An ore image denoising and grade prediction method and system

Abstract

The present invention is applicable to the technical field of automatic ore sorting, and provides an ore image denoising and grade prediction method and system, including the following steps: respectively collecting a low-speed ore transmission color image and a high-speed ore transmission color image through a dual-energy XRT machine, and each transmission color image consists of two groups of images with high energy and low energy; determining the low-speed low-energy ore transmission color image as a pure image, determining the high-speed low-energy ore transmission color image as a noise image, improving and reconstructing the training samples based on the residual network structure, re-integrating the denoised low-energy and high-energy ore transmission color images, performing calculations at the pixel level, creating a full-zero matrix M with the same resolution as the ore transmission color image, creating a three-dimensional full-zero matrix with the same size as the ore transmission color image, performing channel fusion and converting it into an ore grade map. The present invention effectively solves the problem of difficult ore image denoising, improves the recognition accuracy and real-time performance of XRT equipment, and is convenient for visual inspection of ore grades.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic ore sorting, and specifically relates to a method and system for ore image denoising and grade prediction. Background Technique

[0002] Dual-energy XRT imaging material recognition technology is a technology that realizes material recognition by using the different absorption degrees of high- and low-energy X-rays by materials. Using a dual-energy X-ray detector, two sets of images of high energy and low energy of the material can be obtained, and an R-value model is established by analyzing these two sets of images. In the actual process, noise will be introduced in almost every link of X-ray imaging, including: quantum fluctuation noise generated by unstable ray sources, electronic noise generated by image sensors, noise generated by digital systems, ray scattering noise, etc. Especially the low-energy image noise is more obvious, and ensuring the authenticity of the data characteristics of high- and low-energy images is a relatively important link in model making and recognition. At present, in the research on the separation of waste metals using dual-energy X-rays, it is mainly based on the R-value algorithm. The R-value algorithm is based on the attenuation formula of a single energy spectrum, and the R-value model is established on the theory of monochromatic light. However, the X-rays in the actual process are continuous, which will change the absorption rate of materials, resulting in the R-value model being inaccurate in many cases. This method cannot effectively separate ores and waste rocks for many types of ores, nor can it obtain the grade information of ores. Therefore, it is necessary to provide a method and system for ore image denoising and grade prediction to solve the above problems. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a method and system for ore image denoising and grade prediction to solve the problems existing in the above background technology.

[0004] The present invention is realized as follows. A method for ore image denoising and grade prediction, the method includes the following steps:

[0005] Collect low-speed ore transmission images and high-speed ore transmission images respectively through a dual-energy XRT machine. The ore transmission images include quartz transmission images and metal block transmission images, and each transmission image is composed of two sets of images of high energy and low energy;

[0006] Determine the low-speed low-energy ore transmission image as a pure image, determine the high-speed low-energy ore transmission image as a noise image, and set the pure image and the noise image as the training samples of the denoising model;

[0007] Improve and reconstruct the training samples based on the residual network structure to improve the denoising accuracy and inference speed of the denoising model;

[0008] Extract the pixels of the high-energy and low-energy images of the quartz transparent image and the metal block transparent image, use Python or Matlab to make a frequency scatter plot, and determine the coincidence of metal and quartz;

[0009] Use the extracted pixels as training samples and train a model using the LGBM algorithm under the skleran framework;

[0010] Re-integrate the denoised low-energy and high-energy ore transparent images, perform calculations at the pixel level, create a all-zero matrix M with the same resolution as the ore transparent image, perform prediction calculations for each pixel of the ore transparent image, and re-place the results into the matrix M at the same coordinates to obtain matrix M1;

[0011] Create a three-dimensional all-zero matrix with the same size as the ore transparent image. The value of the first channel is 77 - 77 * M1, the second channel is set to 230, and the third channel is set to 249. Perform channel fusion and convert it into an ore grade map.

[0012] As a further solution of the present invention: The step of improving and reconstructing the training samples based on the residual network structure specifically includes:

[0013] Flip the image pixels so that the ore spots of the stone will not be erased during the max pooling process;

[0014] Use the nearest neighbor method without changing the pixel values of the image;

[0015] Do not fix the width of the image, so that denoising can perform feature processing by single row or multiple rows and run in parallel with the subsequent ore grade calculation link;

[0016] Change the basic network to depth-wise-separable convolution, remove the fully connected layer, use FCN, and change it to an end-to-end structure;

[0017] Adjust the number of sub-structures and the number of feature maps, perform quantization processing on the model and convert it into a mobile mode.

[0018] As a further solution of the present invention: The method further includes performing data augmentation and normalization processing on the ore transparent image as the input data of the denoising model.

[0019] As a further solution of the present invention: In the ore grade map, the red dots represent pure ore spots, changing from red to yellow and then to green indicates a gradual decrease in the ore content, and the blue represents waste spots.

[0020] Another object of the present invention is to provide an ore image denoising and grade prediction system, and the system includes:

[0021] An ore penetration image acquisition module, which is used to acquire low-speed ore penetration images and high-speed ore penetration images respectively through a dual-energy XRT machine. The ore penetration images include quartz penetration images and metal block penetration images, and each penetration image consists of two groups of high-energy and low-energy images;

[0022] A training sample determination module, which is used to determine that the low-speed low-energy ore penetration image is a pure image, determine that the high-speed low-energy ore penetration image is a noise image, and set the pure image and the noise image as the training samples of the denoising model;

[0023] An improvement and reconstruction module, which improves and reconstructs the training samples based on the residual network structure to improve the denoising accuracy and inference speed of the denoising model;

[0024] A frequency scatter plot module, which is used to extract the pixels of the high-energy images and low-energy images of the quartz penetration images and metal block penetration images, make a frequency scatter plot using python or matlab, and determine the coincidence situation of metals and quartz;

[0025] An LGBM training module, which is used to use the extracted pixels as training samples and train a model using the LGBM algorithm under the skleran framework;

[0026] A recombined matrix module, which is used to re-integrate the denoised low-energy and high-energy ore penetration images, perform calculations at the pixel level, create a full-zero matrix M with the same resolution as the ore penetration image, perform prediction calculations for each pixel of the ore penetration image, and re-place the results into the matrix M at the same coordinates to obtain the matrix M1;

[0027] An ore grade map generation module, which is used to create a three-dimensional full-zero matrix with the same size as the ore penetration image. The value of the first channel is 77 - 77 * M1, the second channel is set to 230, and the third channel is set to 249, and perform channel fusion and convert it into an ore grade map.

[0028] As a further solution of the present invention: the improvement and reconstruction module includes:

[0029] An image pixel flipping unit, which is used to flip the image pixels so that the ore spots of the stones will not be erased during the maximum pooling process;

[0030] A pixel value fixing unit, which is used to use the nearest neighbor method without changing the pixel values of the image;

[0031] An acceleration processing unit, which does not fix the width of the image, enables the denoising to perform feature processing row by row or multiple rows, and runs in parallel with the subsequent ore grade calculation link;

[0032] A training saving unit is used to change the base network to depth-wise-separable convolution, remove the fully connected layer, use FCN, and change to an end-to-end structure;

[0033] A structure adjustment unit is used to adjust the number of sub-structures and the number of feature maps, perform quantization processing on the model, and convert it into a mobile mode.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] Based on the residual structure, the present invention improves and reconstructs the algorithm, which can be specifically used for XRT low-energy image denoising, while maintaining accuracy and real-time performance. The output image of the denoising model has very little interference and retains the vast majority of the original ore spots, without distorting the original data, and greatly improves the accuracy of data extraction; by extracting the pixels of the high- and low-energy images of quartz and metal, making a frequency scatter plot, and using the LGBM algorithm to analyze and make a model, the LGBM algorithm has the characteristics of fast speed, high accuracy, and small memory occupancy, and is not prone to overfitting, especially suitable for XRT devices. Re-extract the high- and low-energy image data of the ore after denoising, perform calculations at the pixel level, reconstruct the matrix M1, and then convert it into a color image through a formula to obtain an ore grade image that can be directly viewed. It solves the problem that the convolutional neural network is not suitable for XRT device image denoising, ensures the denoising accuracy and real-time performance, and at the same time improves the problem that XRT devices cannot predict the ore grade. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of a method for ore image denoising and grade prediction.

[0037] Figure 2 It is an effect diagram of denoising using a denoising model in a method for ore image denoising and grade prediction.

[0038] Figure 3 It is an ore grade map in a method for ore image denoising and grade prediction. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] The following describes the specific implementation of the present invention in detail with reference to specific embodiments.

[0041] Such as Figure 1 、 Figure 2 and Figure 3As shown in the figure, an embodiment of the present invention provides a method for ore image denoising and grade prediction. The method includes the following steps:

[0042] S100, respectively collect a low-speed ore transmission color image and a high-speed ore transmission color image through a dual-energy XRT machine. The ore transmission color image includes a quartz transmission color image and a metal block transmission color image. Each transmission color image is composed of two groups of high-energy and low-energy images;

[0043] S200, determine the low-speed low-energy ore transmission color image as a pure image, determine the high-speed low-energy ore transmission color image as a noise image, and set the pure image and the noise image as training samples of the denoising model;

[0044] S300, improve and reconstruct the training samples based on the residual network structure to improve the denoising accuracy and inference speed of the denoising model;

[0045] S400, extract the pixels of the high-energy images and low-energy images of the quartz transmission color image and the metal block transmission color image, use python or matlab to make a frequency scatter plot, and determine the coincidence of metal and quartz;

[0046] S500, use the extracted pixels as training samples and train the model using the LGBM algorithm under the skleran framework;

[0047] S600, re-integrate the denoised low-energy and high-energy ore transmission color images, perform calculations at the pixel level, create a full-zero matrix M with the same resolution as the ore transmission color image, perform prediction calculations for each pixel of the ore transmission color image, and re-place the results into the matrix M with the same x and y coordinates to obtain matrix M1;

[0048] S700, create a three-dimensional full-zero matrix with the same size as the ore transmission color image. The value of the first channel is 77 - 77 * M1, the second channel is set to 230, and the third channel is set to 249. Perform channel fusion and convert it into an ore grade map.

[0049] In the embodiment of the present invention, the LGBM algorithm under the skleran framework is used to train the model. The LGBM algorithm has the characteristics of fast speed, high accuracy, and small memory occupation, and it is not prone to overfitting. The present invention uses the residual structure as a carrier to improve and reconstruct the entire algorithm, enabling it to be specifically used for XRT low-energy image denoising while maintaining accuracy and real-time performance. By means of data augmentation and batch normalization operations, the training speed and generalization ability of the training model are improved, and various methods such as adjusting the substructure are used to achieve the optimal denoising effect and time-consuming requirements. The output image of the denoising model has very few interferences and retains the vast majority of the original ore spots, without causing distortion to the original data, greatly improving the accuracy of data extraction. Also, by denoising the image row by row, the time consumption is further reduced, and denoising and prediction can be carried out simultaneously in real time. By extracting the pixels of the high- and low-energy images of quartz and metal, making a frequency scatter plot, and using the LGBM algorithm to analyze and make a model, this algorithm has the characteristics of fast speed, high accuracy, and small memory occupation, and it is not prone to overfitting, which is especially suitable for XRT devices. The high- and low-energy image data of the ore after denoising is re-extracted, calculated at the pixel level, the matrix M1 is reconstructed, and then transformed into a color image through a formula to obtain an ore grade image that can be directly viewed. It solves the problem that the convolutional neural network is not suitable for XRT device image denoising, ensures the denoising accuracy and real-time performance, and at the same time improves the problem that the XRT device cannot predict the ore grade.

[0050] In the embodiment of the present invention, the red dots in the ore grade map represent pure ore spots. The change from red to yellow and then to green indicates a gradual decrease in the ore content, and the blue represents waste spots. Thus, the ore grade can be directly judged by the naked eye. The method also includes performing data augmentation and standardization processing on the ore transparent image as the input data of the denoising model, and through continuous parameter tuning, training to obtain the optimal denoising model.

[0051] In the embodiment of the present invention, the step of improving and reconstructing the training samples based on the residual network structure specifically includes:

[0052] S301, flipping the image pixels so that the ore spots of the stone will not be erased during the max-pooling process;

[0053] S302, using the nearest neighbor method without changing the pixel values of the image;

[0054] S303, not fixing the width of the image, enabling the denoising to perform feature processing row by row or multiple rows, and running in parallel with the subsequent ore grade calculation link;

[0055] S304, changing the basic network to depth-wise-separable convolution to reduce the inference time, removing the fully connected layer, using FCN, and changing to an end-to-end structure, which is beneficial for training and time saving;

[0056] S305 adjusts the number of substructures and the number of feature maps, which is beneficial to the inference speed. The model is quantized and converted into a mobile mode to further improve the inference speed.

[0057] It should be noted that since each pixel value represents the ore content, the pixel values of the image cannot be changed; denoising can perform feature processing on a single row or multiple rows and run in parallel with the subsequent ore grade calculation step, reasonably utilizing computer resources and having a relatively fast acceleration effect.

[0058] The embodiment of the present invention also provides an ore image denoising and grade prediction system, which includes:

[0059] An ore transparent image acquisition module for respectively acquiring a low-speed ore transparent image and a high-speed ore transparent image through a dual-energy XRT machine. The ore transparent image includes a quartz transparent image and a metal block transparent image, and each transparent image consists of two groups of images, namely high-energy and low-energy images;

[0060] A training sample determination module for determining the low-speed low-energy ore transparent image as a pure image, determining the high-speed low-energy ore transparent image as a noise image, and setting the pure image and the noise image as the training samples of the denoising model;

[0061] An improvement and reconstruction module for improving and reconstructing the training samples based on the residual network structure to improve the denoising accuracy and inference speed of the denoising model;

[0062] A frequency scatter plot module for extracting the pixels of the high-energy and low-energy images of the quartz transparent image and the metal block transparent image, using python or matlab to create a frequency scatter plot, and determining the coincidence situation of metals and quartz;

[0063] An LGBM training module for using the extracted pixels as training samples and training a model using the LGBM algorithm under the skleran framework;

[0064] A recombined matrix module for re-integrating the denoised low-energy and high-energy ore transparent images, performing calculations at the pixel level, creating a all-zero matrix M with the same resolution as the ore transparent image, performing prediction calculations for each pixel of the ore transparent image, and re-placing the results into the matrix M at the same coordinates to obtain matrix M1;

[0065] An ore grade map generation module for creating a three-dimensional all-zero matrix with the same size as the ore transparent image, setting the value of the first channel as 77 - 77*M1, setting the second channel as 230, and setting the third channel as 249, performing channel fusion and converting it into an ore grade map.

[0066] In the embodiment of the present invention, the improvement and reconstruction module includes:

[0067] An image pixel flipping unit for flipping image pixels so that the ore spots of the stone will not be erased during the max pooling process;

[0068] A pixel value fixing unit for using the nearest neighbor method without changing the pixel values of the image;

[0069] An acceleration processing unit that does not fix the width of the image, enabling denoising to perform feature processing row by row or multiple rows, and running in parallel with the subsequent ore grade calculation step;

[0070] A training saving unit for changing the base network to depth-wise-separable convolution, removing the fully connected layer, using FCN, and changing to an end-to-end structure;

[0071] A structure adjustment unit for adjusting the number of sub-structures and the number of feature maps, performing quantization processing on the model, and converting it into a mobile mode.

[0072] The above only describes the preferred embodiments of the present invention in detail and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0073] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown in sequence according to the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily need to be executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not necessarily need to be sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0074] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0075] After considering the specification and the disclosure of the embodiments, those skilled in the art will readily conceive of other embodiments of the present disclosure. The present application is intended to cover any variations, uses, or adaptations of the present disclosure, which follow the general principles of the present disclosure and include known common knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and the embodiments are only to be regarded as exemplary, and the true scope and spirit of the present disclosure are pointed out by the claims.

Claims

1. A method for ore image denoising and grade prediction, characterized in that, The method includes the following steps: Collect low-speed ore transmission images and high-speed ore transmission images respectively through a dual-energy XRT machine. The ore transmission images include quartz transmission images and metal block transmission images, and each transmission image consists of two groups of images, namely high-energy and low-energy images; Determine that the low-speed low-energy ore transmission image is a pure image, determine that the high-speed low-energy ore transmission image is a noise image, and set the pure image and the noise image as the training samples of the denoising model; Improve and reconstruct the training samples based on the residual network structure to improve the denoising accuracy and inference speed of the denoising model. The step of improving and reconstructing the training samples based on the residual network structure specifically includes: flipping the image pixels so that the ore spots of the stones will not be erased during the max-pooling process; using the nearest neighbor method without changing the pixel values of the image; Do not fix the width of the image, so that denoising can perform feature processing by single row or multiple rows and run in parallel with the subsequent ore grade calculation link; change the basic network to depth-wise-separable convolution, remove the fully connected layer, use FCN, and change it to an end-to-end structure; adjust the number of sub-structures and the number of feature maps, perform quantization processing on the model and convert it to the mobile mode; Extract the pixels of the high-energy images and low-energy images of the quartz transmission images and metal block transmission images, make a frequency scatter plot, and determine the overlap situation of metals and quartz; Use the extracted pixels as training samples and train the model using the LGBM algorithm under the skleran framework; Re-integrate the denoised low-energy and high-energy ore transmission images, perform calculations at the pixel level, create a full-zero matrix M with the same resolution as the ore transmission image, perform prediction calculations for each pixel of the ore transmission image, and re-place the results into the matrix M at the same coordinates to obtain matrix M1; Create a 3D full-zero matrix with the same size as the ore transmission image. The value of the first channel is 77 - 77 * M1, the second channel is set to 230, and the third channel is set to 249, perform channel fusion and convert it into an ore grade map.

2. The ore image denoising and grade prediction method according to claim 1, wherein, The method further includes performing data augmentation and normalization processing on the ore transmission images as the input data of the denoising model.

3. The ore image denoising and grade prediction method according to claim 1, characterized in that In the ore grade map, the red dots represent pure ore spots, changing from red to yellow and then to green indicates a gradual decrease in the ore content, and the blue represents waste spots.

4. An ore image denoising and grade prediction system, characterized in that, The system includes: An ore transmission image acquisition module for collecting low-speed ore transmission images and high-speed ore transmission images respectively through a dual-energy XRT machine. The ore transmission images include quartz transmission images and metal block transmission images, and each transmission image consists of two groups of images, namely high-energy and low-energy images; A training sample determination module for determining that the low-speed low-energy ore transmission image is a pure image, determining that the high-speed low-energy ore transmission image is a noise image, and setting the pure image and the noise image as the training samples of the denoising model; An improvement and reconstruction module for improving and reconstructing the training samples based on the residual network structure to improve the denoising accuracy and inference speed of the denoising model; Frequency scatter plot module, which is used to extract the pixels of the high-energy images and low-energy images of the quartz transparent image and the metal block transparent image, make a frequency scatter plot, and determine the coincidence situation of the metal and the quartz; LGBM training module, which is used to use the extracted pixels as training samples and train a model using the LGBM algorithm under the skleran framework; Recombinant matrix module, which is used to re-integrate the denoised low-energy and high-energy ore transparent images, perform calculations at the pixel level, create a full-zero matrix M with the same resolution as the ore transparent image, perform prediction calculations for each pixel of the ore transparent image, and re-place the results into the matrix M at the same coordinates to obtain matrix M1; Ore grade map generation module, which is used to create a three-dimensional full-zero matrix with the same size as the ore transparent image, the value of the first channel is 77 - 77 * M1, the second channel is set to 230, and the third channel is set to 249, perform channel fusion and convert it into an ore grade map. The improvement and reconstruction module includes: Image pixel flipping unit, which is used to flip the image pixels so that the ore spots of the stone will not be erased during the max pooling process; Pixel value fixing unit, which is used to use the nearest neighbor method without changing the pixel values of the image; Acceleration processing unit, which does not fix the width of the image, enables the denoising to perform feature processing by single row or multiple rows, and runs in parallel with the subsequent ore grade calculation link; Training saving unit, which is used to change the basic network to depth-wise-separable convolution, remove the fully connected layer, use FCN, and change it to an end-to-end structure; Structure adjustment unit, which is used to adjust the number of sub-structures and the number of feature maps, perform quantization processing on the model and convert it into a mobile mode.

Images