from the existing technology, the EDRE network of the present invention takes ore edge completion as the core, inputs three-channel image data, and synergistically enhances the dual reconstruction of local texture and global structure through deep separable convolution and self-attention mechanism. It can effectively repair the ore particle boundaries while maintaining the original particle distribution. The final particle size measurement relative error is only 0.42%, which proves that it can improve the accuracy and robustness of ore particle size detection in occluded scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1Schematic diagram of the process of the present invention; Figure 2 Schematic diagram of ore mask data generated for the present invention; Figure 3 This is a schematic diagram of the preprocessing of the ore mask data set according to the present invention; Figure 4 A schematic diagram of the network structure of a method for measuring ore particle size using image completion technology provided by the present invention; Figure 5 This is a schematic diagram comparing the actual effects of the image completion method proposed in the present invention and other mainstream image completion methods; Figure 6 A schematic diagram of particle size detection of an ore particle size measurement method using image completion technology provided by the present invention. DETAILED DESCRIPTION

[0016] In order to better explain the present invention, the image completion ore particle size detection method for obscured irregular ore particles of the present invention will be explained in more detail below with reference to the accompanying drawings.

[0017] like Figure 1-4 As shown, the present invention discloses a method for detecting ore particle size based on the EDRE ore image completion network model. Specifically, the present invention implements ore particle size detection based on image completion through the following steps: Step 1: Collect original ore image data An industrial camera was used to capture ore samples under actual working conditions; each image was 512 × 512 pixels in size.

[0018] Step 2: Generate ore mask dataset The original image is occluded using a structured mask to construct an ore mask dataset. The mask is constructed by first binarizing the original ore image to extract the foreground area and generate a basic mask. The ore contour mask is then translated and directional overlay is added to simulate the stacked occlusion situation in real life.

[0019] Step 3: Preprocess the mask dataset The preprocessing process includes grayscale and edge processing of the ore mask dataset; inputting the original ore image, converting the original image into a mask grayscale image, and then using the mask image to perform pixel-by-pixel multiplication operations on the grayscale image; at the same time, edge detection is performed on the original image to generate an edge map, and then using the mask map to perform pixel-by-pixel multiplication operations on the edge map, and finally outputting the mask grayscale image and the mask edge image.

[0020] Step 4: Input the original ore image and the masked ore dataset to be detected into the improved ore edge completion algorithm EDRE network model to obtain the image result of the restored ore edge; Based on the original ore image and the dataset of the irregular ore mask facing the occlusion, the model is trained by means of random pairing combination; wherein the dataset is divided into a training set, a verification set and a test set according to a 7:2:1 ratio; the model training process is realized by adopting a PyTorch framework; the model construction process is that the generator encodes the input image through a convolution block, then extracts deep image features by means of a TE module composed of an RFE and a TR, and finally recovers the missing edge information through a decoding part; and then the generation result and the original image are input into the discriminator for discriminating the error of the generation result.

[0021] Step five: ore particle size calculation According to the ore image recovered by using the image completion algorithm of the present application, the geometric fitting method is used to obtain the size information of a single particle, and then the long axis L and the short axis S of the minimum circumscribed rectangle of each particle are calculated, and finally the average particle size D of the ore particle is obtained.

[0022] Step six: model verification As Figure 5 shown, it is a comparison chart of the actual effect of the method proposed in the present application and other three kinds of currently mainstream image completion methods. In order to make the comparison representative, the present application selects three currently mainstream image completion models, which are EdgeConnect, STNet and CTSTNet respectively. The comparison process is to extract the edge completion module of the three models respectively, and apply them individually to the ore image completion task of the present application for comparison. From left to right in the figure, they are: the original ore image, the edge information missing image input into the completion network, the repaired ore image result generated by using the present application, the repaired ore image result generated by using EdgeConnect, the repaired ore image result generated by using STNet and the repaired ore image result generated by using CTSTNet; it can be seen directly from the figure that in the input image, there is obvious missing of the ore edge, and the contour information of some areas is missing; the EdgeConnect and CTSTNet methods are prone to problems such as blur, discontinuity and inconsistent edge thickness when facing large-area missing of the edge and complex areas difficult to accurately fit the real edge; STNet is relatively complete, but it is prone to obvious deviation in sharp turning areas; and by means of the method proposed in the present application, the missing edge is reconstructed, and finally a continuous, smooth and highly fitted edge curve of the real contour is generated. The area indicated by the box in the figure further highlights the advantages of the present method.

[0023] Step seven: model result verification To verify that the present invention can improve the accuracy of particle size detection, the experiment selected clean coal and black gangue with a particle size of 13mm-25mm. The particle size detection method based on the image completion EDRE model was used to compare the particle size of the original ore image and the particle size of the ore image with lost edge information. The results are shown in Table 1: Table 1 Comparison of average particle size

[0024] According to Table 1 and particle size detection Figure 6 As shown in the figure, the loss of edge information will significantly reduce the accuracy of particle size detection. However, using the method of the present invention, the measurement error is only 0.42%, which is close to the original result. It can be concluded that the method proposed in the present invention has good accuracy and robustness.

[0025] Currently, this method can be deployed on-site in coal preparation plants. By integrating cameras and an edge completion algorithm platform, it can realize an online particle size detection system. It is suitable for promotion to mines, coal washing plants, iron ore mines, and other ore processing links, and has broad application prospects.

[0026] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention fall within the protection scope and disclosure scope of the present invention.

Claims

1. A method for measuring ore particle size using image completion technology, characterized by: The following steps are involved: (1) Collecting raw ore image data: Using a high-performance industrial camera as the image acquisition core to capture ore samples under actual working conditions; And use LED white light as a fixed light source for collection; (2) Generate an ore mask dataset: Use a structural mask to mask the original image and construct an ore mask dataset. The mask is constructed by first binarizing the original ore image to extract the foreground area and generate a basic mask. Then, the ore outline mask is translated and a directional overlay operation is added to simulate the stacking occlusion in real situations. (3) Preprocessing of the ore mask dataset: The preprocessing of the ore mask dataset is the process of processing the mask grayscale image and the mask edge image; (4) Using image completion module processing: using a mineral edge completion algorithm EDRE network model, the mineral edge completion algorithm EDRE network model adopts the network structure of generator and discriminator, including: convolution block, feature extraction network TE module and discriminator; The generator takes the mask image, edge image, and grayscale image as input, and uses convolution blocks and a self-built feature extraction module (TE) to encode and extract features from the image. The decoder receives the high-dimensional features processed by the intermediate module, gradually maps the high-dimensional features back to the image space through convolution operations, and outputs a reconstructed ore image after edge loss. Finally, the completed ore image and the original edge image are sent to the discriminator to determine whether the output result is reasonable. (5) Ore particle size calculation: After completing the ore image, the particle size information is obtained using the geometric fitting method.

2. The method for measuring ore particle size using image completion technology according to claim 1, characterized in that: The industrial camera in step (1) has a high-speed shutter, a low-noise sensor and an automatic exposure control function.

3. The method for measuring ore particle size using image completion technology according to claim 2, characterized in that: In step (3): a. Mask grayscale image processing: Input the original ore image, convert the original image into a mask grayscale image, then use the mask image to perform pixel-by-pixel multiplication on the grayscale image, and finally output the grayscale image after mask processing; the specific process is expressed as follows: ; Where B is the image mask, is the mask grayscale image, is the final output mask grayscale image.

4. The method for measuring ore particle size using image completion technology according to claim 3, characterized in that: In step (4): a. The convolutional module consists of convolution, spectral normalization (SN), and ReLU activation functions. The model adopts a symmetrical structure, which can achieve good structural coordination between feature encoding and decoding, and between the backbone network and branch modules. b. The self-built feature extraction module (TE) is divided into two submodules: the local feature enhancement module (RFE) and the global context modeling module (TR). The local feature enhancement module (RFE) uses depthwise separable convolution to extract texture edge information within occluded areas, improving detail recovery. It also uses the GELU activation function for feature transformation and finally combines residual connections to ensure feature integrity. The calculation process is expressed as follows: ； Among them, given the input feature map , is a set of real numbers, B is the batch size, C is the number of channels, H is the height of the feature map, W is the width of the feature map, DWConv represents the depth-wise separable convolution operation, is a 1×1 convolution, GELU is the activation function, To encode intermediate features locally; The TR module obtains the global information of the ore through the global self-attention mechanism. The calculation process is expressed as: ; Among them, given the input feature map , is the sequence form of feature images, 、 、 is a learnable linear change matrix, Q, K, and V are The query matrix, key matrix, and value matrix are obtained by learning linear transformation mapping. is the scaling factor, It is a multi-head attention mechanism. In order to add residual connection and LayerNorm normalization to the self-attention mechanism, MLP is a feedforward network. is the global fusion feature; Finally, the information extracted by the two is integrated to achieve complete reconstruction of the edge structure; the self-built feature extraction module TE is expressed as: ； in, is the local encoding intermediate feature, is the global fusion feature, is the final fusion result; c. The discriminator uses the PatchGAN architecture, which consists of several stacked convolutional blocks to extract discriminative features at different scales. The final layer outputs the discriminant result. Each convolutional block consists of convolution, spectral normalization (SN), and LeakyReLU activation function. The generated result and the original edge image are first input, and then an adversarial learning mechanism is introduced. d. In the model loss process, a joint loss function is used for optimization, mainly including adversarial loss, feature matching loss and reconstruction loss.

5. The method for measuring ore particle size using image completion technology according to claim 4, characterized in that: In step (5): the major axis and minor axis of the minimum circumscribed rectangle of each particle connected domain are calculated as L and S respectively, the geometric mean particle size D is used to represent the equivalent particle size of the particles, and then the relative error is calculated; the specific calculation formula is: ; Among them, RE is the relative error, is the average particle size measured in the present invention, is the actual measured average particle size.

Images