Gold ore belt ore automatic sorting system based on image recognition

Abstract

The invention discloses an automatic gold ore belt ore sorting system based on image recognition, which relates to the technical field of image recognition vision, and comprises an image acquisition module for continuously scanning gold ore on the surface of a transmission belt according to an equal displacement trigger mode under a stable light source condition based on a synchronous parameter set to obtain an original ore image flow, carrying out image preprocessing on each frame of ore image, and obtaining an ore candidate region image set and ore position information through connected domain segmentation; the recognition and sorting module is used for inputting the ore candidate area image set into an image recognition model, generating an ore category probability graph through pixel-level segmentation, mapping and dispersing the ore category probability graph into a sorting control grid according to sorting execution control parameters and ore position information, and generating an initial ore sorting scheme; the synchronous consistency, the recognition reliability and the execution stability of automatic ore sorting are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition and vision technology, and in particular to an automatic sorting system for gold mine belt ore based on image recognition. Background Technology

[0002] As mineral resource development continues to advance towards large-scale and intelligent operations, ore sorting, as a crucial link in the mineral processing flow, directly impacts mine production efficiency and resource utilization due to its level of automation and precision. In gold mining and beneficiation, belt conveyors are widely used for raw ore transport and pre-sorting operations due to their strong continuity and adaptability. Existing gold mine belt ore sorting technologies are gradually incorporating image recognition-based detection methods. These methods use industrial cameras to acquire images of the ore surface and analyze the ore's appearance characteristics using image recognition to determine ore type or grade. Compared to traditional manual sorting or sorting based on single physical characteristics, this technology offers significant advantages in information acquisition, non-contact detection, and automation, and has become one of the important development directions in the field of intelligent mineral processing.

[0003] In existing image recognition-based gold mine conveyor belt ore sorting technologies, the spatiotemporal matching accuracy between image recognition results and actual sorting execution remains a significant factor limiting system performance. Due to the continuous motion of the conveyor belt, after image acquisition in the imaging area, the ore requires a certain amount of time and displacement to reach the sorting execution area. Without precise modeling of conveyor belt operating parameters and the spatial relationship between the imaging and execution areas, discrepancies between image recognition results and actual sorting locations can easily arise. Furthermore, existing technologies often focus on improving the recognition accuracy of the image recognition model itself, while paying insufficient attention to the synchronous mapping and execution control of image recognition results in dynamic conveyor belt scenarios. This makes it difficult to fully leverage the overall effectiveness of image recognition technology in automated ore sorting under high-speed operation or densely distributed ore conditions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an automatic sorting system for gold mine belt ore based on image recognition to solve the problem of difficulty in achieving accurate synchronization and reliable execution of image recognition results.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an automatic sorting system for gold mine conveyor belt ore based on image recognition, which includes a synchronous modeling module, which collects the operating parameters of the conveyor belt and measures the imaging execution zone spacing, establishes the displacement-time mapping relationship of the conveyor belt, and generates a synchronous parameter set; The image acquisition module, based on a synchronization parameter set, continuously scans the gold ore on the surface of the conveyor belt under stable light source conditions using an equal displacement triggering method to acquire the original ore image stream. It also performs image preprocessing on each frame of the ore image and obtains the ore candidate region image set and ore location information through connected component segmentation. The identification and sorting module inputs the image set of ore candidate areas into the image recognition model, generates an ore category probability map through pixel-level segmentation, and maps and discretizes the ore category probability map into the sorting control grid according to the sorting execution control parameters and ore location information to generate an initial ore sorting scheme. The decision execution module performs uncertainty assessment and rejection re-inspection on the initial ore sorting plan, determines the ore sorting plan, and executes automatic ore sorting in the sorting execution area according to the ore sorting plan, generating an ore sorting report.

[0007] As a preferred embodiment of the image recognition-based automatic gold mine conveyor ore sorting system of the present invention, the specific steps for collecting conveyor belt operating parameters, measuring the imaging execution zone spacing, establishing a conveyor belt displacement-time mapping relationship, and generating a synchronization parameter set are as follows. Collect the operating parameters of the conveyor belt, and time-stamp the operating parameters to form traceable operating data. Based on the traceable operating data, determine the motion change relationship of the conveyor belt at each time point. The imaging area is determined by projecting the camera's field of view onto the conveyor belt. The sorting execution area is determined by the effective range of the sorting mechanism on the conveyor belt. The distance between the corresponding baselines of the imaging area and the sorting execution area in the transmission direction is measured to obtain the imaging execution area spacing.

[0008] As a preferred embodiment of the image recognition-based automatic gold mine conveyor belt ore sorting system of the present invention, the specific steps for establishing the conveyor belt displacement-time mapping relationship and generating a synchronization parameter set are as follows: The displacement of the transport belt is updated time-by-time according to the motion change relationship. The corresponding time is determined by the imaging execution zone spacing as the judgment condition, and the time mapping relationship of the transport belt displacement is obtained. The transmission belt operating parameters, imaging execution zone spacing, and transmission belt displacement time mapping relationship are uniformly encapsulated to form a synchronization parameter set.

[0009] As a preferred embodiment of the image recognition-based automatic gold ore sorting system for conveyor belts according to the present invention, the following steps are taken: based on a synchronization parameter set, under stable light source conditions, the gold ore on the surface of the conveyor belt is continuously scanned in an equal displacement triggering manner to acquire the original ore image stream. The transmission belt displacement-time mapping relationship in the synchronization parameter set is invoked to convert the continuous motion process of the transmission belt into a predictable sequence of displacement nodes and use it as the sole triggering reference. The conveyor belt operation process is tracked in real time, and an imaging trigger signal is generated when the conveyor belt displacement reaches the preset equal displacement condition. The gold ore on the surface of the conveyor belt is imaged sequentially using an imaging trigger signal while the light source is in a stable working state. The ore images are then sorted and numbered according to the displacement trigger sequence to form an original ore image stream.

[0010] As a preferred embodiment of the image recognition-based automatic gold mine belt ore sorting system of the present invention, the specific steps for image preprocessing of each frame of ore image are as follows: Establish temporal associations between the current frame and adjacent frames before and after it, according to the temporal order of the original ore image stream; Based on the temporal correlation, the current frame is subjected to brightness normalization processing, and the normalized frame is subjected to multi-scale edge enhancement and background suppression processing to obtain a standard mineral image.

[0011] As a preferred embodiment of the image recognition-based automatic gold mine belt ore sorting system of the present invention, the specific steps for obtaining the ore candidate region image set and ore location information through connected component segmentation are as follows: The standard ore image is binarized, and adjacent pixels are aggregated and labeled based on pixel connectivity to form an initial set of connected regions. The initial set of connected regions is temporally consistent with the connected regions at corresponding positions in adjacent frames. Spatially overlapping connected regions in continuously shifted frames are retained, and single-frame noise regions are removed to obtain ore candidate regions. The candidate region images are cropped according to the pixel range of each ore candidate region in the current frame, and the candidate region images are collected in the order of acquisition to form a set of ore candidate region images; Simultaneously, the position information of each ore candidate region in the original ore image stream is recorded and mapped to the transport belt displacement coordinate system through a synchronization parameter set to generate ore position information.

[0012] As a preferred embodiment of the image recognition-based automatic ore sorting system for gold mine conveyor belts described in this invention, the constant displacement condition is determined by continuously collecting and tracking the operating parameters of the conveyor belt, discretizing the continuous displacement process into displacement nodes with fixed step lengths, and setting the displacement interval between adjacent displacement nodes.

[0013] As a preferred embodiment of the image recognition-based automatic gold mine belt ore sorting system of the present invention, the steps of inputting the set of ore candidate region images into the image recognition model and generating an ore category probability map through pixel-level segmentation are as follows: The image set of ore candidate regions is reorganized according to the acquisition order and displacement trigger number, and based on the temporal consistency relationship, the ore candidate regions of the same ore are associated as candidate target temporal instances across frames. Each candidate target time-series instance is bound to a pixel position in the original ore image stream, and the pixel position is mapped to the transport band displacement coordinate system corresponding to the ore position information to obtain a set of candidate target time-series instances; The set of candidate target time series instances is input into the image recognition model. A cross-frame memory representation is established for each candidate target time series instance, and pixel-level segmentation is performed at the candidate region scale to output a ore category probability map.

[0014] As a preferred embodiment of the image recognition-based automatic gold ore sorting system for belt conveyors described in this invention, the steps for mapping and discretizing the ore category probability map onto the sorting control grid based on sorting execution control parameters and ore location information to generate an initial ore sorting scheme are as follows: For multiple candidate target time series instances that exist in parallel within the same displacement node, the set interactive recognition mechanism is invoked to interactively model the intermediate recognition representations of each candidate target time series instance, forming context consistency constraints. The pixel-level segmentation and decoding process is constrained and corrected based on context consistency constraints, and the ore category probability map is corrected for consistency to obtain a stable ore probability map. A three-level mapping chain is established based on ore location information and sorting execution control parameters; based on the three-level mapping chain, the ore category probability corresponding to the pixel in the stable ore probability map is mapped and discretized to the sorting control grid; The probability of ore category carried by each grid node in the sorting control grid is discretized at the grid level to generate an initial ore sorting scheme.

[0015] As a preferred embodiment of the image recognition-based automatic gold ore sorting system for belt conveyors described in this invention, the steps of performing uncertainty assessment and rejection / re-inspection of the initial ore sorting scheme to determine the ore sorting scheme, and then executing automatic ore sorting in the sorting execution area according to the ore sorting scheme to generate an ore sorting report are as follows. Based on the consistency between the image recognition confidence distribution and temporal sequence of each ore in the initial ore sorting scheme, the degree of uncertainty of the initial sorting results is evaluated. Based on the degree of uncertainty, the initial sorting results that do not meet the preset confidence conditions are rejected, and a re-inspection is triggered to determine the ore sorting scheme. According to the ore sorting plan, the corresponding automatic ore sorting operation is performed in the sorting execution area to obtain the automatic ore sorting results. By summarizing and recording, an ore sorting report is generated.

[0016] The beneficial effects of this invention are as follows: by constructing a transport belt displacement-time mapping and using an equal displacement triggering method for ore imaging, the ore image data based on image recognition is accurately synchronized with the actual position of the ore; furthermore, by generating an ore category probability map through pixel-level segmentation in image recognition and mapping it to the sorting control grid, a sorting scheme that can be directly executed is formed, which effectively improves the synchronization consistency, recognition reliability and execution stability of automatic ore sorting. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of an automatic ore sorting system for gold mine belt conveyors based on image recognition.

[0019] Figure 2 A flowchart for generating the basic data for synchronization parameters.

[0020] Figure 3 This is a flowchart for continuous imaging acquisition triggered by equal displacement.

[0021] Figure 4 A flowchart generated for the initial ore sorting scheme. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an automatic sorting system for gold mine belt ore based on image recognition, including the following steps: The synchronous modeling module collects the operating parameters of the transport belt and measures the distance between imaging execution zones, establishes the displacement-time mapping relationship of the transport belt, and generates a synchronous parameter set.

[0026] The operating parameters of the conveyor belt are collected and time-stamped to form traceable operating data. Based on the traceable operating data, the motion change relationship of the conveyor belt at each time point is determined.

[0027] Furthermore, during the operation of the conveyor belt, conveyor belt operation parameters, including conveyor belt speed, cumulative displacement or displacement increment, and corresponding drive state parameters, are continuously collected at a fixed sampling period. Each conveyor belt operation parameter sampling record is written with a corresponding time stamp, and traceable operation data is formed by binding the conveyor belt operation parameter sampling records with time stamps. Based on the time stamps, the traceable operation data is organized sequentially to establish a conveyor belt operation parameter time index table, and motion parameter conversion processing is performed on adjacent operation parameter sampling records according to the time index order, thereby determining the motion change relationship of the conveyor belt at each time point.

[0028] It should be noted that the fixed sampling period is set according to the conveyor belt running speed and the subsequent imaging triggering accuracy.

[0029] The imaging area is determined by projecting the camera's field of view onto the conveyor belt. The sorting execution area is determined by the effective range of the sorting mechanism on the conveyor belt. The distance between the corresponding baselines of the imaging area and the sorting execution area in the transmission direction is measured to obtain the imaging execution area spacing.

[0030] Furthermore, based on the projection range of the camera's field of view onto the conveyor belt, the starting and ending boundaries of the camera's field of view covering the surface of the conveyor belt are determined along the transmission direction, and the imaging area is determined by the projection coverage range between the starting and ending boundaries. Based on the effective range of the sorting execution mechanism's sorting action on the conveyor belt, the starting and ending boundaries of the sorting execution mechanism's action are determined along the transmission direction, and the sorting execution area is determined by the effective range between the starting and ending boundaries of the sorting execution mechanism's action. Corresponding baselines in the transmission direction are selected in the imaging area and the sorting execution area, respectively. The baseline corresponding to the imaging area is the reference boundary line of the imaging area in the transmission direction, and the baseline corresponding to the sorting execution area is the reference boundary line of the sorting execution area in the transmission direction. The imaging execution area spacing is obtained by measuring the linear distance between the baseline corresponding to the imaging area and the baseline corresponding to the sorting execution area under the same transmission direction coordinates.

[0031] It should be noted that the baseline corresponding to the imaging area is obtained by calibrating the projection range of the camera's field of view on the surface of the transmission belt, and preferably the starting boundary line or the ending boundary line of the projection range in the transmission direction is selected as the baseline corresponding to the imaging area. The baseline corresponding to the sorting execution area is obtained by structural calibration or control parameter analysis of the effective range of the sorting execution mechanism on the conveyor belt, and preferably the starting boundary line or the ending boundary line of the effective range in the transmission direction is selected as the baseline corresponding to the sorting execution area.

[0032] The displacement of the transport belt is updated hourly according to the motion change relationship, and the corresponding time is determined by the imaging execution zone spacing as the criterion, so as to obtain the time mapping relationship of the transport belt displacement.

[0033] Furthermore, based on the motion change relationship of the transport belt at each time point, the cumulative displacement of the transport belt is updated hourly according to the time mark order of the transport belt operation parameter time index table. During the hourly update process, the cumulative displacement of the transport belt is compared with the imaging execution zone spacing. When the cumulative displacement of the transport belt meets the displacement condition corresponding to the imaging execution zone spacing, the corresponding time mark or interpolation time is determined, thereby establishing the transport belt displacement time mapping relationship corresponding to the imaging execution zone spacing.

[0034] The transmission belt operating parameters, imaging execution zone spacing, and transmission belt displacement time mapping relationship are uniformly encapsulated to form a synchronization parameter set.

[0035] Furthermore, the transmission belt operating parameters, imaging execution zone spacing, and transmission belt displacement time mapping relationship are organized according to a unified data structure, and a clear correspondence is established among the three. Each set of transmission belt operating parameters is bound and encapsulated with the corresponding imaging execution zone spacing and transmission belt displacement time mapping relationship to form a synchronization parameter set.

[0036] The image acquisition module, based on a synchronization parameter set, continuously scans the gold ore on the surface of the transport belt under stable light conditions using an equal displacement triggering method to acquire the original ore image stream. It then performs image preprocessing on each frame of the ore image and obtains the ore candidate region image set and ore position information (including the displacement trigger number of the ore candidate region in the original ore image stream, the pixel spatial position in the corresponding frame, and the transmission direction displacement position in the transport belt displacement coordinate system after mapping by the synchronization parameter set) through connected component segmentation.

[0037] The transmission belt displacement-time mapping relationship in the synchronization parameter set is invoked to convert the continuous motion process of the transmission belt into a predictable sequence of displacement nodes, which serves as the sole triggering reference.

[0038] Furthermore, the transmission belt displacement-time mapping relationship is read from the synchronization parameter set, and the continuous change process of the cumulative displacement of the transmission belt is discretized into a set of displacement nodes sorted in ascending order according to the displacement conditions and corresponding time markers recorded in the transmission belt displacement-time mapping relationship. At the same time, the corresponding time marker or corresponding time given by the transmission belt displacement-time mapping relationship is associated with each displacement node. The displacement nodes and their corresponding time markers or corresponding times are aggregated to form a predictable displacement node sequence, and the predictable displacement node sequence is used as the unique triggering reference.

[0039] The system tracks the operation of the conveyor belt in real time and generates an imaging trigger signal when the conveyor belt displacement reaches a preset constant displacement condition.

[0040] Furthermore, during the operation of the conveyor belt, operating parameters are continuously collected and written into the time stamp sequence corresponding to the synchronization parameter set. Based on the cumulative displacement or displacement increment in the operating parameters, the cumulative displacement of the conveyor belt is updated in real time to achieve real-time tracking of the conveyor belt operation process. The cumulative displacement of the conveyor belt obtained in real time is compared with the target displacement node corresponding to the preset equal displacement conditions one by one. When the cumulative displacement of the conveyor belt reaches or crosses the target displacement node, an imaging trigger signal corresponding to the target displacement node is immediately generated. The imaging trigger signal, the target displacement node number, and the time stamp are recorded synchronously to keep the triggering order traceable.

[0041] It should be noted that the equal displacement condition is determined by continuously collecting and tracking the operating parameters of the conveyor belt, discretizing the continuous displacement process into displacement nodes with fixed step sizes, and setting the displacement interval between adjacent displacement nodes. The equal displacement condition is used to define the fixed displacement step size between adjacent triggers. The unique triggering reference refers to the triggering reference system constructed with the transport belt displacement as the sole criterion. The displacement node sequence formed by discretely based on the equal displacement condition constitutes the specific executable triggering reference. During the operation of the transport belt, all imaging triggers take the displacement nodes in the displacement node sequence as the triggering target. Reaching the preset equal displacement condition is only used as a judgment rule to determine whether the current displacement has reached the next displacement node, thereby ensuring that the imaging trigger is always uniformly controlled by the displacement node sequence.

[0042] The gold ore on the surface of the conveyor belt is imaged sequentially using an imaging trigger signal while the light source is in a stable working state. The ore images are then sorted and numbered according to the displacement trigger sequence to form an original ore image stream.

[0043] Furthermore, under the condition that the stable light source remains unchanged, the imaging trigger signal is received and an imaging acquisition is completed on the gold ore on the surface of the conveyor belt each time the imaging trigger signal arrives, and the corresponding ore image frame is obtained; each ore image frame is bound and recorded with the displacement node number and time stamp corresponding to the imaging trigger signal, and the ore image frames are sorted and numbered according to the displacement trigger order, and the ore image frames are continuously collected in the order of numbering to form the original ore image stream.

[0044] Establish temporal associations between the current frame and adjacent frames before and after it, based on the temporal order of the original ore image stream.

[0045] Furthermore, the ore image frames are traversed sequentially according to the numbering order of the original ore image stream. Each ore image frame is determined as the current frame, and the previous and next ore image frames adjacent to the current frame number are located respectively. Based on the continuity of the displacement triggering order, a one-to-one index association is established between the current ore image frame and the previous and next ore image frames. The current frame number, the previous frame number, the next frame number, and the corresponding time stamp are written into the timing association record, thereby forming the timing association relationship between the current frame and the adjacent frames before and after.

[0046] Based on the temporal correlation, the current frame is subjected to brightness normalization processing, and the normalized frame is subjected to multi-scale edge enhancement and background suppression processing to obtain a standard mineral image.

[0047] Furthermore, the correspondence between the current ore image frame, the previous ore image frame, and the next ore image frame in the temporal correlation relationship is read, the brightness distribution differences within the temporal correlation range are statistically analyzed, and brightness normalization processing is performed on the current ore image frame using a unified brightness reference. Multi-scale edge response calculation and overlay enhancement are performed on the current ore image frame after brightness normalization processing, and background suppression processing is performed simultaneously to reduce the influence of background texture and noise in the transmission band, and a standard ore image is output.

[0048] The standard ore image is binarized, and adjacent pixels are aggregated and labeled based on pixel connectivity to form an initial set of connected regions.

[0049] Furthermore, the pixel values ​​of the standard ore image are compared pixel by pixel with the binarization threshold. Pixels that meet the target foreground conditions are marked as foreground pixels and the remaining pixels are marked as background pixels to obtain a binarized image. On the binarized image, connected component labeling and aggregation are performed on adjacent foreground pixels based on pixel connectivity rules, and the set of pixels with the same connected region identifier is summarized to form an initial connected region set.

[0050] It should be noted that the binarization threshold is obtained by statistically calculating the pixel brightness distribution of the standard ore image. Specifically, the pixel brightness histogram is calculated in the standard ore image, the segmentation position is determined based on the difference in brightness distribution between the foreground ore region and the background region of the transport belt, and the threshold is set based on the brightness value corresponding to the segmentation position. The target foreground conditions are obtained by determining the pixel feature differences between the ore region and the background region of the transport zone in the standard ore image; Pixel connectivity rules are obtained by pre-selecting and fixing the neighborhood connectivity method used during the image processing stage, and are invoked as the basis for determining connected region aggregation before performing connected component labeling processing on the standard ore image.

[0051] The initial set of connected regions is temporally consistent with the connected regions at corresponding positions in adjacent frames. Spatially overlapping connected regions in continuously shifted frames are retained, and single-frame noise regions are removed to obtain ore candidate regions.

[0052] Furthermore, based on temporal correlation, the initial connected region sets of the current frame ore image frame and adjacent frames ore image frames are located. Taking the pixel range corresponding to the connected region identifier of the current frame as a reference, the spatial overlap relationship is calculated in the initial connected region sets of adjacent frames and the temporal consistency check is completed. Connected regions that meet the spatial overlap condition in continuous displacement frames are retained and connected regions that only appear in a single frame are removed to obtain ore candidate regions.

[0053] It should be noted that the spatial overlap relationship is calculated in the initial connected region set of adjacent frames, and the expression is: ; In the formula, It is the spatial overlap ratio between the connected regions in the current frame of the mineral image and the connected regions in the adjacent frame of the mineral image, used to quantify the spatial consistency of the two connected regions at the pixel level. It is the pixel range corresponding to a certain connected region identifier in the current frame of the mineral image frame. The pixel range consists of all pixel positions belonging to the same connected region identifier. It is the pixel range corresponding to a certain connected region identifier in the ore image frame adjacent to the current frame. The pixel range consists of all pixel positions belonging to the same connected region identifier in the adjacent frames. It is the temporal index of the current ore image frame in the original ore image stream; It is the temporal index of the ore image frame adjacent to the current ore image frame. This indicates the previous or next frame of the ore image. ; It should also be noted that the spatial overlap condition refers to the fact that the pixel range corresponding to the connected region identifier in the current frame of the ore image overlaps with the pixel range corresponding to the connected region identifier in the previous frame or the next frame of the ore image in terms of spatial position. The spatial overlap condition is obtained by comparing the pixel ranges corresponding to the connected region identifiers in different temporal ore image frames pixel by pixel and calculating the ratio of pixel intersection or overlap area.

[0054] The candidate region images are cropped according to the pixel range of each ore candidate region in the current frame, and the candidate region images are collected in the order of acquisition to form an ore candidate region image set.

[0055] Furthermore, the connected region identifiers of the ore candidate regions are traversed in the current frame of the ore image. The pixel range corresponding to each connected region identifier is read and the bounding rectangle clipping boundary of the pixel range is calculated. The candidate region image is clipped from the current frame of the ore image according to the bounding rectangle clipping boundary, and the candidate region image is bound and recorded with the displacement trigger number and connected region identifier of the current frame of the ore image. The candidate region images are sequentially appended and collected according to the acquisition order of the original ore image stream to form a set of ore candidate region images.

[0056] Simultaneously, the position information of each ore candidate region in the original ore image stream is recorded and mapped to the transport belt displacement coordinate system through a synchronization parameter set to generate ore position information.

[0057] Furthermore, when cropping candidate region images, the displacement trigger number, circumscribed rectangle cropping boundary, and pixel range center position of the ore candidate region in the current frame of the ore image are recorded. The transmission band displacement time mapping relationship and imaging execution interval are read from the synchronization parameter set, and the cumulative displacement of the transmission band corresponding to the displacement trigger number is used as the displacement reference. Based on the projection range of the camera field of view on the transmission band, the pixel coordinates of the pixel range center position of the ore candidate region in the transmission direction are converted into displacement increments, and superimposed with the displacement reference and mapped to the transmission band displacement coordinate system to generate ore position information that corresponds one-to-one with the ore candidate region.

[0058] The conveyor belt displacement coordinate system is established by calibrating the conveyor belt operation process, with the transmission direction as the main axis and a preset reference position as the zero point, and the displacement information of the synchronization parameter set is mapped to a unified displacement reference frame.

[0059] The identification and sorting module inputs the image set of ore candidate areas into the image recognition model, generates an ore category probability map through pixel-level segmentation, and maps and discretizes the ore category probability map into the sorting control grid according to the sorting execution control parameters and ore location information to generate an initial ore sorting scheme.

[0060] The image set of ore candidate regions is reorganized according to the acquisition order and displacement trigger number, and based on the temporal consistency relationship, the ore candidate regions of the same ore are associated as candidate target temporal instances across frames.

[0061] Furthermore, the candidate ore region image set is traversed according to the acquisition order of the original ore image stream, and reorganized into a candidate region sequence arranged in ascending order of displacement trigger number using displacement trigger number as the index key; between the candidate region sequences corresponding to adjacent displacement trigger numbers, cross-frame matching is performed based on the spatial overlap relationship of pixel range, and the ore candidate regions that meet the spatial overlap condition are associated and aggregated in order of displacement trigger number to form a temporal instance of candidate target.

[0062] Each candidate target time-series instance is bound to a pixel position in the original ore image stream, and the pixel position is mapped to the transport belt displacement coordinate system corresponding to the ore position information to obtain a set of candidate target time-series instances.

[0063] Furthermore, the candidate target time series instances are traversed one by one, and the pixel range of the ore candidate region corresponding to each displacement trigger number in the candidate target time series instance is read. The pixel range is then bound and recorded with the ore image frame number and the outer rectangle clipping boundary in the original ore image stream to complete the pixel position binding of the candidate target time series instances. Based on the pixel position binding, according to the mapping method of the synchronization parameter set in the ore position information generation step, the pixel position is converted into the transmission direction displacement increment and superimposed with the cumulative displacement of the transmission band corresponding to the displacement trigger number and mapped to the transmission band displacement coordinate system, thus forming a set of candidate target time series instances.

[0064] The set of candidate target time series instances is input into the image recognition model. A cross-frame memory representation is established for each candidate target time series instance, and pixel-level segmentation is performed at the candidate region scale to output a ore category probability map.

[0065] Furthermore, the candidate target temporal instance set is input into the image recognition model one by one according to the candidate target temporal instance identifier. The image recognition model extracts features from the candidate region image sequence corresponding to the continuous displacement trigger number and merges and accumulates the features of the previous frame and the current frame to form a cross-frame memory representation of each candidate target temporal instance. Based on the cross-frame memory representation, pixel-level segmentation and decoding are performed on the candidate region image at the candidate region scale. The probability value of the corresponding ore category is output for each pixel of the candidate region image, and the pixel-level probability values ​​are organized into an ore category probability map according to the ore category channel.

[0066] It should be noted that for each pixel in the candidate region image, the probability value corresponding to the mineral category is output, and the expression is: ; In the formula, The pixel position in the candidate region image Belongs to the category of minerals The probability value is used to characterize the pixel position. Classified as ore The degree of confidence; It is an ore category identifier used to distinguish different ore categories in pixel-level segmentation tasks; It is a pixel position index in the candidate region image, used to identify the specific pixel position in the candidate region image where pixel-level segmentation is performed; Pixel position In ore categories The corresponding category score is used to characterize the pixel location. With ore category The matching strength; Pixel position In ore categories The corresponding category score is used to characterize the pixel location. With ore category The matching strength; It is an ore category index used to traverse pixel positions. All corresponding ore categories; This represents the total number of ore categories, used to indicate the number of ore categories participating in the probability normalization calculation in the pixel-level segmentation output; where, Category scores are obtained by using image recognition models to determine pixel locations during pixel-level segmentation and decoding. Cross-frame memory features through mineral categories The corresponding unnormalized category score output by the classification channel; The category score is the score given by the image recognition model for pixel location during pixel-level segmentation and decoding. Cross-frame memory features through various ore categories The corresponding classification channels output unnormalized category scores in parallel.

[0067] It should also be noted that the image recognition model is a pixel-level segmentation and recognition network for the set of candidate ore region images and the set of candidate target temporal instances. The image recognition model takes the candidate region image as input, outputs the ore category probability map at the pixel position scale, and fuses cross-frame information to form a cross-frame memory representation under the condition of continuous frame input of candidate target temporal instances, which is used to improve the stability of the ore category probability map in continuous displacement frames.

[0068] The image recognition model comprises a feature extraction and encoding layer, a cross-frame memory representation layer, a pixel-level segmentation and decoding layer, and a mineral category probability output layer. Specifically: the feature extraction and encoding layer consists of multiple convolutional sub-layers sequentially connected to downsampling sub-layers, used to form multi-scale encoded features; the cross-frame memory representation layer consists of a temporal feature fusion structure, used to fuse the encoded features of the current frame with the memory features corresponding to the previous shift trigger number to generate updated memory features; the pixel-level segmentation and decoding layer consists of multiple upsampling sub-layers alternately connected to convolutional sub-layers, and is concatenated and fused with the output features of the feature extraction and encoding layer at the corresponding level; the mineral category probability output layer consists of a pixel-level classification mapping structure connected to a Softmax normalization structure.

[0069] The feature extraction and encoding layer extracts local texture features and high-level semantic features of the candidate region image step by step through multi-layer convolution and downsampling operations. The cross-frame memory representation layer performs temporal fusion operations on the frame features corresponding to the continuous displacement trigger numbers, merges the memory features of the previous frame with the encoded features of the current frame, and updates the memory state to form a cross-frame memory representation. The pixel-level segmentation and decoding layer restores the cross-frame memory representation to the spatial resolution of the candidate region image step by step through upsampling and convolution operations, and generates an ore category score vector at each pixel position. The ore category probability output layer performs Softmax normalization on the ore category score vector to obtain the ore category probability value at each pixel position and assembles it into an ore category probability map.

[0070] The feature extraction and encoding layer and the cross-frame memory representation layer are directly connected through a feature interface. The output port of the feature extraction and encoding layer corresponds one-to-one with the input port of the cross-frame memory representation layer. The output port of the cross-frame memory representation layer is directly connected to the input port of the pixel-level segmentation and decoding layer, forming a sequential cascaded structure. The intermediate output ports of different spatial resolutions in the feature extraction and encoding layer are connected in parallel to the corresponding upsampling sub-layers in the pixel-level segmentation and decoding layer for feature concatenation. The final output port of the pixel-level segmentation and decoding layer is directly connected to the input port of the ore category probability output layer, thus forming an image recognition model that includes sequential connections and cross-layer parallel connections.

[0071] Training the image recognition model: The set of candidate ore region images is used as the training input, and a pixel-level ore category annotation map is configured for each candidate ore region image as a supervision signal. At the same time, the continuous frame input is organized according to the temporal instance set of candidate targets to meet the temporal learning conditions of cross-frame memory representation level. In the forward computation stage, the continuous frame candidate region images are input into the image recognition model to output the ore category probability map. In the loss calculation stage, the pixel-level segmentation loss is calculated by the ore category probability map and the pixel-level ore category annotation map. In the parameter update stage, the pixel-level segmentation loss is backpropagated and the image recognition model parameters are updated. Iterative training continues until the pixel-level segmentation index of the validation set converges.

[0072] For multiple candidate target time series instances that exist in parallel within the same displacement node, the set interactive recognition mechanism is invoked to interactively model the intermediate recognition representations of each candidate target time series instance, forming context consistency constraints.

[0073] Furthermore, intermediate identification representations corresponding to parallel candidate target temporal instances are aggregated within the same displacement node. The set interaction identification mechanism is invoked to perform pairwise similarity calculation and weighted information interaction on the intermediate identification representations to obtain updated intermediate identification representations that fuse the semantic information of neighboring candidate targets. Contextual consistency constraints for constraining subsequent pixel-level segmentation and decoding are then generated based on the updated intermediate identification representations.

[0074] It should be noted that the ensemble interaction recognition mechanism refers to a recognition method that jointly models the intermediate recognition representations of multiple candidate target time series instances within the same displacement node. By performing similarity calculation and information interaction on the intermediate recognition representations between candidate target time series instances, the intermediate recognition representation of each candidate target time series instance simultaneously contains its own feature information and the contextual feature information of other candidate targets within the same displacement node. The ensemble interactive recognition mechanism is obtained by pre-constructing an interactive computational structure that supports multi-instance feature input in the image recognition model. During the training phase of the image recognition model, the interactive computational structure learns from samples where multiple ores appear simultaneously within the same displacement node, thereby forming a stable interactive modeling capability between candidate targets.

[0075] The pixel-level segmentation and decoding process is constrained and corrected based on context consistency constraints, and the ore category probability map is corrected for consistency to obtain a stable ore probability map.

[0076] Furthermore, context consistency constraints are input into the pixel-level segmentation and decoding process. During the decoding stage, based on the semantic consistency results among multiple candidate target temporal instances within the same displacement node, adjustment weights are applied to the output scores of the pixel-level classification channels. The pixel-level category scores corresponding to ore categories that match the context consistency constraints are enhanced, while the pixel-level category scores corresponding to ore categories that deviate from the context consistency constraints are suppressed. After decoding is completed, the category probabilities in the ore category probability map that are inconsistent with the context consistency constraints are further normalized and corrected, so that each candidate target maintains a consistent and stable semantic discrimination result within the same displacement node, thereby obtaining a stable ore probability map.

[0077] It should be noted that the pixel-level segmentation and decoding process refers to the feature decoding process in the image recognition model, which involves restoring the cross-frame memory representation to the spatial resolution of the candidate region image through upsampling and convolution operations, and generating the corresponding ore category score at each pixel location.

[0078] A three-level mapping chain is established based on ore location information and sorting execution control parameters; based on the three-level mapping chain, the ore category probability corresponding to the pixel in the stable ore probability map is mapped and discretized to the sorting control grid.

[0079] Furthermore, based on the transport belt displacement coordinates in the ore location information and the pixel positions of the candidate region image, a first mapping relationship is established from the pixel positions of the stable ore probability map to the transport belt displacement coordinates. Based on the mapping relationship between the imaging execution interval and the transport belt displacement time in the synchronization parameter set, a second mapping relationship is established to map the transport belt displacement coordinates to the arrival time or arrival displacement corresponding to the sorting execution area. Based on the grid division rules of the sorting control grid and the correspondence between the execution units in the sorting execution control parameters, a third mapping relationship is established from the sorting execution area to the grid nodes, thus forming a three-level mapping chain. The ore category probability is read pixel by pixel in the stable ore probability map. According to the three-level mapping chain, the transport belt displacement coordinates corresponding to the pixels are converted into target grid nodes in the sorting control grid. The ore category probabilities are aggregated and statistically analyzed according to the target grid nodes and then discretized and allocated, so that the ore category probabilities in the stable ore probability map are mapped and discretized to the sorting control grid.

[0080] It should be noted that the grid division rule is generated by calibrating the structure and function of the sorting execution mechanism, collecting the actual installation position, start and end boundaries of action, and effective sorting coverage of each sorting execution unit in the transport belt coordinate system; combining the response timing and minimum control resolution of the execution unit, the sorting execution area is discretized in the transmission direction and width direction; and the discretization result is then correlated with the control number of each execution unit.

[0081] The probability of ore category carried by each grid node in the sorting control grid is discretized at the grid level to generate an initial ore sorting scheme.

[0082] Furthermore, the probability vector of the ore category carried by each grid node in the sorting control grid is read, and the ore category or rejection status corresponding to the grid node is determined by grid-level discretization. The judgment results of all grid nodes are then collected in the order of the sorting control grid to generate an initial ore sorting scheme.

[0083] The decision execution module performs uncertainty assessment and rejection re-inspection on the initial ore sorting plan, determines the ore sorting plan, and executes automatic ore sorting in the sorting execution area according to the ore sorting plan, generating an ore sorting report.

[0084] Based on the consistency between the image recognition confidence distribution and temporal sequence of each ore in the initial ore sorting scheme, the uncertainty of the initial sorting results is evaluated.

[0085] Furthermore, for each sorting result in the initial ore sorting scheme, the corresponding ore category probability vector is read to form an image recognition confidence distribution. This is combined with the ore category probability changes under continuous displacement trigger numbers in the candidate target time-series instance set to calculate a time-series consistency index. First, the distribution concentration of the ore category probability vector is quantified to characterize the confidence strength of the current sorting result in category determination. Then, the stability of the ore category probability changes under continuous displacement trigger numbers is quantified to characterize the consistency of the sorting results in the time-series dimension. Subsequently, the above two types of quantification results are normalized and weighted and fused according to preset weights to form a comprehensive evaluation value used to characterize the reliability of the sorting results. This comprehensive evaluation value is mapped to an uncertainty metric value, thus forming the degree of uncertainty of the initial sorting results.

[0086] Based on the degree of uncertainty, the initial sorting results that do not meet the preset confidence conditions are rejected, and a re-inspection is triggered to determine the ore sorting scheme.

[0087] Furthermore, the uncertainty level of each sorting result in the initial ore sorting scheme is compared with the preset confidence conditions. The initial sorting results that do not meet the preset confidence conditions are rejected and a re-inspection is triggered. The ore sorting scheme is determined based on the re-inspection results and the initial sorting results that meet the preset confidence conditions.

[0088] It should be noted that the preset confidence conditions are set by statistically analyzing the range of values ​​of the image recognition confidence distribution and temporal consistency index on training samples and historical sorting results, and based on the balance requirements of sorting accuracy and rejection rate.

[0089] According to the ore sorting plan, the corresponding automatic ore sorting operation is performed in the sorting execution area to obtain the automatic ore sorting results. By summarizing and recording, an ore sorting report is generated.

[0090] Furthermore, based on the sorting execution instructions in the ore sorting scheme, and combined with the ore location information and sorting execution control parameters, the corresponding automatic ore sorting operation is performed on the ore that has reached the sorting execution position in the sorting execution area, and the execution time, execution position, and execution result are recorded simultaneously to obtain the automatic ore sorting result; the automatic ore sorting result is summarized according to the sorting execution number and time sequence, and recorded in correspondence with the judgment result of the ore sorting scheme to generate an ore sorting report.

[0091] In summary, this invention achieves precise synchronization between image recognition-based ore image data and the actual position of the ore by constructing a conveyor belt displacement-time mapping and using an equal displacement triggering method for ore imaging; furthermore, it generates an ore category probability map through pixel-level segmentation in image recognition and maps it to a sorting control grid to form a directly executable sorting scheme, thereby effectively improving the synchronization consistency, identification reliability, and execution stability of automatic ore sorting under continuous conveyor belt operation conditions.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic sorting system for gold mine belt ore based on image recognition, characterized in that: include, The synchronous modeling module collects the operating parameters of the transport belt and measures the distance between imaging execution zones, establishes the displacement-time mapping relationship of the transport belt, and generates a synchronous parameter set. The image acquisition module, based on a synchronization parameter set, continuously scans the gold ore on the surface of the conveyor belt under stable light source conditions using an equal displacement triggering method to acquire the original ore image stream. It also performs image preprocessing on each frame of the ore image and obtains the ore candidate region image set and ore location information through connected component segmentation. The identification and sorting module inputs the image set of ore candidate areas into the image recognition model, generates an ore category probability map through pixel-level segmentation, and maps and discretizes the ore category probability map into the sorting control grid according to the sorting execution control parameters and ore location information to generate an initial ore sorting scheme. The decision execution module performs uncertainty assessment and rejection re-inspection on the initial ore sorting plan, determines the ore sorting plan, and executes automatic ore sorting in the sorting execution area according to the ore sorting plan, generating an ore sorting report.

2. The automatic sorting system for gold mine belt ore based on image recognition as described in claim 1, characterized in that: The specific steps for collecting the transport belt's operating parameters, measuring the imaging execution zone spacing, establishing the transport belt displacement-time mapping relationship, and generating a synchronization parameter set are as follows. Collect the operating parameters of the conveyor belt, and time-stamp the operating parameters to form traceable operating data. Based on the traceable operating data, determine the motion change relationship of the conveyor belt at each time point. The imaging area is determined by projecting the camera's field of view onto the conveyor belt. The sorting execution area is determined by the effective range of the sorting mechanism on the conveyor belt. The distance between the corresponding baselines of the imaging area and the sorting execution area in the transmission direction is measured to obtain the imaging execution area spacing.

3. The automatic sorting system for gold mine belt ore based on image recognition as described in claim 1, characterized in that: The specific steps for establishing the transmission band displacement-time mapping relationship and generating a synchronization parameter set are as follows. The displacement of the transmission belt is updated every moment according to the motion change relationship. The corresponding time is determined by the imaging execution zone spacing as the judgment condition, and the time mapping relationship of the transmission belt displacement is obtained. The transmission belt operating parameters, imaging execution zone spacing, and transmission belt displacement time mapping relationship are uniformly encapsulated to form a synchronization parameter set.

4. The automatic gold mine belt ore sorting system based on image recognition as described in claim 3, characterized in that: The method involves continuously scanning the gold ore surface of the transport belt under stable light source conditions using an equal displacement triggering method, based on a synchronization parameter set, to acquire the original ore image stream. The specific steps are as follows: The transmission belt displacement-time mapping relationship in the synchronization parameter set is invoked to convert the continuous motion process of the transmission belt into a predictable sequence of displacement nodes and use it as the sole triggering reference. The conveyor belt operation process is tracked in real time, and an imaging trigger signal is generated when the conveyor belt displacement reaches the preset equal displacement condition. The gold ore on the surface of the conveyor belt is imaged sequentially using an imaging trigger signal while the light source is in a stable working state. The ore images are then sorted and numbered according to the displacement trigger sequence to form an original ore image stream.

5. The automatic sorting system for gold mine belt ore based on image recognition as described in claim 1, characterized in that: The specific steps for image preprocessing of each frame of the ore image are as follows: Establish temporal associations between the current frame and adjacent frames before and after it, according to the temporal order of the original ore image stream; Based on the temporal correlation, the current frame is subjected to brightness normalization processing, and the normalized frame is subjected to multi-scale edge enhancement and background suppression processing to obtain a standard mineral image.

6. The automatic gold mine belt ore sorting system based on image recognition as described in claim 5, characterized in that: The specific steps for obtaining the ore candidate region image set and ore location information through connected component segmentation are as follows: The standard ore image is binarized, and adjacent pixels are aggregated and labeled based on pixel connectivity to form an initial set of connected regions. The initial set of connected regions is temporally consistent with the connected regions at corresponding positions in adjacent frames. Spatially overlapping connected regions in continuously shifted frames are retained, and single-frame noise regions are removed to obtain ore candidate regions. The candidate region images are cropped according to the pixel range of each ore candidate region in the current frame, and the candidate region images are collected in the order of acquisition to form a set of ore candidate region images; Simultaneously, the position information of each ore candidate region in the original ore image stream is recorded and mapped to the transport belt displacement coordinate system through a synchronization parameter set to generate ore position information.

7. The automatic sorting system for gold mine belt ore based on image recognition as described in claim 4, characterized in that: The constant displacement condition is determined by continuously collecting and tracking the operating parameters of the conveyor belt, discretizing the continuous displacement process into displacement nodes with fixed step sizes, and setting the displacement interval between adjacent displacement nodes.

8. The automatic sorting system for gold mine belt ore based on image recognition as described in claim 6, characterized in that: The specific steps for inputting the set of candidate ore region images into the image recognition model and generating an ore category probability map through pixel-level segmentation are as follows. The image set of ore candidate regions is reorganized according to the acquisition order and displacement trigger number, and based on the temporal consistency relationship, the ore candidate regions of the same ore are associated as candidate target temporal instances across frames. Each candidate target time-series instance is bound to a pixel position in the original ore image stream, and the pixel position is mapped to the transport band displacement coordinate system corresponding to the ore position information to obtain a set of candidate target time-series instances; The set of candidate target time series instances is input into the image recognition model. A cross-frame memory representation is established for each candidate target time series instance, and pixel-level segmentation is performed at the candidate region scale to output a ore category probability map.

9. The automatic gold mine belt ore sorting system based on image recognition as described in claim 6 or 8, characterized in that: The initial ore sorting scheme is generated by mapping and discretizing the ore category probability map onto the sorting control grid based on the sorting execution control parameters and ore location information. The specific steps are as follows. For multiple candidate target time series instances that exist in parallel within the same displacement node, the set interactive recognition mechanism is invoked to interactively model the intermediate recognition representations of each candidate target time series instance, forming context consistency constraints. The pixel-level segmentation and decoding process is constrained and corrected based on context consistency constraints, and the ore category probability map is corrected for consistency to obtain a stable ore probability map. A three-level mapping chain is established based on ore location information and sorting execution control parameters; Based on a three-level mapping chain, the ore category probabilities corresponding to pixels in the stable ore probability map are mapped and discretized to the sorting control grid. The probability of ore category carried by each grid node in the sorting control grid is discretized at the grid level to generate an initial ore sorting scheme.

10. The automatic sorting system for gold mine belt ore based on image recognition as described in claim 9, characterized in that: The process involves uncertainty assessment and rejection / re-inspection of the initial ore sorting scheme to determine the final ore sorting plan. Automatic ore sorting is then performed in the sorting execution area according to the chosen plan, generating an ore sorting report. The specific steps are as follows: Based on the consistency between the image recognition confidence distribution and temporal sequence of each ore in the initial ore sorting scheme, the degree of uncertainty of the initial sorting results is evaluated. Based on the degree of uncertainty, the initial sorting results that do not meet the preset confidence conditions are rejected, and a re-inspection is triggered to determine the ore sorting scheme. According to the ore sorting plan, the corresponding automatic ore sorting operation is performed in the sorting execution area to obtain the automatic ore sorting results. By summarizing and recording, an ore sorting report is generated.

Images