Method and system for automatically detecting molybdenum ore resources based on improved YOLOv10

Abstract

The invention discloses an improved YOLOv10-based molybdenum ore resource automatic detection method and system, and belongs to the technical field of computer vision beneficiation, and the improved YOLOv10-based molybdenum ore resource automatic detection method and system comprise a YOLOv10-based improved model and an X-ray sorting technology. According to the method, a complete automatic detection and separation process is formed by organically combining a YOLOv10-based improved model and an X-ray separation technology, and a formed molybdenum ore resource detection and separation scheme is not limited by physical and chemical properties of molybdenum ore; compared with the traditional manual detection and sorting mode, the molybdenum ore resource sorting method has the advantages of greatly reducing manual intervention and operation time, and remarkably improving the working efficiency of the whole molybdenum ore resource sorting process, and has the remarkable advantages in the aspects of detection accuracy, sorting efficiency, resource utilization rate, adaptability to different working conditions and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision mineral processing, and specifically relates to a method and system for automatically detecting molybdenum ore resources based on improved YOLOv10. Background Art

[0002] In recent years, with the continuous development of deep learning methods and machine vision technology at home and abroad, the cross-integration of deep learning and mineral processing technology has provided a new idea for the construction of smart mines. Deep learning technology and machine vision have begun to be applied in the field of mineral processing. PATEL et al. developed an iron ore classification technology based on online vision. They built a model using multi-class support vector machine (SVM) specifically for the classification of iron ore. A total of 2,200 images were obtained for the development of the iron ore classification model, and the classification error rate was only 0.27%. However, this method is based on artificially designed feature extraction strategies, which are not well applicable to molybdenum ore. It is difficult to accurately extract effective features when facing molybdenum ore. Moreover, for molybdenum ore, more complex factors need to be considered in the preprocessing stage. The huge preprocessing workload of the model will become more unbearable and cannot meet the actual needs of molybdenum ore sorting. AGRAWAL et al. proposed a deep learning mineral classification framework called mineral-ResNet, and experimentally verified it with AVIRIS hyperspectral images of the copper mining area. The framework showed an overall accuracy of 92.16% and a Kappa coefficient of 0.89, showing effective recognition capabilities for most minerals. However, compared with the minerals in the copper mining area, the spectral characteristics of molybdenum ore are quite different, making it difficult for the mineral-ResNet framework to accurately capture the characteristic information of molybdenum ore. In addition, the distribution and morphology of molybdenum ore may also be different from those of the minerals in the copper mining area, which will cause the framework to deviate when processing molybdenum ore data. EBRAHIMI et al. developed a computer vision scheme combining hierarchical analysis method and feature mapping technology, which showed good results in identifying 16 common ores. However, the chemical stability of molybdenum ore is relatively high, and the formation and change of its surface oxide layer are different from other common ores. This will cause the color change of molybdenum ore to be relatively small under different environmental conditions, unlike some ores that will produce obvious color changes during oxidation. This will affect the change pattern of its image features, making it impossible for computer vision methods based on common ore features to accurately identify molybdenum ore. SI NGH et al. used image data and radial basis function neural network to classify the feed ore of ferromanganese smelter, achieving an identification accuracy of 88.71%. However, since molybdenum ore mainly exists in the form of molybdenite, which has low hardness and is easy to produce powder, affecting image clarity and feature stability, this method is difficult to accurately sort molybdenum ore. Wen Dapeng et al. used laser induced breakdown spectroscopy (LIBS) technology combined with principal component analysis-particle swarm optimization-support vector (PCA-PSO-SVM) algorithm to classify and identify 12 types of ores, and achieved good results.

[0003] However, molybdenite has a unique hexagonal layered structure, which means that the molybdenum atoms are arranged in a special, layered manner inside it. This structure makes the arrangement of molybdenum atoms different from that of the other 12 types of ores. Under laser excitation, its internal electron transitions and the formation of plasma will have unique spectral characteristics. The layered structure makes the paths of energy transfer and photon emission more complicated, and the intensity and position of the resulting spectral lines are different from those of general ores. At the same time, some current concentrators have many deficiencies in the processes of hand selection, flotation, gravity selection and magnetic selection of molybdenum ore waste. On the one hand, the overall work efficiency is low and it is difficult to meet the needs of efficient production; on the other hand, there are omissions and misselections in the accuracy of selection, and it is impossible to accurately screen and process molybdenum ore resources. In addition, the grade and recovery rate of molybdenum ore are generally not high, resulting in a waste of resources. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method and system for automatic detection of molybdenum ore resources based on improved YOLOv10.

[0005] The technical solution adopted to solve the above technical problems is: an automatic detection method and system for molybdenum ore resources based on improved YOLOv10, including a model improved based on YOLOv10 and X-ray sorting technology. The X-ray sorting technology is used as a front-end detection method to perform preliminary physical property detection on molybdenum ore and convert the detection results into data form that can be used for subsequent model analysis;

[0006] The improved model based on YOLOv10 serves as the intelligent analysis core of the back-end, and conducts in-depth mining and precise judgment on the data transmitted from the front-end, thereby realizing the accurate classification and positioning of molybdenum ore, providing a reliable basis for the final sorting operation.

[0007] Through the above technical solutions, targeted optimization measures such as enhancing feature extraction capabilities, optimizing multi-scale feature fusion, improving upsampling effects, and optimizing prediction box regression are aimed at realizing automated detection to improve work efficiency, improve the accuracy of molybdenum ore resource detection to reduce missed selections and wrong selections, and help improve the grade control and recovery rate of molybdenum ore, thereby improving resource utilization and reducing resource waste.

[0008] Furthermore, the improved model based on YOLOv10 first uses CloAtt to enhance the C2F module in Backbone. In the new structure, a branch uses deep separable convolution to capture high-frequency local information and uses context-aware weights to enhance these features. High-frequency feature information such as local texture and shape of molybdenum ore targets, as well as context information, are crucial for locating and identifying targets. This branch can reduce the impact of interference factors on feature extraction, while enhancing key features related to the target, and improving the accuracy of the model in distinguishing between the target and the background.

[0009] The other branch obtains low-frequency global information through the attention mechanism, which facilitates the understanding of the overall layout of the image and the relationship between regions. In the molybdenum mine X-ray image, the global information helps the model to grasp the distribution of molybdenum mine targets and their relative relationship with the background as a whole, avoiding misjudgment. Finally, the information of the two branches is fused to integrate high-frequency local information and low-frequency global information.

[0010] In the global branch, in order to obtain low-frequency global information, Kglobal and Vglobal obtained by linear transformation are first downsampled, and then the traditional attention process is performed. The specific formula is:

[0011] Q global ,K global ,V global =FC(X in )

[0012] X G =Attn(Q global ,Pool(K global ),Pool(V global) ))

[0013] First, Qlocal, Klocal, and Vlocal are generated through linear transformation. Subsequently, Vlocal is processed through deep convolution (DWconv) to achieve effective aggregation of local information. This design allows the model to capture high-frequency local features in the image in detail while maintaining sensitivity to global contextual information. The specific formula is:

[0014] Q lobal ,K lobal ,V lobal =FC(X in )

[0015] V DW =DWconv(V local ).

[0016] The above technical solution allows the model to capture the high-frequency local features in the image in detail while maintaining sensitivity to global contextual information, which is crucial to improving the detection accuracy of targets in molybdenum mine X-ray images.

[0017] Furthermore, in order to further enhance the model's ability to process local features, two depth-wise separable convolutions (DW-Convs) are first used to aggregate Qlocal and Klocal, respectively. In order to obtain context-aware weights between -1 and 1, a series of transformations are performed on the Hadamard product of Qlocal and Klocal, including normalization and the application of activation functions. Finally, by applying these context-aware weights to local features, the model's ability to capture local details is further enhanced, thereby improving the detection accuracy of targets in molybdenum mine X-ray images. The detailed steps of the entire process are as follows:

[0018] Q L =DWconv(Q local )

[0019] K L =DWconv(K local )

[0020] A L =FC(Swish(FC(Q L E L )))

[0021]

[0022] X L =AttneV DW

[0023] After completing the above local feature enhancement steps, the features extracted by the two branches are further integrated by connecting their outputs in the channel dimension. This connection operation enables the model to consider both high-frequency local information and low-frequency global information, thereby obtaining a more comprehensive image representation. Then, in order to extract the final output from these integrated features, a fully connected (FC) module is introduced. The FC module generates a comprehensive, high-dimensional feature representation by learning the complex relationship between the input features. The calculation process of the final output can be expressed as:

[0024] X C =concat(X G ,X L )

[0025] X Out =FC(X C )

[0026] Through the above enhancements, the C2F module can more keenly capture the subtle feature information of the texture, shape, and color of the ore when processing images.

[0027] Through the above technical solutions, the enhanced C2F module can better extract the unique crystal structure characteristics of molybdenum ore, providing strong support for the accurate judgment of the model.

[0028] Furthermore, the SPPF module on Backbone is replaced by the FocalModulation module. This module uses the focal modulation mechanism to replace the traditional self-attention mechanism. The self-attention mechanism requires complex query-key interaction and query-value aggregation for each query token with other tokens to calculate the attention score and capture context information. Its mathematical expression is as follows:

[0029] y i =M1(T1(x i ,X),X)

[0030] The focus modulation mechanism consists of three key components: focus contextualization, gated aggregation, and element-wise affine transformation. These components work together to effectively encode visual context and adaptively inject this information into query tokens to enhance features. The focus modulation mechanism first aggregates spatial context information into modulators at different granularities, and then injects this modulated information into query tokens based on the query dependencies. Specifically, the mathematical expression of focus modulation is as follows:

[0031] y i =q(x i )e M2(x i ,X)

[0032] Among them, M represents the aggregation process, T represents the interaction process, and q(x i ) is a query mapping function, and the Hadamard product is represented by ⊙. In the aggregation process M2, context calculations are reduced by sharing operators (such as deep convolutions, etc.), while the M1 in the self-attention mechanism is more expensive to calculate because it needs to sum the non-shared attention scores of different queries. The aggregation process M2 includes two steps:

[0033] Step 1: Hierarchical contextualization to extract contextual information at different granularity levels from local to global scope;

[0034] Step 2: All contextual features at different granularity levels are condensed into a single feature vector, the modulator, through gated aggregation.

[0035] Through the above technical solution, the detection rate of small-sized molybdenum ore targets that are difficult to detect in images can be significantly improved. At the same time, the molybdenum ore targets can be more accurately located in complex backgrounds and background interference can be reduced.

[0036] Further, hierarchical contextualization: given an input feature map X, it is first projected into a new feature space with a linear layer Z0=fz(X)€RH*W*C and then L depth-wise convolutions are used to obtain a hierarchical representation of the context. The output Zl is expressed as:

[0037]

[0038] Compared with conventional convolution, it has channel characteristics and hierarchical context to generate L-level feature maps. At the lth level, the size of the effective receptive field rl is:

[0039]

[0040] rl is much larger than the convolution kernel size kl. The larger receptive field captures more global context at a coarser granularity. In order to capture the global context of the entire input, global average pooling ZL+1=AvgPool(ZL) is applied to the L-th level feature map, thereby obtaining a total (L+1) feature map They jointly capture both local and long-range contexts at different levels of granularity.

[0041] Through the above technical solutions, depth-wise convolution is applied for hierarchical contextualization because it is learnable and structure-aware compared to pooling. Compared with regular convolution, it has channel characteristics and therefore has lower computational cost.

[0042] Further, gated aggregation: The (L+1) feature maps obtained by hierarchical contextualization are compressed into a modulator, i.e., a single feature vector. In an image, the relationship between a token and its surrounding context often depends on the content of the query. For example, queries for salient visual objects may rely heavily on local fine-grained features, but queries for background scenes may rely mainly on global coarse-grained features. Therefore, a gating mechanism is used to control the amount of feature maps aggregated from different levels Zl for each query. Specifically, a linear layer is used to obtain spatially and level-aware gating weights G:

[0043] G=f g (X)€R H*W*(L+1)

[0044] Then, a weighted sum is performed by element-wise multiplication to obtain a single feature map Zout of the same size as the input X:

[0045]

[0046] Among them, Gl€RH*W*1 is a channel at level l. So far, all aggregations are spatial aggregations. In order to establish the relationship between different channels, another linear layer h() is used to obtain the modulator M:

[0047] M=h(Z out )€R H*W*C

[0048] Combining the interaction and aggregation formulas, the overall focus modulation formula can be expressed as:

[0049]

[0050] Through the above technical solution, the (L+1) feature map obtained by hierarchical contextualization is compressed into a modulator, that is, a single feature vector.

[0051] Furthermore, the structure and aggregation process of FocalModulation includes that the input is first processed through a linear layer, then, information is selectively aggregated through hierarchical contextualization and gated aggregation, and finally, the output is generated through the interaction between the modulator and the query;

[0052] FocalModulation replaces the original spatial pyramid pooling fast (SPPF), improving the accuracy of the network model without changing the number of parameters and the amount of calculation. In the actual mining and transportation process, the size and shape of molybdenum ore vary greatly. Based on its advanced modulation mechanism, the FocalModulation module can dynamically adjust according to the importance of different scale features in the ore image, deeply fuse the features of different scales, and thus more accurately identify molybdenum ores of different sizes and shapes, improving the model's adaptability to complex and diverse molybdenum ore samples;

[0053] In addition, the Upsample module of the original Head part is replaced with the Dysample module. DySample is an ultra-lightweight and efficient dynamic upsampler. When processing molybdenum ore X-ray images, it can reduce the amount of calculation and processing time while ensuring a certain detection performance. Using DySample as an upsampler to replace the original Upsample module can improve the running speed of the model. DySample can also reduce the model size, which is conducive to system deployment and maintenance, reduce hardware costs, and improve the overall stability and reliability of the system.

[0054] The module structure of the DySample upsampler includes a feature map X of size C*H1*W1 and a grid of size 2g*H2*W2, where the first dimension 2g represents the x and y coordinates of the original feature map X. By using the grid_sample function and the position in the point sampling set δ, the feature map X is resampled to X' of size C*H2*W2. The specific formula is as follows:

[0055] X' = ​​grid_sample(x,δ)

[0056] The generation process of the point sampling set δ based on the dynamic range factor includes, first, given the upsampling scale factor s and the feature map X of size C*H*W, using a linear layer with input and output channel numbers C and 2gs2 respectively to generate an offset O of size 2gs2*H*W. The specific formula is as follows:

[0057] O=linear(X)

[0058] Then, through pixel shufle, it is regenerated into a high-resolution original sampling grid G ​​of size 2g*sH*sW. Then the point sampling set δ is the sum of the offset O and the original sampling grid G. The specific expression is as follows:

[0059] δ=O+G

[0060] The Dysample module can significantly improve the accuracy of upsampling while ensuring computational efficiency. Through this high-precision upsampling process, the processed image can more clearly present the detailed features of molybdenum ore. These detailed information further assists the model in accurate detection and reduces the possibility of misjudgment. Finally, the GIOU loss function is used to replace the original target positioning loss term. The loss function is replaced by GIOU. The original YOLOv10 loss function consists of target positioning loss, confidence loss and classification loss:

[0061] Loss=Loss loc +Loss conf +Loss cls

[0062] The original YOLOv10 model uses CIOU to calculate the target positioning loss, and uses Binary Cross Entropy to calculate the target confidence loss and classification loss. The target positioning loss refers to the difference between the center point coordinate value of the target prediction area and the center point coordinate value of the actual object, and the difference between the width and height of the prediction box and the width and height of the actual box. IOU (Intersection Over Union) is the intersection-and-union ratio between the predicted box and the true value box, which is used to measure the overlap rate between the predicted box and the true value box. Generally speaking, the IOU value is proportional to the model prediction accuracy, but when the predicted box and the true value box do not intersect, its value is 0. At this time, it cannot correctly reflect the distance between the two, and the gradient cannot be regressed, making training impossible. In X-ray-based molybdenum ore image detection, if this happens, it is difficult for the model to accurately mark the true position of the molybdenum ore target, so IOU cannot be used directly as a loss function. Since the regression of the bounding box is related to the overlapping area, aspect ratio and center point distance of the two boxes, CIOU takes into account the consistency of the overlapping area and the distance between the center point of the bounding box and the aspect ratio, which enables CIOU to measure the similarity of the bounding box more comprehensively than IOU, thereby providing more accurate regression guidance. Its calculation formula is:

[0063] L CIOU =1-IOU(A,B)+p 2 (a,b) / c 2 +αv

[0064] Where A is the predicted box; B is the real box; a and b represent the centroids of the predicted box and the true box, respectively; p represents the Euclidean distance between the two centroids; c represents the diagonal distance of the minimum closure area surrounding the two boxes; v is used to compare the consistency of the aspect ratio, and α is the weight coefficient of v.

[0065] α and v are defined as follows:

[0066] α=v / (1-IOC+v)

[0067] v=4 / Π 2 (arctan(w gt / h gt )-arctan(w / h)) 2

[0068] Where wgt and hgt are the width and height of the true value box, w and h are the width and height of the predicted box;

[0069] The calculation formula of GIOU loss function is as follows:

[0070] G=IC / A∩B / C|

[0071] Where G is the GIOU value; A is the predicted box; B is the real box; C is the smallest box containing A and B; I is the IOU value; its loss function is:

[0072] L GIOU =lG

[0073] Since the GIOU evaluation method introduces the minimum intersection C that includes A and B, the objective function can still be optimized when A and B do not overlap. The calculation steps of the GIOU evaluation method are:

[0074] Input: predicted box A, = (x1, p, y1, p, x2, p, y2, p), real box B = (x1, g, y1, g, x2, g, y2, g); Output: LGIOU;

[0075] Calculate the area of ​​A and B:

[0076] A=(x 2,p -x 1,p )*(y 2,p -y 1,p )

[0077] B=(x 2,g -x 1,g )*(y 2,g -y 1,g )

[0078] Calculate the overlapping area Ix,y of A and B:

[0079] x1=max(x1' ,p ,x 1,g ),x2=min(x' 2,p ,x 2,g )

[0080] y1=max(y1' ,p ,y 1,g ),x2=min(y' 2,p ,y 2,g )

[0081]

[0082] Find the minimum area D of A∩B:

[0083] x 1,c =max(x1' ,p ,x 1,g ),x 2,c =min(x' 2,p ,x 2,g )

[0084] y 1,c =max(y1',p ,y 1,g ),x 2,c =min(y' 2,p ,y 2,g )

[0085] D=(x 2,c -x 1,c )*(y 2,c -y 1,c )

[0086] Calculate IOU:

[0087] I=I x,y / (A+BI x,y )

[0088] Calculate GIOU:

[0089] G=I-(B c -(A+BI x,y )) / D.

[0090] Calculate the final loss:

[0091] L GIOU =1-G.

[0092] Through the above technical scheme, in molybdenum ore detection, accurate target positioning is critical for distinguishing different ore individuals and determining their specific positions in the image. Compared with the traditional loss function, the GIOU loss function can more accurately measure the difference between the predicted box and the real box, thereby more effectively adjusting the position and size of the predicted box. When faced with the complex arrangement of molybdenum ore on the conveyor belt, by using the GIOU loss function, the model can more accurately locate the position of each ore, ensuring that subsequent sorting operations can be accurately performed on each ore individual, thereby improving the accuracy and reliability of the entire sorting process.

[0093] Furthermore, the X-ray sorting technology includes the following specific steps:

[0094] Step 1: Ore pretreatment: In the initial stage of the entire molybdenum ore resource detection and sorting process, the ore needs to be pretreated comprehensively and carefully. This step mainly includes crushing, screening and cleaning:

[0095] Crushing: The raw ore usually exists in the form of large blocks, which is not conducive to the subsequent X-ray penetration and sorting operations. The raw ore is crushed by crushing equipment to gradually reduce its particle size, making it easier for subsequent processing. The degree of crushing needs to be reasonably adjusted according to the specific requirements of the sorting equipment and the characteristics of the ore. The particle size of the crushed ore should be controlled within a certain range, which can ensure that the X-ray can penetrate effectively and facilitate the subsequent screening operation;

[0096] Screening operation: The particle size of the crushed ore is not uniform, so screening operation is required. By setting screens with different apertures, the crushed ore is screened according to the particle size, and the part with too large or too small particle size is removed. Only the ore with the particle size meeting the requirements can enter the subsequent X-ray irradiation link, which can ensure that the X-ray can obtain a more consistent penetration effect when penetrating the ore, thereby improving the accuracy of detection;

[0097] Cleaning stage: During the mining and transportation of raw ore, some impurities such as soil and dust often adhere to the surface of the ore. These impurities will affect the penetration effect of X-rays and subsequent detection results. Therefore, after screening, the ore needs to be cleaned by washing it with clean water or cleaning solution to thoroughly remove impurities on its surface to ensure that the ore reaches the particle size range and purity requirements suitable for subsequent X-ray sorting;

[0098] Step 2: X-ray penetration and detection: The pre-treated ore is sent to the X-ray sorting equipment and irradiated with X-rays. Since X-rays have strong penetrating power, most of their energy can pass through the gaps between the atoms of the ore. However, the complexity of the internal structure of the ore and the difference in density will cause X-rays to encounter different degrees of obstruction during the penetration process. The distribution of different mineral components in the ore and the difference in crystal structure will lead to different degrees of difficulty in X-ray penetration. The detector on the other side of the ore can accurately detect the slight differences in the degree of X-ray penetration. The detector uses a highly sensitive detection element and can accurately measure the intensity changes of X-rays. When X-rays penetrate the ore, their intensity will change. The detector can accurately capture the slight differences in the degree of X-ray penetration based on these changes and convert them into electrical signals for analysis. These electrical signals carry important information about the internal structure and density differences of the ore, and are the basic data source for subsequent image processing and target detection.

[0099] Step 3: Image processing and target detection: The electrical signal from the detector is transmitted to the image processing unit. Here, the electrical signal needs to go through the following processing to be converted into an image form that can be analyzed. First, filtering is performed. By using a suitable filtering algorithm, the noise interference in the electrical signal is removed to make the electrical signal purer. Then, enhancement processing is performed. By adjusting the intensity, contrast and other parameters of the electrical signal, the converted image can more clearly present the internal structure and density difference of the ore. After these processes, the electrical signal is finally converted into a grayscale image that can present the internal structure and density difference of the ore. Then, the improved YOL is used. The target detection algorithm in the Ov10 model analyzes the grayscale image. The model first extracts and analyzes the features of each pixel in the image, and uses the improved modules to work together to fully mine the feature information in the image. Then, based on these feature information, the target detection algorithm distinguishes between concentrates and tailings, and accurately marks the position of each metal ore on the conveyor belt. When processing a grayscale image containing multiple molybdenum ores, the model can accurately identify which ores are concentrates and which are tailings, and give the precise coordinate position of each ore in the image. These position information will be directly applied to the subsequent sorting execution link;

[0100] Step 4: Sorting execution: Based on the results of the improved YOLOv10 model marking the ore position and category, when the ore is sent to the actuator through the conveyor belt, the actuator is controlled to spray high-pressure gas, and the ore marked as concentrate is sprayed into the concentrate bin, while the ore marked as tailings falls directly into the tailings bin. In this process, the actuator uses a high-precision gas injection device, which can accurately control the injection direction and intensity of the high-pressure gas according to the ore position and category information marked by the model, ensuring that only the ore marked as concentrate will be sprayed into the concentrate bin, and the tailings will fall into the tailings bin accurately.

[0101] Through the above technical solution, many problems existing in the detection and sorting process of molybdenum ore resources can be effectively solved, such as low detection accuracy and low sorting efficiency, providing strong technical support for the development and utilization of molybdenum ore resources.

[0102] The beneficial effects of the present invention are as follows: The present invention organically combines the improved model based on YOLOv10 with X-ray sorting technology to form a complete set of automatic detection and sorting processes. In this process, from the pretreatment of the raw ore to X-ray irradiation, image processing, target detection, and finally sorting execution, each link is closely connected and has a high degree of automation. The formed molybdenum ore resource detection and sorting scheme is not limited by the physical and chemical properties of molybdenum ore, and shows significant advantages in detection accuracy, sorting efficiency, resource utilization, and adaptability to different working conditions. Compared with traditional manual detection and sorting methods, it greatly reduces manual intervention and operation time, and significantly improves the work efficiency of the entire molybdenum ore resource sorting, which is helpful to further improve resource utilization in the process of molybdenum ore resource development and utilization, and realize the sustainable development of molybdenum ore resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 is a diagram of an X-ray sorting device of the present invention;

[0104] Figure 2 is the image detection and control flow chart of the present invention;

[0105] Figure 3 This is the original YOLOv10 network structure diagram;

[0106] Figure 4 This is a comparison chart between the C2F module and the C2F-CloAtt module;

[0107] Figure 5 It is the structural diagram of FocalModulation;

[0108] Figure 6 It is a diagram of the module structure of the DySample upper sampler of the present invention;

[0109] Figure 7 It is a diagram of the generation process of the point sampling set δ based on the dynamic range factor;

[0110] Figure 8 This is the improved YOLOv10 network structure diagram of the present invention. DETAILED DESCRIPTION

[0111] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0112] like Figure 1-Figure 8As shown, the method and system for automatic detection of molybdenum ore resources based on improved YOLOv10 of this embodiment include a model improved based on YOLOv10 and X-ray sorting technology. The X-ray sorting technology is used as a front-end detection means to perform preliminary physical property detection on the molybdenum ore and convert the detection results into a data form that can be used for subsequent model analysis;

[0113] The improved model based on YOLOv10 serves as the intelligent analysis core of the back-end, and conducts in-depth mining and precise judgment on the data transmitted from the front-end, thereby realizing the accurate classification and positioning of molybdenum ore, providing a reliable basis for the final sorting operation.

[0114] In order to further improve the model's ability to process local features, two depth-wise separable convolutions (DW-Convs) are first used to aggregate Qlocal and Klocal respectively. In order to obtain context-aware weights between -1 and 1, a series of transformations are performed on the Hadamard product of Qlocal and Klocal, including normalization and the application of activation functions. Finally, by applying these context-aware weights to local features, the model's ability to capture local details is further enhanced, thereby improving the detection accuracy of targets in molybdenum mine X-ray images. The detailed steps of the whole process are as follows:

[0115] Q L =DWconv(Q local )

[0116] K L =DWconv(K local )

[0117] A L =FC(Swish(FC(Q L E L )))

[0118]

[0119] X L =AttneV DW

[0120] After completing the above local feature enhancement steps, the features extracted by the two branches are further integrated by connecting their outputs in the channel dimension. This connection operation enables the model to consider both high-frequency local information and low-frequency global information, thereby obtaining a more comprehensive image representation. Then, in order to extract the final output from these integrated features, a fully connected (FC) module is introduced. The FC module generates a comprehensive, high-dimensional feature representation by learning the complex relationship between the input features. The calculation process of the final output can be expressed as:

[0121] X C =concat(X G ,X L )

[0122] X Out =FC(X C )

[0123] Through the above enhancements, the C2F module can more keenly capture the subtle feature information of the texture, shape, and color of the ore when processing images.

[0124] The FocalModulation module is used to replace the SPPF module on Backbone. This module uses the focal modulation mechanism to replace the traditional self-attention mechanism. The self-attention mechanism requires complex query-key interaction and query-value aggregation for each query token (QueryToken) with other tokens to calculate the attention score and capture context information. Its mathematical expression is as follows:

[0125] y i =M1(T1(x i ,X),X)

[0126] The focus modulation mechanism consists of three key components: focus contextualization, gated aggregation, and element-wise affine transformation. These components work together to effectively encode visual context and adaptively inject this information into query tokens to enhance features. The focus modulation mechanism first aggregates spatial context information into modulators at different granularities, and then injects this modulated information into query tokens based on the query dependencies. Specifically, the mathematical expression of focus modulation is as follows:

[0127] y i =q(x i )e M2(x i ,X)

[0128] Among them, M represents the aggregation process, T represents the interaction process, and q(x i ) is a query mapping function, and the Hadamard product is represented by ⊙. In the aggregation process M2, context calculations are reduced by sharing operators (such as deep convolutions, etc.), while the M1 in the self-attention mechanism is more expensive to calculate because it needs to sum the non-shared attention scores of different queries. The aggregation process M2 includes two steps:

[0129] Step 1: Hierarchical contextualization to extract contextual information at different granularity levels from local to global scope;

[0130] Step 2: All contextual features at different granularity levels are condensed into a single feature vector, the modulator, through gated aggregation.

[0131] Hierarchical contextualization: Given an input feature map X, it is first projected into a new feature space with a linear layer Z0=fz(X)€RH*W*C and then L depth-wise convolutions are used to obtain a hierarchical representation of the context. The output Zl is represented as:

[0132]

[0133] Compared with conventional convolution, it has channel characteristics and hierarchical context to generate L-level feature maps. At the lth level, the size of the effective receptive field rl is:

[0134]

[0135] rl is much larger than the convolution kernel size kl. The larger receptive field captures more global context at a coarser granularity. In order to capture the global context of the entire input, global average pooling ZL+1=AvgPool(ZL) is applied to the L-th level feature map, thereby obtaining a total (L+1) feature map They jointly capture both local and long-range contexts at different levels of granularity.

[0136] Gated Aggregation: The (L+1) feature maps obtained through hierarchical contextualization are compressed into a modulator, i.e., a single feature vector. In an image, the relationship between a token and its surrounding context often depends on the content of the query. For example, a query for a salient visual object may rely heavily on local fine-grained features, but mainly rely on global coarse-grained features for a background scene. Therefore, a gating mechanism is used to control the amount of feature maps aggregated from different levels Z l for each query. Specifically, a linear layer is used to obtain spatially and level-aware gating weights G:

[0137] G=f g (X)€R H*W*(L+1)

[0138] Then, a weighted sum is performed by element-wise multiplication to obtain a single feature map Zout of the same size as the input X:

[0139]

[0140] Among them, G l€RH*W*1 is a channel at the lth level. So far, all aggregations are spatial aggregations. In order to establish the relationship between different channels, another linear layer h() is used to obtain the modulator M:

[0141] M=h(Z out )€RH*W*C

[0142] Combining the interaction and aggregation formulas, the overall focus modulation formula can be expressed as:

[0143]

[0144] like Figure 4 As shown in the figure, the improved model based on YOLOv10 first uses CloAtt to enhance the C2F module in Backbone. In the new structure, one branch uses deep separable convolution to capture high-frequency local information and uses context-aware weights to enhance these features. The high-frequency feature information such as the local texture and shape of the molybdenum ore target and the context information are crucial for locating and identifying the target. This branch can reduce the impact of interference factors on feature extraction, while enhancing key features related to the target, and improving the accuracy of the model in distinguishing the target from the background.

[0145] The other branch obtains low-frequency global information through the attention mechanism, which facilitates the understanding of the overall layout of the image and the relationship between regions. In the molybdenum mine X-ray image, the global information helps the model to grasp the distribution of molybdenum mine targets and their relative relationship with the background as a whole, avoiding misjudgment. Finally, the information of the two branches is fused to integrate high-frequency local information and low-frequency global information.

[0146] In the global branch, in order to obtain low-frequency global information, Kglobal and Vglobal obtained by linear transformation are first downsampled, and then the traditional attention process is performed. The specific formula is:

[0147] Q global ,K global ,V global =FC(X in )

[0148] X G =Attn(Q global ,Pool(K global ),Pool(V global) ))

[0149] First, Qlocal, Klocal, and Vlocal are generated through linear transformation. Subsequently, Vlocal is processed through deep convolution (DWconv) to achieve effective aggregation of local information. This design allows the model to capture high-frequency local features in the image in detail while maintaining sensitivity to global contextual information. The specific formula is:

[0150] Q lobal ,K lobal ,V lobal =FC(X in )

[0151] V DW =DWconv(V local ).

[0152] like Figure 5-Figure 8 As shown in Figure 1, the structure and aggregation process of FocalModulation includes that the input is first processed by a linear layer, then the information is selectively aggregated through hierarchical contextualization and gated aggregation, and finally, the output is generated through the interaction between the modulator and the query;

[0153] FocalModulation replaces the original spatial pyramid pooling fast (SPPF), improving the accuracy of the network model without changing the number of parameters and the amount of calculation. In the actual mining and transportation process, the size and shape of molybdenum ore vary greatly. Based on its advanced modulation mechanism, the FocalModulation module can dynamically adjust according to the importance of different scale features in the ore image, deeply fuse the features of different scales, and thus more accurately identify molybdenum ores of different sizes and shapes, improving the model's adaptability to complex and diverse molybdenum ore samples;

[0154] In addition, the Upsample module of the original Head part is replaced with the Dysample module. DySample is an ultra-lightweight and efficient dynamic upsampler. When processing molybdenum ore X-ray images, it can reduce the amount of calculation and processing time while ensuring a certain detection performance. Using DySample as an upsampler to replace the original Upsample module can improve the running speed of the model. DySample can also reduce the model size, which is conducive to system deployment and maintenance, reduce hardware costs, and improve the overall stability and reliability of the system.

[0155] The module structure of the DySample upsampler includes a feature map X of size C*H1*W1 and a grid of size 2g*H2*W2, where the first dimension 2g represents the x and y coordinates of the original feature map X. By using the grid_sample function and the position in the point sampling set δ, the feature map X is resampled to X' of size C*H2*W2. The specific formula is as follows:

[0156] X' = ​​grid_sample(x,δ)

[0157] The generation process of the point sampling set δ based on the dynamic range factor includes, first, given the upsampling scale factor s and the feature map X of size C*H*W, using a linear layer with input and output channel numbers C and 2gs2 respectively to generate an offset O of size 2gs2*H*W. The specific formula is as follows:

[0158] O=linear(X)

[0159] Then, through pixel shufle, it is regenerated into a high-resolution original sampling grid G ​​of size 2g*sH*sW. Then the point sampling set δ is the sum of the offset O and the original sampling grid G. The specific expression is as follows:

[0160] δ=O+G

[0161] The Dysample module can significantly improve the accuracy of upsampling while ensuring computational efficiency. Through this high-precision upsampling process, the processed image can more clearly present the detailed features of molybdenum ore. These detailed information further assists the model in accurate detection and reduces the possibility of misjudgment. Finally, the GIOU loss function is used to replace the original target positioning loss term. The loss function is replaced by GIOU. The original YOLOv10 loss function consists of target positioning loss, confidence loss and classification loss:

[0162] Loss=Loss loc +Loss conf +Loss cls

[0163] The original YOLOv10 model uses CIOU to calculate the target positioning loss, and uses Binary Cross Entropy to calculate the target confidence loss and classification loss. The target positioning loss refers to the difference between the center point coordinate value of the target prediction area and the center point coordinate value of the actual object, and the difference between the width and height of the prediction box and the width and height of the actual box. IOU (Intersection Over Union) is the intersection-and-union ratio between the predicted box and the true value box, which is used to measure the overlap rate between the predicted box and the true value box. Generally speaking, the IOU value is proportional to the model prediction accuracy, but when the predicted box and the true value box do not intersect, its value is 0. At this time, it cannot correctly reflect the distance between the two, and the gradient cannot be regressed, making training impossible. In X-ray-based molybdenum ore image detection, if this happens, it is difficult for the model to accurately mark the true position of the molybdenum ore target, so IOU cannot be used directly as a loss function. Since the regression of the bounding box is related to the overlapping area, aspect ratio and center point distance of the two boxes, CIOU takes into account the consistency of the overlapping area and the distance between the center point of the bounding box and the aspect ratio, which enables CIOU to measure the similarity of the bounding box more comprehensively than IOU, thereby providing more accurate regression guidance. Its calculation formula is:

[0164] L CIOU =1-IOU(A,B)+p 2 (a,b) / c 2 +αv

[0165] Where A is the predicted box; B is the real box; a and b represent the centroids of the predicted box and the true box, respectively; p represents the Euclidean distance between the two centroids; c represents the diagonal distance of the minimum closure area surrounding the two boxes; v is used to compare the consistency of the aspect ratio, and α is the weight coefficient of v.

[0166] α and v are defined as follows:

[0167] α=v / (1-IOC+v)

[0168] v=4 / Π 2 (arctan(w gt / h gt )-arctan(w / h)) 2

[0169] Where wgt and hgt are the width and height of the true value box, w and h are the width and height of the predicted box;

[0170] The calculation formula of GIOU loss function is as follows:

[0171] G=IC / A∩B / C|

[0172] Where G is the GIOU value; A is the predicted box; B is the real box; C is the smallest box containing A and B; I is the IOU value; its loss function is:

[0173] L GIOU =lG

[0174] Since the GIOU evaluation method introduces the minimum intersection C that includes A and B, the objective function can still be optimized when A and B do not overlap. The calculation steps of the GIOU evaluation method are:

[0175] Input: predicted box A, = (x1, p, y1, p, x2, p, y2, p), real box B = (x1, g, y1, g, x2, g, y2, g); Output: LGIOU;

[0176] Calculate the area of ​​A and B:

[0177] A=(x 2,p -x 1,p )*(y 2,p -y 1,p )

[0178] B=(x 2,g -x 1,g )*(y 2,g -y 1,g )

[0179] Calculate the overlapping area Ix,y of A and B:

[0180] x1=max(x1' ,p ,x 1,g ),x2=min(x' 2,p ,x 2,g )

[0181] y1=max(y1' ,p ,y 1,g ),x2=min(y' 2,p ,y 2,g )

[0182]

[0183] Find the minimum area D of A∩B:

[0184] x 1,c =max(x1' ,p ,x 1,g ),x 2,c =min(x' 2,p ,x 2,g )

[0185] y 1,c =max(y1' ,p ,y 1,g ),x 2,c =min(y' 2,p ,y 2,g )

[0186] D=(x 2,c -x 1,c )*(y 2,c -y 1,c )

[0187] Calculate IOU:

[0188] I=I x,y / (A+BI x,y )

[0189] Calculate GIOU:

[0190] G=I-(B c -(A+BI x,y )) / D.

[0191] Calculate the final loss:

[0192] L GIOU =1-G.

[0193] X-ray sorting technology includes the following specific steps:

[0194] Step 1: Ore pretreatment: In the initial stage of the entire molybdenum ore resource detection and sorting process, the ore needs to be pretreated comprehensively and carefully. This step mainly includes crushing, screening and cleaning:

[0195] Crushing: The raw ore usually exists in the form of large blocks, which is not conducive to the subsequent X-ray penetration and sorting operations. The raw ore is crushed by crushing equipment to gradually reduce its particle size, making it easier for subsequent processing. The degree of crushing needs to be reasonably adjusted according to the specific requirements of the sorting equipment and the characteristics of the ore. The particle size of the crushed ore should be controlled within a certain range, which can ensure that the X-ray can penetrate effectively and facilitate the subsequent screening operation;

[0196] Screening operation: The particle size of the crushed ore is not uniform, so screening operation is required. By setting screens with different apertures, the crushed ore is screened according to the particle size, and the part with too large or too small particle size is removed. Only the ore with the particle size meeting the requirements can enter the subsequent X-ray irradiation link, which can ensure that the X-ray can obtain a more consistent penetration effect when penetrating the ore, thereby improving the accuracy of detection;

[0197] Cleaning stage: During the mining and transportation of raw ore, some impurities such as soil and dust often adhere to the surface of the ore. These impurities will affect the penetration effect of X-rays and subsequent detection results. Therefore, after screening, the ore needs to be cleaned by washing it with clean water or cleaning solution to thoroughly remove impurities on its surface to ensure that the ore reaches the particle size range and purity requirements suitable for subsequent X-ray sorting;

[0198] Step 2: X-ray penetration and detection: The pre-treated ore is sent to the X-ray sorting equipment and irradiated with X-rays. Since X-rays have strong penetrating power, most of their energy can pass through the gaps between the atoms of the ore. However, the complexity of the internal structure of the ore and the difference in density will cause X-rays to encounter different degrees of obstruction during the penetration process. The distribution of different mineral components in the ore and the difference in crystal structure will lead to different degrees of difficulty in X-ray penetration. The detector on the other side of the ore can accurately detect the slight differences in the degree of X-ray penetration. The detector uses a highly sensitive detection element and can accurately measure the intensity changes of X-rays. When X-rays penetrate the ore, their intensity will change. The detector can accurately capture the slight differences in the degree of X-ray penetration based on these changes and convert them into electrical signals for analysis. These electrical signals carry important information about the internal structure and density differences of the ore, and are the basic data source for subsequent image processing and target detection.

[0199] Step 3: Image processing and target detection: The electrical signal from the detector is transmitted to the image processing unit. Here, the electrical signal needs to go through the following processing to be converted into an image form that can be analyzed. First, filtering is performed. By using a suitable filtering algorithm, the noise interference in the electrical signal is removed to make the electrical signal purer. Then, enhancement processing is performed. By adjusting the intensity, contrast and other parameters of the electrical signal, the converted image can more clearly present the internal structure and density difference of the ore. After these processes, the electrical signal is finally converted into a grayscale image that can present the internal structure and density difference of the ore. Then, the improved YOL is used. The target detection algorithm in the Ov10 model analyzes the grayscale image. The model first extracts and analyzes the features of each pixel in the image, and uses the improved modules to work together to fully mine the feature information in the image. Then, based on these feature information, the target detection algorithm distinguishes between concentrates and tailings, and accurately marks the position of each metal ore on the conveyor belt. When processing a grayscale image containing multiple molybdenum ores, the model can accurately identify which ores are concentrates and which are tailings, and give the precise coordinate position of each ore in the image. These position information will be directly applied to the subsequent sorting execution link;

[0200] Step 4: Sorting execution: Based on the results of the improved YOLOv10 model marking the ore position and category, when the ore is sent to the actuator through the conveyor belt, the actuator is controlled to spray high-pressure gas, and the ore marked as concentrate is sprayed into the concentrate bin, while the ore marked as tailings falls directly into the tailings bin. In this process, the actuator uses a high-precision gas injection device, which can accurately control the injection direction and intensity of the high-pressure gas according to the ore position and category information marked by the model, ensuring that only the ore marked as concentrate will be sprayed into the concentrate bin, and the tailings will fall into the tailings bin accurately.

[0201] The above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention.

Claims

1. Based on the improved YOLOv10, the automatic detection method and system for molybdenum ore resources are characterized by: Including the improved model based on YOLOv10 and X-ray sorting technology. As the front-end detection method, X-ray sorting technology is responsible for the preliminary physical property detection of molybdenum ore and converting the detection results into data for subsequent model analysis; The improved model based on YOLOv10 serves as the intelligent analysis core of the back-end, and conducts in-depth mining and precise judgment on the data transmitted from the front-end, thereby realizing the accurate classification and positioning of molybdenum ore, providing a reliable basis for the final sorting operation.

2. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 1, characterized in that: The improved model based on YOLOv10 first uses CloAtt to enhance the C2F module in Backbone. In the new structure, a branch uses deep separable convolution to capture high-frequency local information and uses context-aware weights to enhance these features. The local texture and shape high-frequency feature information of the molybdenum mine target, as well as the context information, are crucial for locating and identifying the target. This branch can reduce the impact of interference factors on feature extraction, while enhancing key features related to the target, and improving the accuracy of the model in distinguishing the target from the background. The other branch obtains low-frequency global information through the attention mechanism to understand the overall layout of the image and the relationship between regions. In the molybdenum mine X-ray image, the global information helps the model to grasp the distribution of molybdenum mine targets and their relative relationship with the background as a whole, avoiding misjudgment. Finally, the information of the two branches is fused to integrate high-frequency local information and low-frequency global information. In the global branch, in order to obtain low-frequency global information, Kglobal and Vglobal obtained by linear transformation are first downsampled, and then the traditional attention process is performed. The specific formula is: Q global ,K global ,V global =FC(X in ) X G =Attn(Q global ,Pool(K global ),Pool(V global) )) First, Qlocal, Klocal and Vlocal are generated through linear transformation. Subsequently, Vlocal is processed by deep convolution (DWconv) to achieve effective aggregation of local information. This design allows the model to capture high-frequency local features in the image in detail while maintaining sensitivity to global contextual information. The specific formula is: Q lobal ,K lobal ,V lobal =FC(X in ) V DW =DWconv(V local )。 3. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 2, characterized in that: In order to further improve the model's ability to process local features, two depth-wise separable convolutions (DW-Convs) are first used to aggregate Qlocal and Klocal respectively. In order to obtain context-aware weights between -1 and 1, the Hadamard product of Qlocal and Klocal is transformed, including normalization and the application of activation functions. Finally, by applying these context-aware weights to local features, the model's ability to capture local details is further enhanced, thereby improving the detection accuracy of targets in molybdenum mine X-ray images. The detailed steps of the whole process are as follows: Q L =DWconv(Q local ) K L =DWconv(K local ) A L =FC(Swish(FC(Q L e K L ))) X L =Attne V DW After completing the above local feature enhancement steps, the features extracted by the two branches are further integrated by connecting their outputs in the channel dimension. This connection operation enables the model to consider both high-frequency local information and low-frequency global information, thereby obtaining a more comprehensive image representation. Then, in order to extract the final output from these integrated features, a fully connected (FC) module is introduced. The FC module generates a comprehensive, high-dimensional feature representation by learning the complex relationship between the input features. The calculation process of the final output is expressed as: X C =concat(X G ,X L ) X Out =FC(X C ) Through the above enhancements, the C2F module can more keenly capture the subtle feature information of the texture, shape, and color of the ore when processing images.

4. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 3, characterized in that: The FocalModulation module is used to replace the SPPF module on Backbone. This module uses the focal modulation mechanism to replace the traditional self-attention mechanism. The self-attention mechanism requires complex query-key interaction and query-value aggregation for each query token with other tokens to calculate the attention score and capture context information. Its mathematical expression is as follows: the i =M1(T1(x i ,X),X) The focus modulation mechanism consists of three key components: focus contextualization, gated aggregation, and element-wise affine transformation. These components work together to effectively encode visual context and adaptively inject this information into query tokens to enhance features. The focus modulation mechanism first aggregates spatial context information into modulators at different granularities, and then injects this modulated information into query tokens based on the query dependencies. Specifically, the mathematical expression of focus modulation is as follows: y i =q(x i )e M2(x i ,X) Among them, M represents the aggregation process, T represents the interaction process, and q(x i ) is a query mapping function, and the Hadamard product is represented by ⊙. In the aggregation process M2, context calculation is reduced by sharing operators (such as deep convolution), while M1 in the self-attention mechanism is more expensive because it needs to sum the non-shared attention scores of different queries. The aggregation process M2 includes two steps: Step 1: Hierarchical contextualization to extract contextual information at different granularity levels from local to global scope; Step 2: All contextual features at different granularity levels are condensed into a single feature vector, the modulator, through gated aggregation.

5. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 4, characterized in that: Hierarchical contextualization: Given an input feature map X, it is first projected into a new feature space with a linear layer Z0=fz(X)€RH*W*C and then L depth-wise convolutions are used to obtain a hierarchical representation of the context. The output Zl is represented as: Compared with conventional convolution, it has channel characteristics and hierarchical context to generate L-level feature maps. At the lth level, the size of the effective receptive field rl is: rl is much larger than the convolution kernel size kl. The larger receptive field captures more global context at a coarser granularity. In order to capture the global context of the entire input, global average pooling ZL+1=AvgPool(ZL) is applied to the L-th level feature map, thereby obtaining a total (L+1) feature map They jointly capture both local and long-range contexts at different levels of granularity.

6. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 1, characterized in that: Gated Aggregation: The (L+1) feature maps obtained through hierarchical contextualization are compressed into a modulator, i.e., a single feature vector. In an image, the relationship between a token and its surrounding context usually depends on the content of the query. For example, queries that heavily rely on local fine-grained features to encode salient visual objects, but mainly rely on global coarse-grained features to query background scenes, therefore, a gating mechanism is used to control the amount of feature maps aggregated from different levels Zl for each query. Specifically, a linear layer is used to obtain the spatially and level-aware gating weights G: G=f g (X)€R H*W*(L+1) Then, a weighted sum is performed by element-wise multiplication to obtain a single feature map Zout of the same size as the input X: Among them, Gl€RH*W*1 is a channel at level l. So far, all aggregations are spatial aggregations. In order to establish the relationship between different channels, another linear layer h() is used to obtain the modulator M: M=h(Z out )€R H*W*C Combining the interaction and aggregation formulas, the overall focus modulation formula is expressed as:

7. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 6, characterized in that: The structure and aggregation process of FocalModulation includes that the input is first processed through a linear layer, then the information is selectively aggregated through hierarchical contextualization and gated aggregation, and finally, the output is generated through the interaction between the modulator and the query; FocalModulation replaces the original spatial pyramid pooling fast (SPPF), improving the accuracy of the network model without changing the number of parameters and the amount of calculation. In the actual mining and transportation process, the size and shape of molybdenum ore vary greatly. Based on its advanced modulation mechanism, the FocalModulation module can dynamically adjust according to the importance of different scale features in the ore image, deeply fuse the features of different scales, and thus more accurately identify molybdenum ores of different sizes and shapes, improving the model's adaptability to complex and diverse molybdenum ore samples; In addition, the Upsample module of the original Head part is replaced with the Dysample module. DySample is an ultra-lightweight and efficient dynamic upsampler. When processing molybdenum ore X-ray images, it can reduce the amount of calculation and processing time while ensuring a certain detection performance. Using DySample as an upsampler to replace the original Upsample module can improve the running speed of the model. DySample can also reduce the model size, which is conducive to system deployment and maintenance, reduce hardware costs, and improve the overall stability and reliability of the system. The module structure of the DySample upsampler includes a feature map X of size C*H1*W1 and a grid of size 2g*H2*W2, where the first dimension 2g represents the x and y coordinates of the original feature map X. By using the grid_sample function and the position in the point sampling set δ, the feature map X is resampled to X' of size C*H2*W2. The specific formula is as follows: X' = ​​grid_sample(x,δ) The generation process of the point sampling set δ based on the dynamic range factor includes, first, given the upsampling scale factor s and the feature map X of size C*H*W, using a linear layer with input and output channel numbers C and 2gs2 respectively to generate an offset O of size 2gs2*H*W. The specific formula is as follows: O=linear(X) Then, through pixel shufle, it is regenerated into a high-resolution original sampling grid G ​​of size 2g*sH*sW. Then the point sampling set δ is the sum of the offset O and the original sampling grid G. The specific expression is as follows: δ=O+G The Dysample module can significantly improve the accuracy of upsampling while ensuring computational efficiency. Through this high-precision upsampling process, the processed image can more clearly present the detailed features of molybdenum ore. These detailed information further assists the model in accurate detection and reduces the possibility of misjudgment. Finally, the GIOU loss function is used to replace the original target positioning loss term. The loss function is replaced by GIOU. The original YOLOv10 loss function consists of target positioning loss, confidence loss and classification loss. composition: Loss=Loss loc +Loss conf +Loss cls The original YOLOv10 model uses CIOU to calculate the target positioning loss, and uses Binary Cross Entropy to calculate the target confidence loss and classification loss. The target positioning loss refers to the difference between the center point coordinate value of the target prediction area and the center point coordinate value of the actual object, and the difference between the width and height of the prediction box and the width and height of the actual box. IOU (Intersection Over Union) is the intersection-and-union ratio between the predicted box and the true value box, which is used to measure the overlap rate between the predicted box and the true value box. Generally speaking, the IOU value is proportional to the model prediction accuracy, but when the predicted box and the true value box do not intersect, its value is 0. At this time, it cannot correctly reflect the distance between the two, and the gradient cannot be regressed, making training impossible. In X-ray-based molybdenum ore image detection, if this happens, it is difficult for the model to accurately mark the true position of the molybdenum ore target, so IOU cannot be used directly as a loss function. Since the regression of the bounding box is related to the overlapping area, aspect ratio and center point distance of the two boxes, CIOU takes into account the consistency of the overlapping area and the distance between the center point of the bounding box and the aspect ratio, which enables CIOU to measure the similarity of the bounding box more comprehensively than IOU, thereby providing more accurate regression guidance. Its calculation formula is: L CIOU =1-IOU(A,B)+p 2 (a,b) / c 2 +αv Where A is the predicted box; B is the real box; a and b represent the centroids of the predicted box and the real box respectively; p represents the Euclidean distance between the two centroids; c represents the diagonal distance of the minimum closure area surrounding the two boxes; v is used to compare the consistency of the aspect ratio, α is the weight coefficient of v, and α and v are defined as follows: α=v / (1-IOC+v) v=4 / Π 2 (arctan(w gt / h gt )-arctan(w / h)) 2 Where wgt and hgt are the width and height of the true value box, w and h are the width and height of the predicted box; The calculation formula of GIOU loss function is as follows: G=I-|C / A∩B| / |C Where G is the GIOU value; A is the predicted box; B is the real box; C is the smallest box containing A and B; I is the IOU value; its loss function is: L GIOU =l-G Since the GIOU evaluation method introduces the minimum intersection C containing A and B, the objective function is still optimized when A and B do not overlap. The calculation steps of the GIOU evaluation method are: Input: predicted box A, = (x1, p, y1, p, x2, p, y2, p), real box B = (x1, g, y1, g, x2, g, y2, g); Output: LGIOU; Calculate the area of ​​A and B: A=(x 2,p -x 1,p )*(and 2,p -and 1,p ) B=(x 2,g -x 1,g )*(y 2,g -y 1,g ) Calculate the overlapping area Ix,y of A and B: x1=max(x1' ,p ,x 1,g ),x2=min(x' 2,p ,x 2,g ) y1=max(y1' ,p ,y 1,g ),x2=min(y' 2,p ,y 2,g ) Find the minimum area D of A∩B: x 1,c =max(x1' ,p ,x 1,g ),x 2,c =min(x' 2,p ,x 2,g ) and 1,c =max(y1' ,p ,and 1,g ),x 2,c =min(y' 2,p ,and 2,g ) D=(x 2,c -x 1,c )*(y 2,c -y 1,c ) Calculate IOU: I=I x,y / (A+B-I x,y ) Calculate GIOU: G=I-(B c -(A+B-I x,y )) / D, Calculate the final loss: L GIOU =1-G。 8. The method and system for automatically detecting molybdenum ore resources based on improved YOLOv10 according to claim 7, characterized in that: The X-ray sorting technology includes the following specific steps: Step 1: Ore pretreatment: In the initial stage of the entire molybdenum ore resource detection and sorting process, the ore needs to be pretreated comprehensively and carefully, which mainly includes crushing, screening and cleaning: Crushing: The raw ore usually exists in the form of large blocks, which is not conducive to the subsequent X-ray penetration and sorting operations. The raw ore is crushed by crushing equipment to gradually reduce its particle size, making it easier for subsequent processing. The degree of crushing needs to be reasonably adjusted according to the specific requirements of the sorting equipment and the characteristics of the ore. The particle size of the crushed ore should be controlled within a certain range, which can ensure that the X-ray can penetrate effectively and facilitate the subsequent screening operation; Screening operation: The particle size of the crushed ore is not uniform, so screening operation is required. By setting screens with different apertures, the crushed ore is screened according to the particle size, and the part with too large or too small particle size is removed. Only the ore with the particle size meeting the requirements can enter the subsequent X-ray irradiation link, so as to ensure that the X-ray can obtain a more consistent penetration effect when penetrating the ore, and improve the accuracy of detection; Cleaning stage: During the mining and transportation of raw ore, some impurities such as soil and dust often adhere to the surface of the ore. These impurities will affect the penetration effect of X-rays and subsequent detection results. Therefore, after screening, the ore needs to be cleaned by washing it with clean water or cleaning solution to thoroughly remove impurities on its surface to ensure that the ore reaches the particle size range and purity requirements suitable for subsequent X-ray sorting; Step 2: X-ray penetration and detection: The pre-treated ore is sent to the X-ray sorting equipment and irradiated with X-rays. Since X-rays have strong penetrating power, most of their energy can pass through the gaps between the atoms of the ore. However, the complexity of the internal structure of the ore and the difference in density will cause X-rays to encounter different degrees of obstruction during the penetration process. The distribution of different mineral components in the ore and the difference in crystal structure will lead to different degrees of difficulty in X-ray penetration. The detector on the other side of the ore can accurately detect the slight differences in the degree of X-ray penetration. The detector uses a highly sensitive detection element and can accurately measure the intensity changes of X-rays. When X-rays penetrate the ore, their intensity will change. The detector can accurately capture the slight differences in the degree of X-ray penetration based on these changes and convert them into electrical signals for analysis. These electrical signals carry important information about the internal structure and density differences of the ore, and are the basic data source for subsequent image processing and target detection. Step 3: Image processing and target detection: The electrical signal from the detector is transmitted to the image processing unit. Here, the electrical signal needs to go through the following processing to be converted into an image form that can be analyzed. First, filtering is performed. By using a suitable filtering algorithm, the noise interference in the electrical signal is removed to make the electrical signal purer. Then, enhancement processing is performed. By adjusting the intensity and contrast parameters of the electrical signal, the converted image can more clearly present the internal structure and density difference of the ore. After these processes, the electrical signal is finally converted into a grayscale image that can present the internal structure and density difference of the ore. Then, the improved YOLO is used. The target detection algorithm in the v10 model analyzes the grayscale image. The model first extracts and analyzes the features of each pixel in the image, and uses the improved modules to work together to fully mine the feature information in the image. Then, based on these feature information, the target detection algorithm distinguishes between concentrates and tailings, and accurately marks the position of each metal ore on the conveyor belt. When processing a grayscale image containing multiple molybdenum ores, the model can accurately identify which ores are concentrates and which are tailings, and give the precise coordinate position of each ore in the image. These position information will be directly applied to the subsequent sorting execution link; Step 4: Sorting execution: Based on the results of the improved YOLOv10 model marking the ore position and category, when the ore is sent to the actuator through the conveyor belt, the actuator is controlled to spray high-pressure gas, and the ore marked as concentrate is sprayed into the concentrate bin, while the ore marked as tailings falls directly into the tailings bin. In this process, the actuator uses a high-precision gas injection device, which can accurately control the injection direction and intensity of the high-pressure gas according to the ore position and category information marked by the model, ensuring that only the ore marked as concentrate will be sprayed into the concentrate bin, and the tailings will fall into the tailings bin accurately.

Images