A method and system for analyzing a mineral band of a mineral separation table based on a Mask-RCNN algorithm

Abstract

The application discloses a beneficiation table ore belt analysis method and system based on a Mask-RCNN algorithm, and comprises the following steps: 1) real-time acquisition of a table image: installing and consolidating a camera, aligning a side of the table, and collecting data through the monitoring camera; 2) image processing: compressing and adjusting saturation of each collected frame of image; wherein, for data used for training and verifying a model, manual semantic annotation is carried out by using a VGG Image Annotator tool; compared with a traditional manual operation method, the application can realize rapid and accurate feedback of position information of a concentrate belt and a tailing belt boundary line, can realize real-time calculation of an actual distance between a mineral groove and the concentrate belt boundary line, can realize automatic adjustment of a position of a mineral receiving plate through the distance, can improve a qualified rate of a concentrate grade, and can accelerate the efficiency of separating minerals.

Description

Technical Field

[0001] This invention relates to the fields of industrial contour detection, semantic segmentation, and deep learning, and mainly to a method and system for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm. Background Technology

[0002] Mineral resources are an essential foundation for my country's economic and military development, and the development of related industries provides strong support for China in the new era. Yunnan Province, as a major mineral resource province, is also a bridgehead for mineral technology development. However, Yunnan's industrial and mining enterprises are currently facing problems such as low prices, environmental policy restrictions, and significant resource waste, making technological transformation of these enterprises urgent. Mineral processing technology directly reflects the utilization rate of minerals and the degree of resource waste, determining the production quality of industrial and mining enterprises, while the performance of mineral processing equipment directly affects the efficiency of mineral processing. Shaking tables, as an important gravity separation device in mineral processing equipment, enable complex minerals to loosen and stratify under the thrust of water flow and the asymmetric reciprocating motion of the bed surface, ultimately presenting multiple fan-shaped mineral bands such as concentrate bands, middlings bands, and tailings bands on the bed surface, achieving the separation of minerals of different grades. Current shaking table mineral processing technology is very mature, but the degree of automation is still not high. Due to the influence of many factors such as ore quantity, feed particle size, and feed concentration, the distribution of fan-shaped mineral bands on the shaking table changes in real time. Operators must visually identify the characteristics of the ore zone and manually adjust the position of the receiving plate to minimize mineral waste and achieve a qualified concentrate grade. However, this traditional manual operation method requires frequent adjustments to the receiving plate position, resulting in high labor intensity and increasing labor costs for enterprises year by year. Furthermore, beneficiation indicators are subject to controversy due to numerous subjective human factors.

[0003] Currently, intelligentization is a major trend in the development of traditional industries. Enterprises can achieve high-quality and high-efficiency development by relying on technological innovation and intelligent upgrading. To improve the mineral processing performance of shaking tables and enhance the production quality of mining enterprises, this invention proposes combining deep learning algorithms with shaking table technology. By capturing real-time dynamic information of the ore deposit through a camera, and using computer algorithms to analyze the distribution of the fan-shaped ore bands on the current ore deposit, the invention can accurately identify the boundary between the tailings zone and the concentrate zone, and provide feedback on the location of the boundary at the edge of the ore deposit, enabling timely adjustment of the receiving plate. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method and system for analyzing mineral processing shaking table ore zones based on the Mask-RCNN algorithm. The model algorithm is simple and easy to be ported to the system for application. The Mask-RCNN algorithm is used to identify the boundary line of the ore deposit ore zone.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows:

[0006] A method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm includes the following steps:

[0007] S1: Real-time acquisition of shaking table images: Install and secure the camera, aim it at the side of the shaking table, and collect data through the monitoring camera;

[0008] S2: Image Processing: Compression and saturation adjustment are performed on each frame of the acquired image; among which, for the data used to train and validate the model, manual semantic annotation is performed using the VGG Image Annotator (VIA) tool;

[0009] S3: Construct a semantic segmentation model based on the Mask-RCNN algorithm to obtain information on the intermediate region between the tailings zone and the concentrate zone of the shaking table, as well as the location of the ore bin: Train the Mask-RCNN model using the training set, adjust the learning rate and relevant hyperparameters, and obtain the optimal parameter file for the model; Deploy the Mask-RCNN model using the parameter file obtained from training to achieve semantic segmentation of real-time frame data, identify and mask the intermediate region between the concentrate zone and the tailings zone, as well as the location of the ore bin;

[0010] S4: Obtain edge information and eliminate background interference to detection: Use Sobel edge detection and image binarization to highlight the location of the target boundary line and remove the interference of the image background to its boundary localization;

[0011] S5: Crop the image and detect the boundary line at the edge of the shaking table: Given the fixed position of the edge of the shaking table in the image, set four points to construct a rectangular box that contains complete information about the edge of the shaking table, and then crop the image after S4; through vertical projection, detect five consecutive non-zero points starting from position 0 and the maximum position respectively and record their coordinates to obtain the y-direction position information of the intersection of the two boundary lines and the edge of the shaking table; according to the pre-set multi-scale distance conversion formula, the actual position information of the boundary line on the shaking table can be calculated based on the y-direction information of the intersection point and the ore slot;

[0012] S6: Record the location of the ore bin, the intersection of the two boundary lines at the edge of the shaking table, and the distance of the ore bin from the concentrate zone in the database;

[0013] S7: Configure a Python 3.7 environment on the server, a porting environment, and use a real-time camera to capture images for debugging.

[0014] Furthermore, in step S1, a shaking table working system is set up, and the camera position is fixed so that it is aligned with the side of the shaking table to capture the complete side edge information of the shaking table; for real-time detection data, the current frame is read every 3 to 4 seconds, the camera is released after processing the frame, and a new frame is read again; for known data used for training and verification, by setting, real-time frames of the ore deposit are collected every 15 minutes, data for one week is collected and stored as jpg files, and 500 images are randomly selected for use as training and verification sets.

[0015] Furthermore, step S2 includes the following steps:

[0016] S2-1: Perform a Gaussian transform on each frame of the image captured by the camera, and then downsample by discarding even-numbered rows and columns. The target image size is the same as the original image. To achieve high-resolution image compression and reduce subsequent computational complexity;

[0017] S2-2: Adjust the saturation of the compressed image; First, convert the RGB space image to HSL space to control the upper and lower limits of saturation s; Set the saturation increment i, ranging from -1 to 1, and calculate the adjustment coefficient applicable to which RGB key for saturation adjustment; Where, when the increment i≥0, the saturation is adjusted using an exponential growth method, and when the increment i<0, the saturation is adjusted using a linear reduction method; Since i+s may exceed the upper and lower limits of saturation, when i+s≥1, the complement of s is used as the adjustment basis, and when i+s<1, the reciprocal of the complement of i is used as the adjustment basis;

[0018] The adjusted RGB values ​​are validated, with values ​​less than 0 being assigned 0 and values ​​greater than 1 being assigned 1. Finally, the adjusted RGB image is output for the next step.

[0019] S2-3: For the data used to train and validate the model, after the above operations, manual semantic annotation is required using the VGG ImageAnnotator tool; Specifically, polygon annotation is performed on the image operated on in S2-2 to obtain the middle area of ​​the boundary between concentrate and tailings and the position of the head of the ore bin, with the annotation labels being mime and bunker, respectively.

[0020] Furthermore, step 3 includes the following steps:

[0021] S3-1: For 500 labeled images in a known dataset, randomly divide them into a training set and a validation set in an 8:2 ratio;

[0022] S3-2: Train the semantic segmentation model based on the Mask-RCNN algorithm using the training set, and then evaluate the model performance using the validation set; the Mask-RCNN algorithm adds a fully connected segmentation network to the Faster-RCNN framework to achieve the three tasks of classification, regression and segmentation;

[0023] First, the training set data processed by S2 is input into a pre-trained neural network (CNN) to extract features. The CNN can be divided into two parts: the backbone feature extraction network ResNet and the feature pyramid FPN. The ResNet-FPN module uses ResNet101 as the backbone feature extraction network and performs a simple feature extraction process from bottom to top. The FPN performs nearest neighbor upsampling from top to bottom and fuses the upsampling results with the feature maps generated from bottom to top through lateral connections. Specifically, 1×1 convolution operations are performed on the output feature layers Conv2, Conv3, Conv4, and Conv5 of the ResNet network, and the output channels are all set to 256 channels. Then, they are summed with the upsampled feature maps. After fusion, 3×3 convolution is used to process the features to eliminate the aliasing effect of upsampling. The feature pyramid [P2, P3, P4, P5] is obtained from the above operations.

[0024] Next, the feature pyramid is fed into the Region Proposal Network (RPN) to generate multiple region candidate boxes. The RPN is a lightweight neural network that scans the image through a sliding window to find regions where the target exists. The RPN generates two outputs for the scanned region anchors: the anchor category (foreground or background) and the bounding box fine-tuning. If multiple anchors overlap, the anchor with the highest foreground score is retained, and the rest are discarded to construct region candidate boxes.

[0025] Compared to the Faster-RCNN algorithm, Mask-RCNN proposes to replace ROI pooling with ROI Align, which solves the region mismatch problem caused by the two quantization operations in ROI pooling. ROI Align cancels the quantization operation and maintains the image values ​​on the floating-point boundaries when traversing each candidate region. The candidate region is divided into k×k units, and four fixed coordinate positions are calculated in each unit. The values ​​of the four positions are calculated using bilinear interpolation, and then max pooling is used.

[0026] The output of the ROI Align module is input into the classification / regression module and the semantic segmentation module. The classification and regression modules are processed by the softmax function and the regressor respectively after passing through the fully connected layer. The semantic segmentation module is a convolutional network that takes the positive region selected by the ROI classifier as input and generates their 28×28 low-resolution masks.

[0027] The Mask-RCNN model is trained using a labeled training set to obtain the optimal parameter file, which is stored in .log file format. Then, the Mask-RCNN model is deployed using this parameter file. Finally, the classification / regression module obtains the location information of the ore bins, and the semantic segmentation module realizes the semantic segmentation of real-time frame data, and masks the intermediate area between the concentrate zone and the tailings.

[0028] Furthermore, step 4 includes the following steps:

[0029] S4-1: Use Sobel edge detection to obtain the edge information of the Mask-RCNN model mask; the Mask-RCNN model processes and analyzes the real-time shaking table frame to obtain a mask representing the middle region of the concentrate zone and tailings zone; the Sobel operator performs edge detection on this mask to obtain the boundary lines of the concentrate zone and tailings zone on the shaking table; the Sobel operator includes two sets of 3×3 matrices, one horizontal and one vertical; performing planar convolution with the image yields approximate values ​​of the horizontal and vertical brightness differences; as shown in the following formula, A is the original image, G... X and G Y These are the grayscale values ​​of the image after horizontal and vertical edge detection;

[0030] G X = [[-1,0,+1],[-2,0,+2],[-1,0,+1]]×A;

[0031] G Y = [[+1,+2,+1],[0,0,0],[-1,-2,-1]]×A;

[0032] Each pixel of the image is edge-detected by combining horizontal and vertical dimensions using the following formula;

[0033]

[0034] S4-2: Perform grayscale transformation on the image after Sobel edge detection. Based on multiple debugging sessions, set the threshold X = 240. For pixels with grayscale values ​​higher than 240, reset the grayscale value to 255. For pixels with grayscale values ​​less than 240, reset the grayscale value to 0 to achieve binarization and eliminate the interference of the background on the recognition of the target boundary line.

[0035] Furthermore, step 5 includes the following steps:

[0036] S5-1: Since the positions of the camera and the shaking table are fixed and only change slightly over time, we roughly pinpoint the edge of the ore deposit based on the training set images during model training. We take four positions on the image, which are required to contain complete shaking table edge information, and then perform cropping on the image after S4. By pinpointing the cropping, we reduce the recognition range and reduce computational complexity.

[0037] S5-2: Since the image after cropping has a pixel value of 255 except for the two boundary lines, all other pixels are 0. Therefore, when the image is vertically mapped, the presence of white dots indicates the presence of boundary lines. This invention performs vertical projection on the cropped image and detects five consecutive non-zero points starting from position 0 and the maximum position, recording their coordinate positions as the y-direction position information of the intersection of the two boundary lines and the edge of the rocker.

[0038] S5-3: The staff sets 5 points at equal intervals along the edge of the shaking table and determines the actual distance of the 5 points from the tail of the shaking table; in the photo taken by the camera, the y-axis position information of the five points is obtained, and a multi-scale distance conversion formula for the segmented shaking table is constructed based on the y-axis position information and the actual distance; by inputting the y-axis position of the above intersection point and the ore trough, the actual distance to the tail of the shaking table with a deviation within 10 cm can be output.

[0039] Furthermore, in step 6, the time information of the current frame is obtained, and the actual distance difference between the position of the ore bin and the boundary line of the concentrate zone is calculated by applying the multi-scale distance conversion formula of the shaking table. The calculated data is then recorded in a database named mime_sql.

[0040] Finally, in step 7, the Python 3.7 environment and corresponding installation packages are configured on the actual application server, the model is ported, and frame data is collected through a real-time camera for debugging and improvement.

[0041] A mineral processing shaking table ore zone analysis system based on Mask-RCNN algorithm includes a shaking table image acquisition module, an image processing module, a Mask-RCNN semantic segmentation module, an edge detection module, a target localization module, a data storage module, and a system debugging module;

[0042] The shaking table image acquisition module, by fixing the positions of the shaking table and the camera, is able to capture complete side information of the shaking table and perform shaking table image acquisition according to different needs;

[0043] The image processing module compresses and adjusts the saturation of each captured image frame; among which, for the data used to train and validate the model, the VGG Image Annotator tool is used for manual semantic annotation.

[0044] The Mask-RCNN semantic segmentation module consists of two parts: a training phase and a practical application phase. In the training phase, the Mask-RCNN model is trained using the training set to obtain the final parameter file, and the model performance is verified using the validation set. In the practical application phase, real-time frame data is input to detect the intermediate region between the concentrate zone and the tailings zone, as well as the location information of the ore bin.

[0045] The edge detection module uses Sobel edge detection and image binarization to highlight the location of the target boundary line and remove the interference of the image background on its boundary localization.

[0046] The target positioning module crops the image to reduce the detection range; through vertical projection, it detects five consecutive non-zero points starting from position 0 and the maximum position respectively and records their coordinates to obtain the y-direction position information of the intersection of the two boundary lines and the edge of the shaking table; according to the pre-set multi-scale distance conversion formula, the actual position information of the boundary line on the shaking table can be calculated based on the y-direction information of the intersection and the ore slot.

[0047] The data storage module records time information, ore bin location, the intersection of the two boundary lines at the edge of the shaking table, and the distance between the ore bin and the concentrate zone in a database named mime_sql;

[0048] The system debugging module configures a Python 3.7 environment and corresponding installation packages on the actual application server, ports the model, and debugs and improves it by collecting frame data through a real-time camera.

[0049] The beneficial effects of this invention are as follows: Compared with the traditional manual operation method, which requires frequent adjustment of the receiving plate, involves a large amount of manual labor, and is inefficient, this invention achieves rapid and accurate feedback of the position information of the boundary line between the concentrate zone and the tailings zone, and can calculate the actual distance between the ore bin and the boundary line of the concentrate zone in real time. Based on this distance, the receiving plate can be automatically adjusted to improve the qualified rate of concentrate grade and accelerate the efficiency of mineral separation. Attached Figure Description

[0050] Figure 1 This is a system flowchart of a mineral processing shaking table ore zone analysis method based on the Mask-RCNN algorithm;

[0051] Figure 2 This is a schematic diagram of the Mask R-CNN model;

[0052] Figure 3 This is a schematic diagram of the original shaking table image during the analysis of the ore zone on the mineral processing shaking table.

[0053] Figure 4 This is a schematic diagram of semantic segmentation images after Mask R-CNN detection during the analysis of ore zones on a mineral processing shaking table.

[0054] Figure 5 This is a schematic diagram of the image after edge detection during the analysis of the ore zone on a mineral processing shaking table.

[0055] Figure 6 This is a schematic diagram of the final result after edge positioning during the analysis of the ore zone on the mineral processing shaking table. Detailed Implementation

[0056] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0057] Reference Figure 1 and Figure 3-6 A method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm includes the following steps:

[0058] S1: Real-time acquisition of shaking table images: Install and secure the camera, aim it at the side of the shaking table, and collect data through the monitoring camera;

[0059] S2: Image Processing: Compression and saturation adjustment are performed on each captured image frame; among which, for the data used to train and validate the model, manual semantic annotation is performed using the VGG Image Annotator tool;

[0060] S3: Construct a semantic segmentation model based on the Mask-RCNN algorithm to obtain information on the intermediate region between the tailings zone and the concentrate zone of the shaking table, as well as the location of the ore bin: Train the Mask-RCNN model using the training set, adjust the learning rate and relevant hyperparameters, and obtain the optimal parameter file for the model; Deploy the Mask-RCNN model using the parameter file obtained from training to achieve semantic segmentation of real-time frame data, identify and mask the intermediate region between the concentrate zone and the tailings zone, as well as the location of the ore bin;

[0061] S4: Obtain edge information and eliminate background interference to detection: Use Sobel edge detection and image binarization to highlight the location of the target boundary line and remove the interference of the image background to its boundary localization;

[0062] S5: Crop the image and detect the boundary line at the edge of the shaking table: Given the fixed position of the edge of the shaking table in the image, set four points to construct a rectangular box that contains complete information about the edge of the shaking table, and then crop the image after S4; through vertical projection, detect five consecutive non-zero points starting from position 0 and the maximum position respectively and record their coordinates to obtain the y-direction position information of the intersection of the two boundary lines and the edge of the shaking table; according to the pre-set multi-scale distance conversion formula, the actual position information of the boundary line on the shaking table can be calculated based on the y-direction information of the intersection point and the ore slot;

[0063] S6: Record the location of the ore bin, the intersection of the two boundary lines at the edge of the shaking table, and the distance of the ore bin from the concentrate zone in the database;

[0064] S7: Configure a Python 3.7 environment on the server, a porting environment, and use a real-time camera to capture images for debugging.

[0065] Furthermore, in step S1, a shaking table working system is set up, and the camera position is fixed so that it is aligned with the side of the shaking table to capture the complete side edge information of the shaking table; for real-time detection data, the current frame data is read every 3 to 4 seconds, the camera is released after processing and analyzing the frame information, and a new frame data is read again; for known data used for training and verification, by setting, real-time frame data of the ore deposit is collected every 15 minutes, data for one week is collected, stored as a jpg file, and 500 images are randomly selected as the training set and verification set.

[0066] Furthermore, step S2 includes the following steps:

[0067] S2-1: Perform a Gaussian transform on each frame of the image captured by the camera, and then downsample by discarding even-numbered rows and columns. The target image size is the same as the original image. To achieve high-resolution image compression and reduce subsequent computational complexity;

[0068] S2-2: Adjust the saturation of the compressed image; First, convert the RGB space image to HSL space to control the upper and lower limits of saturation s; Set the saturation increment i, ranging from -1 to 1, and calculate the adjustment coefficient applicable to which RGB key for saturation adjustment; Where, when the increment i≥0, the saturation is adjusted using an exponential growth method, and when the increment i<0, the saturation is adjusted using a linear reduction method; Since i+s may exceed the upper and lower limits of saturation, when i+s≥1, the complement of s is used as the adjustment basis, and when i+s<1, the reciprocal of the complement of i is used as the adjustment basis;

[0069] The adjusted RGB values ​​are validated, with values ​​less than 0 being assigned 0 and values ​​greater than 1 being assigned 1. Finally, the adjusted RGB image is output for the next step.

[0070] S2-3: For the data used to train and validate the model, after the above operations, manual semantic annotation is required using the VGG ImageAnnotator tool; Specifically, polygon annotation is performed on the image operated on in S2-2 to obtain the middle area of ​​the boundary between concentrate and tailings and the position of the head of the ore bin, with the annotation labels being mime and bunker, respectively.

[0071] Furthermore, refer to Figure 2 and Figure 4 A schematic diagram of the Mask-RCNN model framework, wherein step 3 includes the following steps:

[0072] S3-1: For 500 labeled images in a known dataset, randomly divide them into a training set and a validation set in an 8:2 ratio;

[0073] S3-2: Train the semantic segmentation model based on the Mask-RCNN algorithm using the training set, and then evaluate the model performance using the validation set; the Mask-RCNN algorithm adds a fully connected segmentation network to the Faster-RCNN framework to achieve the three tasks of classification, regression and segmentation;

[0074] First, the training set data processed by S2 is input into a pre-trained neural network (CNN) to extract features. The CNN can be divided into two parts: the backbone feature extraction network ResNet and the feature pyramid FPN. The ResNet-FPN module uses ResNet101 as the backbone feature extraction network and performs a simple feature extraction process from bottom to top. The FPN performs nearest neighbor upsampling from top to bottom and fuses the upsampling results with the feature maps generated from bottom to top through lateral connections. Specifically, 1×1 convolution operations are performed on the output feature layers Conv2, Conv3, Conv4, and Conv5 of the ResNet network, and the output channels are all set to 256 channels. Then, they are summed with the upsampled feature maps. After fusion, 3×3 convolution is used to process the features to eliminate the aliasing effect of upsampling. The feature pyramid [P2, P3, P4, P5] obtained by the above operations is shown.

[0075] Next, the feature pyramid is fed into the Region Proposal Network (RPN) to generate multiple region candidate boxes. The RPN is a lightweight neural network that scans the image through a sliding window to find regions where the target exists. The RPN generates two outputs for the scanned region anchors: the anchor category (foreground or background) and the bounding box fine-tuning. If multiple anchors overlap, the anchor with the highest foreground score is retained, and the rest are discarded to construct region candidate boxes.

[0076] Compared to the Faster-RCNN algorithm, Mask-RCNN proposes to replace ROI pooling with ROI Align, which solves the region mismatch problem caused by the two quantization operations in ROI pooling. ROI Align cancels the quantization operation and maintains the image values ​​on the floating-point boundaries when traversing each candidate region. The candidate region is divided into k×k units, and four fixed coordinate positions are calculated in each unit. The values ​​of the four positions are calculated using bilinear interpolation, and then max pooling is used.

[0077] The output of the ROI Align module is input into the classification / regression module and the semantic segmentation module. The classification and regression modules are processed by the softmax function and the regressor respectively after passing through the fully connected layer. The semantic segmentation module is a convolutional network that takes the positive region selected by the ROI classifier as input and generates their 28×28 low-resolution masks.

[0078] The Mask-RCNN model is trained using a labeled training set to obtain the optimal parameter file, which is stored in .log file format. Then, the Mask-RCNN model is deployed using this parameter file. Finally, the classification / regression module obtains the location information of the ore bins, and the semantic segmentation module realizes the semantic segmentation of real-time frame data, and masks the intermediate area between the concentrate zone and the tailings.

[0079] Furthermore, refer to Figure 5 Step 4 includes the following steps:

[0080] S4-1: Use Sobel edge detection to obtain the edge information of the Mask-RCNN model mask; the Mask-RCNN model processes and analyzes the real-time shaking table frame data to obtain a mask representing the middle region of the concentrate zone and tailings zone; the Sobel operator performs edge detection on this mask to obtain the boundary lines of the concentrate zone and tailings zone on the shaking table; the Sobel operator includes two sets of 3×3 matrices, one horizontal and one vertical; convolving them with the image in a plane, we can obtain approximate values ​​of the brightness difference in the horizontal and vertical directions; as shown in the following formula, A is the original image, G X and G Y These are the grayscale values ​​of the image after horizontal and vertical edge detection;

[0081] G X = [[-1,0,+1],[-2,0,+2],[-1,0,+1]]×A;

[0082] G Y = [[+1,+2,+1],[0,0,0],[-1,-2,-1]]×A;

[0083] Each pixel of the image is edge-detected by combining horizontal and vertical dimensions using the following formula;

[0084]

[0085] S4-2: Perform grayscale transformation on the image after Sobel edge detection. Based on multiple debugging sessions, set the threshold X = 240. For pixels with grayscale values ​​higher than 240, reset the grayscale value to 255. For pixels with grayscale values ​​less than 240, reset the grayscale value to 0 to achieve binarization and eliminate the interference of the background on the recognition of the target boundary line.

[0086] Furthermore, step 5 includes the following steps:

[0087] S5-1: Since the positions of the camera and the shaking table are fixed and only change slightly over time, we roughly pinpoint the edge of the ore deposit based on the training set images during model training. We take four positions on the image, which are required to contain complete shaking table edge information, and then perform cropping on the image after S4. By pinpointing the cropping, we reduce the recognition range and reduce computational complexity.

[0088] S5-2: Since the image after cropping has a pixel value of 255 except for the two boundary lines, all other pixels are 0. Therefore, when the image is vertically mapped, the presence of white dots indicates the presence of boundary lines. This invention performs vertical projection on the cropped image and detects five consecutive non-zero points starting from position 0 and the maximum position, recording their coordinate positions as the y-direction position information of the intersection of the two boundary lines and the edge of the rocker.

[0089] S5-3: The staff sets 5 points at equal intervals along the edge of the shaking table and determines the actual distance of the 5 points from the tail of the shaking table; in the photo taken by the camera, the y-axis position information of the five points is obtained, and a multi-scale distance conversion formula for the segmented shaking table is constructed based on the y-axis position information and the actual distance; by inputting the y-axis position of the above intersection point and the ore trough, the actual distance to the tail of the shaking table with a deviation within 5 cm can be output.

[0090] Furthermore, in step 6, the location of the ore bin, the intersection of the two boundary lines at the edge of the shaking table, and the distance between the ore bin and the concentrate zone are recorded in the database;

[0091] Finally, in step 7, the Python 3.7 environment and corresponding installation packages are configured on the actual application server, the model is ported, and frame data is collected through a real-time camera for debugging and improvement.

[0092] A mineral processing shaking table ore zone analysis system based on Mask-RCNN algorithm includes a shaking table image acquisition module, an image processing module, a Mask-RCNN semantic segmentation module, an edge detection module, a target localization module, a data storage module, and a system debugging module.

[0093] The shaking table image acquisition module, by fixing the positions of the shaking table and the camera, can capture complete side information of the shaking table and perform shaking table image acquisition according to different needs. Specifically:

[0094] S1: Set up a shaking table working system, fix the camera position so that it is aligned with the side of the shaking table, and be able to capture the complete side edge information of the shaking table; for real-time detection data, read the current frame data every 3 to 4 seconds, process and analyze the frame information, release the camera, and reread the new frame data; for known data used for training and validation, by setting, collect real-time frame data of the ore deposit every 15 minutes, collect data for one week, store it as a jpg file, and randomly select 500 images to be used as the training set and validation set;

[0095] The image processing module compresses and adjusts the saturation of each acquired image frame; specifically, for data used to train and validate the model, manual semantic annotation is performed using the VGG Image Annotator tool, including:

[0096] S2-1: Perform a Gaussian transform on each frame of the image captured by the camera, and then downsample by discarding even-numbered rows and columns. The target image size is the same as the original image. To achieve high-resolution image compression and reduce subsequent computational complexity;

[0097] S2-2: Adjust the saturation of the compressed image; First, convert the RGB space image to HSL space to control the upper and lower limits of saturation s; Set the saturation increment i, ranging from -1 to 1, and calculate the adjustment coefficient applicable to which RGB key for saturation adjustment; Where, when the increment i≥0, the saturation is adjusted using an exponential growth method, and when the increment i<0, the saturation is adjusted using a linear reduction method; Since i+s may exceed the upper and lower limits of saturation, when i+s≥1, the complement of s is used as the adjustment basis, and when i+s<1, the reciprocal of the complement of i is used as the adjustment basis;

[0098] The adjusted RGB values ​​are validated, with values ​​less than 0 being assigned 0 and values ​​greater than 1 being assigned 1. Finally, the adjusted RGB image is output for the next step.

[0099] S2-3: For the data used to train and validate the model, after the above operations, manual semantic annotation is required using the VGG ImageAnnotator tool; among them, polygon annotation is performed on the image operated on in S2-2 to obtain the middle area of ​​the boundary between concentrate and tailings and the position of the head of the ore bin, and the annotation labels are mime and bunker respectively.

[0100] The Mask-RCNN semantic segmentation module comprises two parts: a training phase and a practical application phase. In the training phase, the Mask-RCNN model is trained using a training set to obtain the final parameter file, and the model performance is validated using a validation set. In the practical application phase, real-time frame data is input to detect the intermediate region between the concentrate zone and the tailings zone, as well as the location information of the ore bin. Specifically, this includes:

[0101] S3-1: For 500 labeled images in a known dataset, randomly divide them into a training set and a validation set in an 8:2 ratio;

[0102] S3-2: Train the semantic segmentation model based on the Mask-RCNN algorithm using the training set, and then evaluate the model performance using the validation set; the Mask-RCNN algorithm adds a fully connected segmentation network to the Faster-RCNN framework to achieve the three tasks of classification, regression and segmentation;

[0103] First, the training set data processed by S2 is input into a pre-trained neural network (CNN) to extract features. The CNN can be divided into two parts: the backbone feature extraction network ResNet and the feature pyramid FPN. The ResNet-FPN module uses ResNet101 as the backbone feature extraction network and performs a simple feature extraction process from bottom to top. The FPN performs nearest neighbor upsampling from top to bottom and fuses the upsampling results with the feature maps generated from bottom to top through lateral connections. Specifically, 1×1 convolution operations are performed on the output feature layers Conv2, Conv3, Conv4, and Conv5 of the ResNet network, and the output channels are all set to 256 channels. Then, they are summed with the upsampled feature maps. After fusion, 3×3 convolution is used to process the features to eliminate the aliasing effect of upsampling. The feature pyramid [P2, P3, P4, P5] obtained by the above operations is shown.

[0104] Next, the feature pyramid is fed into the Region Proposal Network (RPN) to generate multiple region candidate boxes. The RPN is a lightweight neural network that scans the image through a sliding window to find regions where the target exists. The RPN generates two outputs for the scanned region anchors: the anchor category (foreground or background) and the bounding box fine-tuning. If multiple anchors overlap, the anchor with the highest foreground score is retained, and the rest are discarded to construct region candidate boxes.

[0105] Compared to the Faster-RCNN algorithm, Mask-RCNN proposes to replace ROI pooling with ROI Align, which solves the region mismatch problem caused by the two quantization operations in ROI pooling. ROI Align cancels the quantization operation and maintains the image values ​​on the floating-point boundaries when traversing each candidate region. The candidate region is divided into k×k units, and four fixed coordinate positions are calculated in each unit. The values ​​of the four positions are calculated using bilinear interpolation, and then max pooling is used.

[0106] The output of the ROI Align module is input into the classification / regression module and the semantic segmentation module. The classification and regression modules are processed by a softmax function and a regressor, respectively, after passing through a fully connected layer. The semantic segmentation module is a convolutional network that takes the positive regions selected by the ROI classifier as input and generates their 28×28 low-resolution masks.

[0107] The Mask-RCNN model is trained using a labeled training set to obtain the optimal parameter file, which is stored in .log file format. Then, the Mask-RCNN model is deployed using this parameter file. Finally, the classification / regression module obtains the location information of the ore bins, and the semantic segmentation module realizes the semantic segmentation of real-time frame data, and masks the intermediate area between the concentrate zone and the tailings.

[0108] The edge detection module uses Sobel edge detection and image binarization to highlight the location of the target boundary line and remove interference from the image background for boundary localization; specifically, it includes:

[0109] S4-1: Using Sobel edge detection to obtain edge information from the Mask-RCNN model mask. The Mask-RCNN model processes and analyzes real-time shaking table frame data to obtain a mask representing the area between the concentrate zone and the tailings zone. The Sobel operator performs edge detection on this mask to obtain the boundary lines of the concentrate zone and the tailings zone on the shaking table. The Sobel operator consists of two sets of 3×3 matrices, one horizontal and one vertical. Convolving these matrices with the image in a plane yields approximate brightness differences in the horizontal and vertical directions. As shown in the formula below, A is the original image, G... X and G Y These are the grayscale values ​​of the image after horizontal and vertical edge detection.

[0110] G X = [[-1,0,+1],[-2,0,+2],[-1,0,+1]]×A;

[0111] G Y = [[+1,+2,+1],[0,0,0],[-1,-2,-1]]×A;

[0112] Each pixel of the image is edge-detected by combining horizontal and vertical dimensions using the following formula.

[0113]

[0114] S4-2: Perform grayscale transformation on the image after Sobel edge detection. Based on multiple debugging sessions, set the threshold X = 240. For pixels with grayscale values ​​higher than 240, reset the grayscale value to 255. For pixels with grayscale values ​​less than 240, reset the grayscale value to 0 to achieve binarization and eliminate the interference of the background on the recognition of the target boundary line.

[0115] The target localization module, as shown in S5, crops the image to reduce the detection range. Through vertical projection, it detects five consecutive non-zero points starting from position 0 and the maximum position, recording their coordinates to obtain the y-direction position information of the intersection of the two boundary lines and the edge of the shaking table. Based on a pre-set multi-scale distance conversion formula, the actual position information of the boundary line on the shaking table can be calculated using the y-direction information of the intersection point and the ore slot.

[0116] The data storage module, as shown in S6, records time information, the location of the ore bin, the intersection of the two boundary lines at the edge of the shaking table, and the distance between the ore bin and the concentrate zone in a database named mime_sql.

[0117] The system debugging module is shown in S7. A Python 3.7 environment and corresponding installation packages are configured on the actual application server, the model is ported, and frame data is collected via a real-time camera for debugging and improvement.

[0118] The above describes an embodiment of the mineral processing shaking table ore zone analysis method and system based on the Mask-RCNN algorithm of the present invention. The present invention achieves rapid and accurate feedback of the position information of the boundary line between the concentrate zone and the tailings zone, and can calculate the actual distance between the ore bin and the boundary line of the concentrate zone in real time. Through this distance, the receiving plate can be automatically adjusted to improve the qualified rate of concentrate grade and accelerate the efficiency of mineral separation.

[0119] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm, characterized in that, Includes the following steps: S1: Real-time acquisition of shaking table images: Install and secure the camera, aim it at the side of the shaking table, and collect data through the monitoring camera; S2: Image Processing: Compression and saturation adjustment are performed on each captured image frame; among which, for the data used to train and validate the model, manual semantic annotation is performed using the VGG Image Annotator tool; S3: Construct a semantic segmentation model based on the Mask-RCNN algorithm to obtain information on the intermediate region between the tailings zone and the concentrate zone of the shaking table, as well as the location of the ore bin: Train the Mask-RCNN model using the training set, adjust the learning rate and relevant hyperparameters, and obtain the optimal parameter file for the model; Deploy the Mask-RCNN model using the parameter file obtained from training to achieve semantic segmentation of real-time frame data, identify and mask the intermediate region between the concentrate zone and the tailings zone, as well as the location of the ore bin; S4: Obtain edge information and eliminate background interference to detection: Use Sobel edge detection and image binarization to highlight the location of the target boundary line and remove the interference of the image background to its contour detection; S5: Crop the image and detect the boundary line at the edge of the shaking table: Given the fixed position of the edge of the shaking table in the image, set four points to construct a rectangular box that contains complete information about the edge of the shaking table, and then crop the image after S4; through vertical projection, detect five consecutive non-zero points starting from position 0 and the maximum position respectively and record their coordinates to obtain the y-direction position information of the intersection of the two boundary lines and the edge of the shaking table; according to the pre-set multi-scale distance conversion formula, the actual position information of the boundary line on the shaking table can be calculated based on the y-direction information of the intersection point and the ore slot; S6: Record the location of the ore bin, the intersection of the two boundary lines at the edge of the shaking table, and the distance of the ore bin from the concentrate zone in the database; S7: Configure a Python 3.7 environment on the server, a porting environment, and use a real-time camera to capture images for debugging.

2. The method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, In step S1, a shaking table working system is set up, and the camera position is fixed so that it is aligned with the side of the shaking table to capture the complete side edge information of the shaking table. For real-time detection data, the current frame is read every 3 to 4 seconds, the frame is processed, the camera is released, and a new frame is read again. For known data used for training and validation, real-time frames of the ore deposit are collected every 15 minutes by setting the data, and a week's worth of data is collected and stored as a jpg file. 500 images are randomly selected from these images to be used as the training set and validation set.

3. The method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, Step S2 includes the following steps: S2-1: Perform a Gaussian transform on each frame of the image captured by the camera, and then downsample by discarding even-numbered rows and columns. The target image size is the same as the original image. To achieve high-resolution image compression and reduce subsequent computational complexity; S2-2: Adjust the saturation of the compressed image; First, convert the RGB space image to HSL space to control the upper and lower limits of saturation s; Set the saturation increment i, ranging from -1 to 1, and calculate the adjustment coefficient applicable to which RGB key for saturation adjustment; Where, when the increment i≥0, the saturation is adjusted using an exponential growth method, and when the increment i<0, the saturation is adjusted using a linear reduction method; Since i+s may exceed the upper and lower limits of saturation, when i+s≥1, the complement of s is used as the adjustment basis, and when i+s<1, the reciprocal of the complement of i is used as the adjustment basis; The adjusted RGB values ​​are validated, with values ​​less than 0 being assigned 0 and values ​​greater than 1 being assigned 1. Finally, the adjusted RGB image is output for the next step. S2-3: For the data used to train and validate the model, after the above operations, manual semantic annotation is required using the VGG ImageAnnotator tool; Specifically, polygon annotation is performed on the image operated on in S2-2 to obtain the middle area of ​​the boundary between concentrate and tailings and the position of the head of the ore bin, with the annotation labels being mime and bunker, respectively.

4. The method for analyzing the ore zone of a shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, Step S3 includes the following steps: S3-1: For 500 labeled images in a known dataset, randomly divide them into a training set and a validation set in an 8:2 ratio; S3-2: Train the semantic segmentation model based on the Mask-RCNN algorithm using the training set, and then evaluate the model performance using the validation set; the Mask-RCNN algorithm adds a fully connected segmentation network to the Faster-RCNN framework to achieve the three tasks of classification, regression and segmentation; First, the training set data processed by S2 is input into a pre-trained neural network (CNN) to extract features. The CNN can be divided into two parts: the backbone feature extraction network ResNet and the feature pyramid FPN. The ResNet-FPN module uses ResNet101 as the backbone feature extraction network and performs a simple feature extraction process from bottom to top. The FPN performs nearest neighbor upsampling from top to bottom and fuses the upsampling results with the feature maps generated from bottom to top through lateral connections. Specifically, 1×1 convolution operations are performed on the output feature layers Conv2, Conv3, Conv4, and Conv5 of the ResNet network, and the output channels are all set to 256 channels. Then, they are summed with the upsampled feature maps. After fusion, 3×3 convolution is used to process the features to eliminate the aliasing effect of upsampling. The feature pyramid [P2, P3, P4, P5] obtained by the above operations is shown. Next, the feature pyramid is fed into the Region Proposal Network (RPN) to generate multiple region candidate boxes. The RPN is a lightweight neural network that scans the image through a sliding window to find regions where the target exists. The RPN generates two outputs for the scanned region anchors: the anchor category and the bounding box fine-tuning. If multiple anchors overlap, the anchor with the highest foreground score is retained, and the rest are discarded to construct region candidate boxes. Compared to the Faster-RCNN algorithm, Mask-RCNN proposes to replace ROI pooling with ROI Align, which solves the region mismatch problem caused by the two quantization operations in ROI pooling. ROI Align cancels the quantization operation and maintains the image values ​​on the floating-point boundaries when traversing each candidate region. The candidate region is divided into k×k units, and four fixed coordinate positions are calculated in each unit. The values ​​of the four positions are calculated using bilinear interpolation, and then max pooling is used. The output of the ROI Align module is input into the classification / regression module and the semantic segmentation module. The classification and regression modules are processed by the softmax function and the regressor respectively after passing through the fully connected layer. The semantic segmentation module is a convolutional network that takes the positive region selected by the ROI classifier as input and generates their 28×28 low-resolution masks. The Mask-RCNN model is trained using a labeled training set to obtain the optimal parameter file, which is stored in .log file format. Then, the Mask-RCNN model is deployed using this parameter file. Finally, the classification / regression module obtains the location information of the ore bins, and the semantic segmentation module realizes the semantic segmentation of real-time frame data, and masks the intermediate area between the concentrate zone and the tailings.

5. The method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, Step S4 includes the following steps: S4-1: Use Sobel edge detection to obtain the edge information of the Mask-RCNN model mask; the Mask-RCNN model processes and analyzes the real-time shaking table frame to obtain a mask representing the middle region of the concentrate zone and tailings zone; the Sobel operator performs edge detection on this mask to obtain the boundary lines of the concentrate zone and tailings zone on the shaking table; the Sobel operator includes two sets of 3×3 matrices, one horizontal and one vertical; performing planar convolution with the image yields approximate values ​​of the horizontal and vertical brightness differences; as shown in the following formula, A is the original image, G... X and G Y These are the grayscale values ​​of the image after horizontal and vertical edge detection; G X =[[-1,0,+1],[-2,0,+2],[-1,0,+1]]×A; G Y =[[+1,+2,+1],[0,0,0],[-1,-2,-1]]×A; Each pixel of the image is edge-detected by combining horizontal and vertical dimensions using the following formula; S4-2: Perform grayscale transformation on the image after Sobel edge detection. Based on multiple debugging sessions, set the threshold X = 240. For pixels with grayscale values ​​higher than 240, reset the grayscale value to 255. For pixels with grayscale values ​​less than 240, reset the grayscale value to 0 to achieve binarization and eliminate the interference of the background on the recognition of the target boundary line.

6. The method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, Step S5 includes the following steps: S5-1: Since the positions of the camera and the shaking table are fixed and only change slightly over time, we roughly pinpoint the edge of the ore deposit based on the training set images during model training. We take four positions on the image, which are required to contain complete shaking table edge information, and then perform cropping on the image after S4. By pinpointing the cropping, we reduce the recognition range and reduce computational complexity. S5-2: Since the cropped image has a pixel value of 0 except for the two boundary lines which are 255, the image is vertically mapped. If there are white dots, the boundary lines must exist. The cropped image is vertically projected, and five consecutive non-zero points are detected from position 0 and the maximum position respectively. Their coordinate positions are recorded as the y-direction position information of the intersection of the two boundary lines and the edge of the rocker. S5-3: The staff sets 5 points at equal intervals along the edge of the shaking table and determines the actual distance of the 5 points from the tail of the shaking table; in the photo taken by the camera, the y-axis position information of the five points is obtained, and a multi-scale distance conversion formula for the segmented shaking table is constructed based on the y-axis position information and the actual distance; by inputting the y-axis position of the above intersection point and the ore trough, the actual distance to the tail of the shaking table with a deviation within 10 cm can be output.

7. The method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, In step S6, the time information of the current frame is obtained, and the actual distance difference between the position of the ore bin and the boundary line of the concentrate zone is calculated by applying the multi-scale distance conversion formula of the shaking table. The calculated data is then recorded in a database named mime_sql.

8. The method for analyzing the ore zone of a mineral processing shaking table based on the Mask-RCNN algorithm as described in claim 1, characterized in that, In step S7, the Python 3.7 environment and corresponding installation packages are configured on the actual application server, the model is ported, and frame data is collected through a real-time camera for debugging and improvement.

9. A mineral processing shaking table ore zone analysis system based on Mask-RCNN algorithm, comprising a shaking table image acquisition module, an image processing module, a Mask-RCNN semantic segmentation module, an edge detection module, a target localization module, a data storage module, and a system debugging module: The shaking table image acquisition module, by fixing the positions of the shaking table and the camera, is able to capture complete side information of the shaking table and perform shaking table image acquisition according to different needs; The image processing module compresses and adjusts the saturation of each acquired image frame; wherein... For the data used to train and validate the model, manual semantic annotation was performed using the VGG Image Annotator tool; The Mask-RCNN semantic segmentation module consists of two parts: a training phase and a practical application phase. During the training phase, the Mask-RCNN model is trained using the training set to obtain the final parameter file, and the model performance is verified using the validation set. In the practical application phase, real-time frame data is input to detect the location information of the intermediate area between the concentrate zone and the tailings zone, as well as the location of the ore bin. The edge detection module uses Sobel edge detection and image binarization to highlight the location of the target boundary line and remove the interference of the image background on its contour detection. The target positioning module crops the image to reduce the detection range; through vertical projection, it detects five consecutive non-zero points starting from position 0 and the maximum position respectively and records their coordinates to obtain the y-direction position information of the intersection of the two boundary lines and the edge of the shaking table; according to the pre-set multi-scale distance conversion formula, the actual position information of the boundary line on the shaking table can be calculated based on the y-direction information of the intersection and the ore slot. The data storage module records time information, ore bin location, the intersection of the two boundary lines at the edge of the shaking table, and the distance between the ore bin and the concentrate zone in a database named mime_sql; The system debugging module configures a Python 3.7 environment and corresponding installation packages on the actual application server, ports the model, and uses real-time camera frame data for debugging and improvement.

Images