Lead-zinc ore example segmentation and intelligent preselection method and device

Abstract

The application provides a lead-zinc ore instance segmentation and intelligent preselection method and device. By introducing a gray edge guided adaptive convolution structure XGS-Conv, a gray guided local-global collaborative attention mechanism XGA-Attention and a boundary perception dynamic detection structure BAD-Head, a collaborative processing framework of "boundary perception-feature enhancement-dynamic detection" is formed, and high-precision identification, instance segmentation and dynamic sorting decision of ore targets in a complex low-grade lead-zinc ore X-ray transmission XRT image are realized. In this way, while ensuring the accuracy of ore target identification and the performance of instance segmentation, the network parameter quantity and the calculation complexity are effectively reduced, and the real-time deployment capability of the model in the industrial edge device is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and device for segmenting and intelligent pre-selecting lead-zinc ore instances. Background Technology

[0002] In the development and utilization of lead-zinc ore resources, pre-selection and waste disposal can remove low-grade waste rock in advance during the coarse or medium crushing stage, reducing the proportion of ineffective material entering the subsequent grinding and flotation processes. This reduces grinding energy consumption, reagent consumption, and tailings discharge, thereby improving the comprehensive utilization efficiency of lead-zinc resources. Therefore, intelligent ore identification technology based on photoelectric imaging and artificial intelligence algorithms has become an important research direction in the current field of intelligent mineral processing.

[0003] Currently, X-ray transmission imaging (XRT) technology is widely used in the intelligent pre-selection process of low-grade ores due to its advantages such as non-contact detection, strong penetration, fast recognition speed, and ease of online industrial deployment. Especially for lead-zinc ores containing sulfide minerals such as galena and sphalerite, different minerals have different X-ray absorption capacities. XRT images can reflect the internal mineral composition and density distribution characteristics of the ore to a certain extent, thus possessing high industrial application value.

[0004] However, in the actual sorting process of complex low-grade lead-zinc ores, XRT grayscale images generally suffer from problems such as blurred target boundaries, small grayscale differences, small fine-grained mineral sizes, and overlapping occlusion of ores, making it difficult for traditional image recognition methods to reliably extract ore edge and texture features. Meanwhile, the high-speed operation of industrial conveyor belts places higher demands on the real-time performance, lightweight design, and robustness of the recognition system. Traditional deep learning models often suffer from complex network structures, large parameter scales, and insufficient industrial adaptability, making it difficult to directly meet the needs of online intelligent sorting.

[0005] Furthermore, existing ore identification target detection algorithms can only output the position of the bounding rectangle of the ore, making it difficult to accurately obtain information about the true boundary area of ​​the ore. When there are overlapping, contact, or irregular shapes among the ore targets, traditional target detection algorithms are prone to problems such as boundary overlap, area statistical errors, and difficulty in target separation, which in turn affects the accuracy of subsequent ore area analysis, particle size statistics, and dynamic sorting control.

[0006] Therefore, there is an urgent need for an intelligent pre-selection method for XRT images of complex low-grade lead-zinc ore, which can achieve accurate segmentation, dynamic identification and online sorting decision-making of ore targets while ensuring industrial real-time performance. This will improve the accuracy of ore target identification and intelligent sorting control under complex working conditions, and lay the foundation for efficient pre-selection and waste disposal and industrial application of lead-zinc ore resources. Summary of the Invention

[0007] This application proposes a method and equipment for segmenting and intelligent pre-selection of lead-zinc ore instances, which can solve one of the problems existing in the background art.

[0008] To achieve the above objectives, this application adopts the following technical solution:

[0009] Firstly, a method for segmenting and intelligently pre-selecting lead-zinc ore instances is provided, including:

[0010] Obtain X-ray transmission grayscale images of lead-zinc ore to be processed;

[0011] The X-ray transmission grayscale image of the lead-zinc ore to be processed is processed using a trained intelligent pre-selection model to obtain the pre-selection result;

[0012] The intelligent pre-selection model for lead-zinc ore adopts a deep learning framework. The model comprises a cascaded backbone network, a neck network, and a detection head network. The backbone network and the neck network incorporate a gray-level edge-guided adaptive convolutional structure XGS-Conv. Within XGS-Conv, small receptive field convolutional branches and large receptive field dilated convolutional branches are constructed, and the gray-level boundary response map dynamically guides these two types of branches. Specifically, XGS-Conv is used to extract a multi-scale gray-level boundary response weight map M from the input feature map X. e The multi-scale gray-level boundary response weight map M e Determine the dynamic selection weight W of the receptive field s The weight W is dynamically selected based on the receptive field. s Output feature F of convolution branch with small receptive field s Output features F of large receptive field dilated convolution branches l Dynamic fusion is performed to obtain the boundary enhancement feature F after processing by a gray-scale boundary-guided adaptive convolutional structure. sa And utilize the multi-scale gray-level boundary response weight map M e Enhanced features F at the boundary sa Enhancement is performed to obtain the final enhanced output feature F. enh .

[0013] In one possible design approach of the first aspect, the multi-scale gray-scale boundary response weight map M e for:

[0014] in, and These represent convolution operations at different scales. This represents the dilation convolution operation. This represents the Sigmoid activation function. This represents a multi-scale grayscale boundary response weight map.

[0015] In one possible design approach of the first aspect, the receptive field dynamically selects the weight W. s for:

[0016] in, and These represent average pooling and max pooling operations, respectively. This indicates a feature concatenation operation. This indicates an element-wise multiplication operation. This represents the dynamic selection weight of the receptive field.

[0017] In one possible design approach of the first aspect, the boundary enhancement feature F sa for: .

[0018] In one possible design approach of the first aspect, the final enhanced output feature F enh for:

[0019] in, Indicates the boundary response enhancement coefficient. This indicates the final enhanced output feature.

[0020] In one possible design of the first aspect, the method further includes: constructing a gray-level guided local-global collaborative attention mechanism XGA-Attention in the neck network, wherein the gray-level guided local-global collaborative attention mechanism XGA-Attention enhances the final output feature F output by the gray-level edge-guided adaptive convolutional structure XGS-Conv in the neck network. enh Extracting local spatial response features F loc With global channel response characteristics F glo And utilize the local spatial response feature F loc With the global channel response feature F glo and the multi-scale grayscale boundary response weight map M e Together, we construct the XRT perceptual attention weights A xrt And the attention weight A is perceived by the XRT. xrt With the boundary enhancement features Weighted enhancement is performed to obtain the feature F after joint enhancement of boundary and attention. att .

[0021] In one possible design approach of the first aspect, the boundary and attention-enhanced feature F att for:

[0022] Among them, F att A represents the output feature enhanced by a gray-level edge-guided adaptive convolutional structure and a gray-level-guided local-global collaborative attention mechanism. xrt This represents the XRT perceptual attention weights. This represents a one-dimensional convolution operation. Represents the local spatial attention weights. Represents the global channel attention weight. Represents the local attention fusion coefficient. This represents the grayscale boundary response enhancement coefficient.

[0023] In one possible design of the first aspect, the method further includes: the detection features after adaptive sampling of the boundary processed by the detection head network. Introducing boundary-aware dynamic sampling offset The boundary-aware dynamic sampling offset Detection features obtained through multi-scale dynamic fusion and boundary-aware weight map Determined, the boundary-aware weight map From the boundary uncertainty response diagram and local texture gradient response features Sure.

[0024] In one possible design approach of the first aspect, the detection features are obtained after boundary adaptive sampling. for:

[0025] in, Indicates the current sampling position. Indicates the sampling offset of regular convolution. This represents the boundary-aware dynamic sampling offset. Indicates the convolution kernel weights, This represents the detected features after adaptive sampling of the boundary. This indicates the offset prediction convolution operation. This represents element-wise multiplication. This represents the hyperbolic tangent activation function. This represents the boundary-aware dynamic sampling offset. and These are the weighting coefficients. Represents a boundary-aware weighted graph. Represents the response diagram to boundary uncertainty. Represents local texture gradient response characteristics. Represents the gradient in the horizontal direction. Represents the gradient in the vertical direction. This represents the detection features after multi-scale dynamic fusion.

[0026] The weighted fusion method can simultaneously preserve boundary uncertainty information and local texture gradient information, avoiding the instability of dynamic offset prediction caused by the dominance of a single feature.

[0027] In a second aspect, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the method as described in any possible implementation of the first aspect.

[0028] Based on the above technical solution, by introducing the gray-scale edge-guided adaptive convolutional structure XGS-Conv, the adaptive selection of convolutional features with different receptive fields and the enhancement of boundary response are realized. While ensuring the accuracy of ore target recognition and instance segmentation performance, the number of network parameters and computational complexity are effectively reduced, and the real-time deployment capability of the model in industrial edge devices is improved. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a diagram of the algorithm network structure provided in the embodiments of this application;

[0031] Figure 2 This is an overall flowchart of the intelligent sorting method for lead-zinc ore provided in the embodiments of this application;

[0032] Figure 3This is an XRT grayscale image of lead-zinc ore provided in the embodiments of this application;

[0033] Figure 4 These are the AI ​​model recognition results provided in the embodiments of this application, wherein (a) is the object detection model recognition result; and (b) is the instance segmentation model recognition result.

[0034] Figure 5 This is the video recognition result of the XRT dynamic image of lead-zinc mine provided in the embodiments of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0036] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0038] This embodiment addresses the problems in X-ray transmission (XRT) images of complex low-grade lead-zinc ore, such as weak grayscale differences between ore and wastewater, blurred target boundaries, easy overlap of fine-grained ore, and insufficient real-time sorting stability under industrial conveyor belt conditions. It proposes a lead-zinc ore instance segmentation and intelligent pre-sorting method based on a collaborative mechanism of XRT image instance segmentation and dynamic sorting decision. This method achieves high-precision instance segmentation and recognition of ore targets, regional statistical analysis, and online intelligent sorting control under complex operating conditions by constructing a grayscale edge-aware feature extraction structure, a spatial-channel collaborative enhancement structure, a multi-scale local response enhancement structure, and a dynamic sorting decision output structure.

[0039] like Figure 1As shown, this embodiment constructs a lead-zinc ore instance segmentation dataset based on XRT grayscale images and uses the existing YOLOv11 deep learning framework as the base network for training and recognition of ore targets. Based on this, the feature extraction, region response enhancement, and dynamic detection output processes in the network are collaboratively improved to construct an intelligent recognition and dynamic sorting network structure suitable for complex low-grade lead-zinc ore XRT images. The constructed intelligent recognition network mainly includes a backbone feature extraction network, a neck multi-scale feature fusion network, and a dynamic sorting output network. Specifically, the backbone feature extraction network is used to extract deep texture features and grayscale boundary features of the ore; the neck multi-scale feature fusion network is used to fuse information from different ore regions, enhance local responses, and strengthen boundary regions; and the dynamic sorting output network is used to determine the ore target category, segment boundary regions, output spatial location information, and generate dynamic sorting decisions.

[0040] Due to the common problems in XRT images of complex low-grade lead-zinc ore, such as small differences in ore grayscale, blurred boundary transitions, large variations in target scale, and easy overlap of fine-grained ore, traditional convolutional networks struggle to simultaneously meet the requirements of lightweight deployment and the ability to extract complex and detailed features. Therefore, this embodiment constructs a continuous feature enhancement and sorting decision chain consisting of a grayscale edge-guided adaptive convolutional structure, a grayscale boundary-guided local-global channel collaborative attention mechanism, and a dynamic sorting and detection structure. This achieves a collaborative processing flow of "boundary perception - attention enhancement - dynamic detection - sorting decision".

[0041] First, input the feature map. The convolution branches are then divided into small receptive field convolution and large receptive field dilation convolution branches. When the ore boundary is relatively clear, the local texture information can accurately represent the target outline. In this case, using small receptive field convolution is beneficial to improve the boundary localization accuracy. However, when the ore boundary is blurred or the target overlaps, it is difficult to accurately distinguish the target boundary by relying solely on local texture information. It is necessary to use a wider range of contextual information for auxiliary discrimination.

[0042] Based on the above characteristics, this embodiment constructs a small receptive field convolutional branch and a large receptive field dilated convolutional branch in XGS-Conv, and dynamically guides the two types of branches using grayscale boundary response maps. When the grayscale boundary response is strong, the weight of the small receptive field convolutional branch is increased; when the grayscale boundary response is weak, the weight of the large receptive field dilated convolutional branch is increased, thereby achieving adaptive selection of the receptive field range.

[0043] Compared with traditional fixed convolutional structures, this embodiment does not simply increase the convolutional scale, but dynamically adjusts the contribution ratio of different receptive field features by utilizing the ore boundary state. This achieves synergistic optimization between boundary feature extraction capability and context awareness capability, thereby improving the ability to identify blurred ore, fine-grained ore, and overlapping ore regions in complex low-grade lead-zinc ore XRT images.

[0044] This embodiment introduces a multi-scale gray-level boundary joint response mechanism, which constructs a gray-level boundary response map by jointly building a small receptive field convolutional branch and a large receptive field dilated convolutional branch. The process is represented as follows: (1)

[0045] in, and These represent convolution operations at different scales. This represents the dilation convolution operation. This represents the Sigmoid activation function. This represents a multi-scale grayscale boundary response weight map.

[0046] Furthermore, to improve the network's adaptive perception capability regarding the clarity of boundaries in complex ore regions, this embodiment combines local average response features and local maximum response features to construct a dynamic selection weight for the receptive field, the process of which is expressed as follows: (2)

[0047] in, and These represent average pooling and max pooling operations, respectively. This indicates a feature concatenation operation. This indicates an element-wise multiplication operation. The receptive field is dynamically selected by weights. Through this structure, the network can dynamically adjust the response ratios of different convolutional branches based on the clarity of the ore boundary. When the ore boundary is relatively clear, the dynamic weights tend to strengthen the convolutional branches with smaller receptive fields; when the ore boundary is blurred, the ore targets overlap, or there is partial occlusion, the dynamic weights tend to strengthen the dilated convolutional branches with larger receptive fields, thereby improving the contextual awareness of the ore region under complex working conditions.

[0048] Furthermore, the output features of the small receptive field convolutional branch and the output features of the large receptive field dilated convolutional branch are dynamically fused to obtain boundary enhancement features: (3)

[0049] in, This represents the output features of the convolution branch with a small receptive field. This represents the output features of the large receptive field dilated convolution branch. This represents the boundary enhancement feature after processing by a gray-level boundary-guided adaptive convolution structure. To further enhance the fine-grained texture representation in complex low-grade ore regions, this embodiment further utilizes gray-level boundary response weights to enhance the fused features, the process of which is expressed as follows: (4)

[0050] in, Indicates the boundary response enhancement coefficient. This indicates the final enhanced output features. Through the aforementioned continuous collaborative mechanism, this embodiment does not simply superimpose existing convolutional modules, but establishes a collaborative constraint relationship between grayscale boundary response, multi-scale local statistical features, and dynamic receptive field selection, achieving dynamic adaptive adjustment of the convolutional receptive field range. Compared to traditional fixed convolutional structures, this embodiment can more effectively improve the boundary localization and texture feature representation capabilities of fine-grained ore, blurred-boundary ore, adherent ore, and locally occluded ore regions in complex low-grade lead-zinc ore XRT images, while also meeting the lightweight and real-time requirements of industrial edge deployment.

[0051] This embodiment constructs a grayscale-guided local-global collaborative attention mechanism (XGA-Attention). This mechanism uses the boundary enhancement features F output by the XGS-Conv module. enh As input, local spatial response features and global channel response features are extracted separately, and then combined with grayscale boundary response information for collaborative modeling. The local spatial response branch focuses on ore boundary regions, grayscale transition regions, and fine-grained ore regions; the global channel response branch models the importance of different channel features. Through the collaborative fusion of local spatial response and global channel response, feature enhancement and background suppression of effective ore regions in complex low-grade lead-zinc ore XRT images are achieved. The local region response features and global channel response features can be expressed as follows: (5) (6)

[0052] in, This indicates a local pooling operation. This indicates a global pooling operation. Indicates local region features, This represents the global channel features. The local channel attention weights and global channel attention weights can be expressed as: (7) (8)

[0053] in, This represents a one-dimensional convolution operation. Indicates the local channel attention weight. This represents the global channel attention weight. Based on this, the local channel attention, global channel attention, and grayscale boundary response weights are fused to obtain the XRT perceptual attention weights: (9)

[0054] in, Represents the local attention fusion coefficient. Indicates the grayscale boundary response enhancement coefficient. This represents the XRT perceptual attention weights after fusion. From this, we obtain the features after joint enhancement of boundaries and attention: (10)

[0055] in, This represents the feature map enhanced by grayscale edge-guided adaptive convolution and local-global channel collaborative attention.

[0056] Furthermore, this embodiment constructs a boundary-aware dynamic detection head structure. This structure uses attention-enhanced multi-scale features as input and dynamically allocates response weights to different scale detection layers. This enables the network to adaptively complete ore target category determination, boundary localization, instance segmentation, and sorting signal generation based on the scale changes, boundary ambiguity, and local texture response of ore targets at different particle sizes. The attention-enhanced feature sets at different scales can be represented as: (11)

[0057] in, , , These represent the attention enhancement features of the small-scale, medium-scale, and large-scale detection layers, respectively.

[0058] The dynamic scaling response weight can be expressed as: (12)

[0059] in, Indicates the first Dynamic response weights of each scale feature layer Let i represent the scale response mapping function, where i and j take values ​​of 3, 4, and 5. and These represent the detection layer indices at different scales. Indicates the first The scale detection layer corresponds to features that are enhanced collaboratively enhanced by boundary and attention functions; Indicates the first The scale detection layer corresponds to boundary and attention-based collaborative enhancement features. The dynamically fused detection features can be represented as: (13)

[0060] in, This represents the detection features after multi-scale dynamic fusion.

[0061] Because complex low-grade lead-zinc ore XRT images commonly exhibit blurred ore boundaries, overlapping fine-grained ore, and local occlusion, when the convolutional sampling position is fixed, the convolutional kernel sampling area easily covers both the target and background areas simultaneously. This leads to weakened boundary response, adhesion of instance segmentation regions, and increased target localization error. Therefore, this embodiment first constructs boundary uncertainty response features and local texture gradient response features, and generates boundary-aware weights through a boundary state awareness mechanism to guide subsequent dynamic sampling position adjustments. The ore boundary uncertainty response can be expressed as: (14)

[0062] in, This represents the response map to boundary uncertainty. The local texture gradient response can be expressed as: (15)

[0063] in, Represents the gradient in the horizontal direction. Represents the gradient in the vertical direction. Representing local texture gradient response features. Constructing boundary-aware weights based on boundary uncertainty response and local texture gradient response: (16)

[0064] in, and These are the weighting coefficients, and their sum is 1. This represents the boundary-aware weight map. The boundary-aware weight simultaneously reflects the degree of ore boundary ambiguity and the degree of local texture change. When the ore boundary is blurred, the boundary uncertainty response dominates; when the ore boundary is relatively clear, the local texture gradient response dominates, thus achieving adaptive representation of different ore boundary states. Based on this, the boundary-aware weight is used to guide dynamic sampling offset prediction, and its dynamic offset is expressed as: (17)

[0065] in, This indicates the offset prediction convolution operation. This represents element-wise multiplication. Represents the hyperbolic tangent activation function; This represents the boundary-aware dynamic sampling offset. Compared to traditional convolutional sampling structures, the conventional sampling offset... This only represents the fixed geometric sampling position of the convolution kernel, which cannot be adaptively adjusted according to the ore boundary state. This embodiment introduces a dynamic sampling offset. This is used for boundary-aware dynamic correction of fixed sampling positions, enabling convolutional sampling points to dynamically shift towards the real ore boundary region, thereby improving the feature extraction capability of the boundary region. The boundary adaptive convolutional sampling process based on dynamic offset is represented as: (18)

[0066] in, Indicates the current sampling position. Indicates the sampling offset of regular convolution. This represents the boundary-aware dynamic sampling offset. Indicates the convolution kernel weights, This represents the detected features after adaptive sampling of the boundary. This represents the total number of sampling points in the convolution kernel.

[0067] Based on the detection features obtained through boundary adaptive sampling, the dynamic sorting and detection structure completes ore target category determination, boundary region localization, instance segmentation mask generation, and recognition confidence output. Its final detection output process is represented as follows: (19)

[0068] in, Indicates the type of ore. Represents the target bounding box. Represents an instance segmentation mask. Represents the recognition confidence level. Head(⋅) represents the detection output function of the Boundary Aware Dynamic Detection Head (BAD-Head), which is used to perform class prediction, bounding box regression, instance mask generation, and confidence calculation on the input detection features.

[0069] To ensure the timing synchronization between the identification area and the sorting execution area, this embodiment further constructs a conveyor belt delay control mechanism, the delay control process of which can be expressed as follows: (20)

[0070] in, This indicates the transport distance between the identification area and the sorting execution area. Indicates the speed of the conveyor belt. This indicates the sorting delay time.

[0071] Through the above structure, this embodiment forms a continuous collaborative technology chain consisting of grayscale boundary response guidance, adaptive receptive field selection, local-global channel collaborative attention enhancement, multi-scale dynamic detection, and sorting execution control. This method can adaptively enhance the effective region features of the ore based on the clarity of ore boundaries, differences in grayscale response, and changes in target scale in the XRT image. It also directly uses the instance segmentation results for dynamic sorting decisions, thereby improving the recognition accuracy, region segmentation ability, and online sorting stability of fine-grained ore, overlapping ore, and ore with blurred boundaries in complex low-grade lead-zinc ore XRT images.

[0072] This embodiment constructs the overall network structure of the proposed intelligent lead-zinc ore sorting algorithm based on the YOLOv11 instance segmentation framework, as follows: Figure 1 As shown, the system mainly consists of a backbone network, a neck network, and a detection head network. First, X-ray transmission (XRT) imaging is performed on lead-zinc ore on an industrial conveyor belt to acquire grayscale images of the ore and construct an ore instance segmentation dataset. Then, the backbone network performs deep feature extraction on the ore images through a multi-layer convolutional (Conv) structure. During the feature extraction stage, XGS-Conv (XRT grayscale-guided dynamically switching convolutional structure) is introduced. Through boundary response guidance and dynamic receptive field selection mechanisms, adaptive feature enhancement is achieved for ore edge regions, weak texture regions, and complex grayscale regions, thereby improving the target representation capability in complex low-grade lead-zinc ore XRT images.

[0073] At the end of the backbone network, the SPPF module, a fast spatial pyramid pooling module, extracts multi-scale spatial features of the target within different receptive fields through continuous max pooling operations, and then concatenates and fuses features of different scales to enhance the network's contextual awareness of ore targets of different sizes. The C2PSA module further enhances the global feature representation capability of the ore region, improving the ability to distinguish between ore targets and background regions under complex background conditions. Subsequently, the neck network is used to complete the information fusion between features of different scales. The Upsample module is used to upsample deep feature maps to restore spatial resolution, enabling the fusion of deep semantic information with shallow detail information; the Concat module is used to concatenate and fuse feature maps of different scales, achieving the joint expression of ore edge detail information and high-level semantic information. The neck network introduces XGA-Attention (XRT gray-scale guided collaborative attention mechanism) at key fusion nodes. Through collaborative modeling of local region response information and global channel information, it adaptively assigns weights to ore targets of different granularities, thereby enhancing the local texture representation capability of fine-grained ore, overlapping ore, and ore targets with blurred boundaries.

[0074] Finally, the detection head network incorporates a BAD-Head (Boundary Aware Dynamic Detection Head) structure to dynamically adjust the response to features at different scales, enabling ore target category determination, contour region segmentation, and spatial location information output. The control system generates corresponding sorting control signals based on the model output, driving the air-blowing device or mechanical baffles to automatically separate ore from waste rock, achieving online intelligent pre-sorting and dynamic sorting control of complex low-grade lead-zinc ore resources.

[0075] Compared with the prior art, this embodiment has the following advantages:

[0076] (1) In this embodiment, a lightweight boundary-aware feature extraction structure for XRT grayscale images is constructed. By introducing XGS-Conv (XRT grayscale guided dynamic switching convolution structure), the adaptive selection of convolution features with different receptive fields and the enhancement of boundary response are realized. While ensuring the accuracy of ore target recognition and instance segmentation performance, the number of network parameters and computational complexity are effectively reduced, and the real-time deployment capability of the model in industrial edge devices is improved.

[0077] (2) In this embodiment, by constructing a spatial-channel collaborative feature enhancement mechanism, XGA-Attention (XRT gray-level guided collaborative attention mechanism) is introduced in the process of ore feature extraction and fusion. By using the joint modeling method of local texture response and global channel information, the ability to distinguish between ore targets and waste rock areas under complex background conditions is improved, and the model's ability to express features of weak texture areas, low gray-level difference areas and fine-grained ore targets is enhanced.

[0078] (3) This embodiment improves the recognition accuracy and boundary region segmentation capability of small-scale ore, overlapping ore and partially occluded ore targets under complex conveyor belt conditions by using a local response enhancement and dynamic feature fusion mechanism; at the same time, combined with the BAD-Head (boundary-aware dynamic detection head) structure, it realizes the dynamic adjustment of the feature response of ore targets of different particle sizes, and improves the stability and robustness of ore target detection and instance segmentation results under complex working conditions.

[0079] (4) This embodiment outputs the true boundary area information of the ore based on instance segmentation. Compared with the traditional target detection method that only outputs rectangular bounding boxes, it can more accurately obtain the true contour, area information and spatial distribution characteristics of the ore, thereby improving the accuracy of ore area statistics, particle size analysis, target area calculation and dynamic sorting decision, and providing a more reliable data foundation for subsequent industrial control and equipment parameter optimization.

[0080] (5) This embodiment can realize real-time identification, regional segmentation, dynamic statistical analysis and online intelligent sorting control of ore targets under the dynamic operation of industrial conveyor belts. It can adapt to problems such as weak gray scale difference, blurred boundaries and ore overlap in complex low-grade lead-zinc ore XRT images. It has good engineering adaptability, industrial deployment capability and promotion and application value.

[0081] Specific application examples:

[0082] This application example proposes a lead-zinc ore instance segmentation and intelligent pre-selection method based on XRT image instance segmentation and dynamic sorting decision-making. Figure 2 As shown, the process mainly includes six steps: lead-zinc ore XRT grayscale image acquisition, instance segmentation dataset construction, XRT grayscale boundary enhancement and dynamic recognition network construction, ore target area recognition and boundary segmentation, recognition information analysis and dynamic sorting decision, and intelligent sorting execution control.

[0083] First, an X-ray transmission (XRT) imaging device was used to perform an online scan of the lead-zinc ore on the industrial conveyor belt, acquiring continuous XRT grayscale images, such as... Figure 3 As shown. Due to the different X-ray absorption capabilities of different minerals, different ore targets exhibit different grayscale response characteristics in grayscale images, thus reflecting the internal density distribution and mineral composition characteristics of the ore to a certain extent. To ensure the continuity of image sequences and data consistency in industrial scenarios, the acquired XRT images are stored according to a unified naming rule and divided into training, validation, and test sets.

[0084] To ensure the continuity of image sequences and data consistency in industrial scenarios, the acquired XRT images were stored according to a unified naming convention, and training, validation, and test sets were constructed. For the ore target regions in the images, the Labelme annotation tool was used to perform pixel-level instance mask annotations on the ore boundaries, generating corresponding JSON tag files, which were further converted into the Txt tag format required by the instance segmentation model, thus constructing a complex low-grade lead-zinc ore XRT instance segmentation dataset.

[0085] For the target ore region in the image, the Labelme annotation tool is used to perform pixel-level instance mask annotation on the ore boundary region, generating corresponding JSON label files, which are then further converted into the label format required by the instance segmentation model, thereby constructing a complex low-grade lead-zinc ore XRT instance segmentation dataset. Subsequently, the constructed dataset is input into the ore intelligent recognition network proposed in this application example for training and recognition. The YOLOv11-based intelligent recognition network mainly consists of three parts: a backbone feature extraction network, a neck multi-scale feature fusion network, and a dynamic detection output network. Its overall structure is as follows: Figure 1 As shown.

[0086] Figure 4 This is a schematic diagram of the recognition results of the intelligent ore recognition model in this application example, where... Figure 4 (a) shows the recognition results of a traditional object detection model. Figure 4 (b) shows the instance segmentation and recognition results proposed in this application example. Figure 4 (a) It can only output the bounding rectangle of the ore target and the corresponding confidence information, which is insufficient to accurately describe the true boundary region of the ore; while Figure 4 (b) It can not only determine the target category of ore, but also output the outline of the real pixel area of ​​ore, so as to achieve an accurate description of the boundary area and real morphological features of ore.

[0087] Compared to traditional target detection methods, the instance segmentation and recognition method proposed in this application example can more effectively distinguish between ore and waste targets, and maintains high recognition accuracy in fine-grained ore, ore with blurred boundaries, and partially occluded scenarios. Furthermore, since the instance segmentation model can obtain information about the actual pixel regions of the ore, it can be further used for ore area statistics, grain size analysis, ore region proportion calculation, and dynamic sorting decision analysis, providing a more reliable data foundation for subsequent intelligent sorting control.

[0088] Furthermore, the XGS-Conv grayscale guided dynamic convolution structure constructed in this application example can enhance the ability to extract edge texture features of complex ore regions; the XGA-Attention collaborative enhancement structure can improve the response capability of weak texture ore regions; and the BAD-Head boundary-aware dynamic detection structure can improve the boundary localization accuracy of ore targets of different particle sizes under complex working conditions, thereby achieving high-precision instance segmentation and recognition in XRT images of complex low-grade lead-zinc ore.

[0089] Figure 5 This example demonstrates the XRT dynamic image recognition results for lead-zinc ore based on video stream detection. The system uses continuous video stream input to track, identify regions, and classify moving ore targets in real time, outputting ore type, target location, and corresponding confidence level information in real time.

[0090] from Figure 5 It can be seen that under the continuous operation of industrial conveyor belts, the ore intelligent identification method proposed in this application example can still achieve relatively stable boundary positioning and target classification for ore targets of different particle sizes, indicating that this application example has good dynamic scene adaptability and industrial real-time deployment capability.

[0091] Furthermore, the system can perform real-time statistical analysis of the quantity of ore, target area, proportion of ore types, and ore flow status on the conveyor belt per unit time. Among these, Figure 5The statistical results shown in the lower left corner can output the number of ores and wastes in the current video frame in real time, which can be further used for jet timing control, sorting parameter optimization and industrial sorting status analysis.

[0092] The control system generates corresponding control signals based on the identification results, and combines the spatial location of the ore, the speed of the conveyor belt, and the area of ​​the target area to drive the air blowing device, solenoid valve, or mechanical baffle to complete the automatic separation of the target ore and waste rock, thereby realizing online intelligent pre-selection and waste disposal of complex low-grade lead-zinc ore resources.

[0093] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0094] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0095] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0096] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0097] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0098] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0099] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0100] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for segmenting and intelligent pre-selecting lead-zinc ore instances, characterized in that, include: Obtain X-ray transmission grayscale images of lead-zinc ore to be processed; The X-ray transmission grayscale image of the lead-zinc ore to be processed is processed using a trained intelligent pre-selection model to obtain the pre-selection result; The intelligent pre-selection model for lead-zinc ore adopts a deep learning framework. The model comprises a cascaded backbone network, a neck network, and a detection head network. The backbone network and the neck network incorporate a gray-level edge-guided adaptive convolutional structure XGS-Conv. Within XGS-Conv, small receptive field convolutional branches and large receptive field dilated convolutional branches are constructed, and the gray-level boundary response map dynamically guides these two types of branches. Specifically, XGS-Conv is used to extract a multi-scale gray-level boundary response weight map M from the input feature map X. e The multi-scale gray-level boundary response weight map M e Determine the dynamic selection weight W of the receptive field s The weight W is dynamically selected based on the receptive field. s Output feature F of convolution branch with small receptive field s Output features F of large receptive field dilated convolution branches l Dynamic fusion is performed to obtain the boundary enhancement feature F after processing by a gray-scale boundary-guided adaptive convolutional structure. sa And utilize the multi-scale gray-level boundary response weight map M e Enhanced features F at the boundary sa Enhancement is performed to obtain the final enhanced output feature F. enh ; The method further includes: constructing a gray-level guided local-global collaborative attention mechanism XGA-Attention in the neck network, wherein the gray-level guided local-global collaborative attention mechanism XGA-Attention enhances the final output feature F output by the gray-level edge guided adaptive convolutional structure XGS-Conv in the neck network. enh Extracting local spatial response features F loc With global channel response characteristics F glo And utilize the local spatial response feature F loc With the global channel response feature F glo and the multi-scale grayscale boundary response weight map M e Together, we construct the XRT perceptual attention weights A xrt And the attention weight A is perceived by the XRT. xrt With the boundary enhancement features Weighted enhancement is performed to obtain the feature F after joint enhancement of boundary and attention. att ; The method further includes: the detection features after adaptive sampling of the boundary processed by the detection head network. Introducing boundary-aware dynamic sampling offset The boundary-aware dynamic sampling offset Detection features obtained through multi-scale dynamic fusion and boundary-aware weight map Determined, the boundary-aware weight map From the boundary uncertainty response diagram and local texture gradient response features It is confirmed that the detection features after multi-scale dynamic fusion are... Feature F enhanced by the combined effect of boundary and attention att Sure; The detection head network introduces a boundary-aware dynamic detection head (BAD-Head) structure, which dynamically adjusts the response to features at different scales to achieve ore target category determination, contour region segmentation, and spatial location information output.

2. The method as described in claim 1, characterized in that, The multi-scale gray-level boundary response weight map M e for: in, and These represent convolution operations at different scales. This represents the dilation convolution operation. This represents the Sigmoid activation function. This represents a multi-scale grayscale boundary response weight map.

3. The method as described in claim 2, characterized in that, The receptive field dynamic selection weight W s for: in, and These represent average pooling and max pooling operations, respectively. This indicates a feature concatenation operation. This indicates an element-wise multiplication operation. This represents the dynamic selection weight of the receptive field.

4. The method as described in claim 3, characterized in that, The boundary enhancement feature F sa for: .

5. The method as described in claim 4, characterized in that, The final enhanced output feature F enh for: in, Indicates the boundary response enhancement coefficient. This indicates the final enhanced output feature.

6. The method as described in claim 5, characterized in that, The feature F after the boundary and attention are jointly enhanced att for: Among them, F att A represents the output feature enhanced by a gray-level edge-guided adaptive convolutional structure and a gray-level-guided local-global collaborative attention mechanism. xrt This represents the XRT perceptual attention weights. This represents a one-dimensional convolution operation. Represents the local spatial attention weights. Represents the global channel attention weight. Represents the local attention fusion coefficient. This represents the grayscale boundary response enhancement coefficient.

7. The method as described in claim 1, characterized in that, Detection features after boundary adaptive sampling for: in, Indicates the current sampling position. Indicates the sampling offset of regular convolution. This represents the boundary-aware dynamic sampling offset. Indicates the convolution kernel weights, This represents the detected features after adaptive sampling of the boundary. This indicates the offset prediction convolution operation. This represents element-wise multiplication. This represents the hyperbolic tangent activation function. This represents the boundary-aware dynamic sampling offset. and These are the weighting coefficients. Represents a boundary-aware weighted graph. Represents the response diagram to boundary uncertainty. Represents local texture gradient response characteristics. Represents the gradient in the horizontal direction. Represents the gradient in the vertical direction. This represents the detection features after multi-scale dynamic fusion.

8. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

Images