Artificial intelligence sorting method and device based on multi-image fusion

Abstract

The invention discloses an artificial intelligence sorting method and device based on multi-image fusion, and relates to the field of ore sorting. Acquiring a first wave band image, a second wave band image and an X-ray image of the target ore; the first waveband image is a visible light image; the second wave band image is a near-infrared image; scaling the first wave band image, the second wave band image and the X-ray image to obtain a multi-channel feature map; the multi-channel feature map serves as input of a pre-trained classification model, and the type of the target ore is obtained; and according to the type of the target ore, the target ore is conveyed to the corresponding receiving area. According to the method, the visible light image, the near-infrared image and the X-ray image are comprehensively analyzed and processed through the artificial intelligence classification model, multi-dimensional information can be fused, and therefore tiny differences of ores can be accurately distinguished, accurate sorting can be achieved even under the condition that the ore components are highly similar, and the ore sorting efficiency is improved. And the ore sorting accuracy and efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ore sorting, and in particular to an artificial intelligence sorting method and device based on multi-image fusion. Background Art

[0002] Ore sorting is a key link in mining production. Its main purpose is to separate valuable minerals from waste rocks or other impurities in the ore by physical or chemical methods to improve the grade of the ore and provide high-quality raw materials for subsequent smelting or further processing. The efficiency and accuracy of ore sorting directly affect the utilization rate of ore resources, production costs and environmental impact.

[0003] Traditional ore color sorters usually rely on a single spectrum (such as visible light) for sorting. Patent CN108108678B discloses a tungsten ore identification and sorting method, which can automatically identify and sort ore by collecting and processing visible light (RGB) images of ore. Compared with manual sorting, it is faster, more efficient, less costly and has a higher sorting rate.

[0004] Although this method is suitable for ores with obvious color differences, the recognition accuracy is low when the ores have similar colors or optical properties. For some ores with similar colors, surface textures or optical properties, the resolution of a single spectrum is limited, resulting in common misselection or omission, and low sorting accuracy. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of low ore sorting accuracy mentioned in the above background technology, and to propose an artificial intelligence sorting method and device based on multi-image fusion.

[0006] The first aspect of the present invention provides an artificial intelligence sorting method based on multi-image fusion, the method comprising:

[0007] Acquire a first band image, a second band image and an X-ray image of the target ore; the first band image is a visible light image; the second band image is a near infrared image;

[0008] Scaling the first waveband image, the second waveband image and the X-ray image to obtain a multi-channel feature map;

[0009] Using the multi-channel feature map as an input of a pre-trained classification model to obtain the type of target ore;

[0010] According to the type of the target ore, the target ore is transported to a corresponding receiving area.

[0011] Optionally, the first band image and the second band image of the target ore are collected by a first collection point; and the X-ray image is collected by a second collection point;

[0012] The obtaining of the first waveband image, the second waveband image and the X-ray image of the target ore comprises:

[0013] Acquire visible light images and near infrared images of the target area;

[0014] Performing edge detection and image segmentation on the visible light image of the target area to determine the spatial position of the target ore and obtain a first band image of the target ore;

[0015] Extracting a second band image of the target ore from the near infrared image of the target area according to the spatial position of the target ore;

[0016] Obtaining a time difference between the first band image and the X-ray image according to the position distance between the first acquisition point and the second acquisition point and the moving speed of the conveyor belt;

[0017] An X-ray image of the target ore is obtained according to the acquisition time difference and the spatial position of the target ore.

[0018] Optionally, scaling the first band image, the second band image, and the X-ray image to obtain a multi-channel feature map includes:

[0019] According to a bilinear interpolation algorithm, scaling the first waveband image, the second waveband image and the X-ray image to a preset size;

[0020] The scaled first band image, second band image and X-ray image are spliced ​​to obtain a multi-channel feature map.

[0021] Optionally, the classification model includes a feature extraction layer, a feature fusion layer and a fully connected layer; the feature extraction layer includes a first convolution branch, a second convolution branch, and a third convolution branch; the multi-channel feature map includes an RGB three-channel layer, a near-infrared single-channel layer and an X-ray single-channel layer;

[0022] in:

[0023] The first convolution branch is used to receive the RGB three-channel layer as input and output a first feature tensor;

[0024] The second convolution branch is used to receive the near-infrared single-channel layer as input and output a second feature tensor;

[0025] The third convolution branch is used to receive the X-ray single channel layer as input and output a third feature tensor;

[0026] The feature fusion layer is used to receive the first feature tensor, the second feature tensor and the third feature tensor as input, and output a global feature vector;

[0027] The fully connected layer is used to receive the global feature vector as input and output the type of target ore.

[0028] Optionally, the feature fusion layer receives the first feature tensor, the second feature tensor, and the third feature tensor as input, and outputs a global feature vector including:

[0029] Concatenate the first feature tensor and the second feature tensor in the channel dimension to obtain a fourth feature tensor;

[0030] Using a preset number of 1×1 convolution kernels to perform convolution operations on the fourth feature tensor, respectively, to obtain a fifth feature tensor;

[0031] Performing a nonlinear transformation on the fifth eigentensor using a ReLu function to obtain a sixth eigentensor, and expanding the sixth eigentensor into a one-dimensional first eigenvector;

[0032] Opening the third eigentensor into a one-dimensional second eigenvector;

[0033] The first feature vector and the second feature vector are concatenated to obtain a global feature vector.

[0034] A second aspect of the present invention provides an artificial intelligence sorting device based on multi-image fusion, the device comprising:

[0035] An image acquisition module, used to acquire a first band image, a second band image and an X-ray image of the target ore; the first band image is a visible light image; the second band image is a near infrared image;

[0036] An image standardization module, used for scaling the first band image, the second band image and the X-ray image to obtain a multi-channel feature map;

[0037] An ore classification module, used to use the multi-channel feature map as an input of a pre-trained classification model to obtain the type of target ore;

[0038] The separation module is used to transport the target ore to a corresponding receiving area according to the type of the target ore.

[0039] Optionally, the first band image and the second band image of the target ore are collected by a first collection point; and the X-ray image is collected by a second collection point;

[0040] The image acquisition module comprises:

[0041] A regional image acquisition module is used to acquire visible light images and near infrared images of the target area;

[0042] A first band image acquisition module, used for performing edge detection and image segmentation on the visible light image of the target area, determining the spatial position of the target ore, and obtaining a first band image of the target ore;

[0043] A second band image acquisition module, used to extract a second band image of the target ore from the near infrared image of the target area according to the spatial position of the target ore;

[0044] A time difference determination module, used to obtain the acquisition time difference between the first band image and the X-ray image according to the position distance between the first acquisition point and the second acquisition point and the moving speed of the conveyor belt;

[0045] The X-ray image acquisition module is used to obtain the X-ray image of the target ore according to the acquisition time difference and the spatial position of the target ore.

[0046] Optionally, the image standardization module includes:

[0047] An image scaling module, used for scaling the first waveband image, the second waveband image and the X-ray image to a preset size according to a bilinear interpolation algorithm;

[0048] The image stitching module is used to stitch the scaled first band image, the second band image and the X-ray image to obtain a multi-channel feature map.

[0049] Optionally, the classification model includes a feature extraction layer, a feature fusion layer and a fully connected layer; the feature extraction layer includes a first convolution branch, a second convolution branch, and a third convolution branch; the multi-channel feature map includes an RGB three-channel layer, a near-infrared single-channel layer and an X-ray single-channel layer;

[0050] in:

[0051] The first convolution branch is used to receive the RGB three-channel layer as input and output a first feature tensor;

[0052] The second convolution branch is used to receive the near-infrared single-channel layer as input and output a second feature tensor;

[0053] The third convolution branch is used to receive the X-ray single channel layer as input and output a third feature tensor;

[0054] The feature fusion layer is used to receive the first feature tensor, the second feature tensor and the third feature tensor as input, and output a global feature vector;

[0055] The fully connected layer is used to receive the global feature vector as input and output the type of target ore.

[0056] Optionally, the feature fusion layer is specifically used for:

[0057] Concatenate the first feature tensor and the second feature tensor in the channel dimension to obtain a fourth feature tensor;

[0058] Using a preset number of 1×1 convolution kernels to perform convolution operations on the fourth feature tensor, respectively, to obtain a fifth feature tensor;

[0059] Performing a nonlinear transformation on the fifth eigentensor using a ReLu function to obtain a sixth eigentensor, and expanding the sixth eigentensor into a one-dimensional first eigenvector;

[0060] Opening the third eigentensor into a one-dimensional second eigenvector;

[0061] The first feature vector and the second feature vector are concatenated to obtain a global feature vector.

[0062] Beneficial effects of the present invention:

[0063] The present invention proposes an artificial intelligence sorting method based on multi-image fusion, which includes: acquiring a first band image, a second band image and an X-ray image of a target ore; the first band image is a visible light image; the second band image is a near-infrared image; scaling the first band image, the second band image and the X-ray image to obtain a multi-channel feature map; using the multi-channel feature map as an input of a pre-trained classification model to obtain the type of the target ore; and transporting the target ore to a corresponding receiving area according to the type of the target ore.

[0064] By combining visible light images, near-infrared images and X-ray images, this method can comprehensively utilize ore information in different bands, and then perform comprehensive analysis and processing through an artificial intelligence classification model. It can fuse multi-dimensional information, thereby accurately distinguishing slight differences in ores. Even when the ore compositions are highly similar, accurate sorting can be achieved, significantly improving the accuracy and efficiency of ore sorting. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The present invention will be further described below in conjunction with the accompanying drawings.

[0066] Figure 1 A flowchart of an artificial intelligence sorting method based on multi-image fusion is provided for an embodiment of the present invention;

[0067] Figure 2A structural diagram of an artificial intelligence sorting device based on multi-image fusion is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The embodiment of the present invention provides an artificial intelligence sorting method based on multi-image fusion. Figure 1 , Figure 1 A flowchart of an artificial intelligence sorting method based on multi-image fusion is provided for an embodiment of the present invention. The method comprises the following steps:

[0070] S101, acquiring a first band image, a second band image and an X-ray image of a target ore.

[0071] S102, scaling the first band image, the second band image and the X-ray image to obtain a multi-channel feature map.

[0072] S103, using the multi-channel feature map as an input of a pre-trained classification model to obtain the type of the target ore.

[0073] S104, transporting the target ore to a corresponding receiving area according to the type of the target ore.

[0074] Among them, the first band image is a visible light image; the second band image is a near-infrared image.

[0075] An artificial intelligence sorting method based on multi-image fusion provided by an embodiment of the present invention can comprehensively utilize ore information in different bands by combining visible light images, near-infrared images and X-ray images, and then perform comprehensive analysis and processing through an artificial intelligence classification model. It can fuse multi-dimensional information, thereby accurately distinguishing slight differences in ores. Even when the ore compositions are highly similar, accurate sorting can be achieved, significantly improving the accuracy and efficiency of ore sorting.

[0076] In one implementation, the method can be used to distinguish different types of iron ores such as magnetite, hematite, and limonite; it can also be used to identify the iron content of ore and distinguish high-grade from low-grade ore. The method is highly adaptable and can select the corresponding training method to train the classification model based on different application scenarios to achieve efficient ore sorting.

[0077] In one embodiment, the first band image and the second band image of the target ore are collected by the first collection point; the X-ray image is collected by the second collection point; step S101 includes:

[0078] Step 1: Acquire visible light images and near infrared images of the target area.

[0079] Step 2: Perform edge detection and image segmentation on the visible light image of the target area to determine the spatial position of the target ore and obtain the first band image of the target ore.

[0080] Step three, according to the spatial position of the target ore, extract the second band image of the target ore from the near infrared image of the target area.

[0081] Step 4: Obtain the acquisition time difference between the first waveband image and the X-ray image according to the position distance between the first acquisition point and the second acquisition point and the moving speed of the conveyor belt.

[0082] Step 5: Obtain an X-ray image of the target ore based on the acquisition time difference and the spatial position of the target ore.

[0083] In one implementation, the ore image can be extracted from the entire area image through edge detection and image segmentation, and its bounding rectangle can be generated. The center position coordinates, length, and width triples of the bounding rectangle are used as the spatial position of the target ore. The near-infrared image and the visible light image are collected at the same acquisition point at the same time, and they correspond one-to-one at the pixel level, providing a reliable basis for subsequent fusion.

[0084] In one implementation, the ore on the conveyor belt may first pass through the second collection point and then the first collection point. For example, if the collection time difference between the second collection point and the first collection point is 5 seconds, then after the visible light image of the target ore is collected at the first collection point, the image information collected 5 seconds ago at the second collection point can be obtained according to the timestamp information, and the ore can be selected according to the spatial position of the ore to obtain the X-ray image of the ore.

[0085] In one embodiment, step S102 includes:

[0086] Step 1: scaling the first band image, the second band image and the X-ray image to a preset size according to a bilinear interpolation algorithm;

[0087] Step 2: splice the scaled first band image, second band image and X-ray image to obtain a multi-channel feature map.

[0088] In one implementation, all images are scaled to a preset size using a bilinear interpolation algorithm to ensure that their spatial resolutions are consistent. This uniform size processing facilitates subsequent feature extraction and image fusion, and avoids computational complexity or errors introduced by inconsistent image sizes.

[0089] In one embodiment, the classification model includes a feature extraction layer, a feature fusion layer and a fully connected layer; the feature extraction layer includes a first convolution branch, a second convolution branch, and a third convolution branch; the multi-channel feature map includes an RGB three-channel layer, a near-infrared single-channel layer and an X-ray single-channel layer; wherein:

[0090] The first convolution branch is used to receive the RGB three-channel layer as input and output the first feature tensor.

[0091] The second convolution branch is used to receive the near-infrared single channel layer as input and output the second feature tensor.

[0092] The third convolution branch is used to receive the X-ray single channel layer as input and output the third feature tensor.

[0093] The feature fusion layer is used to receive the first feature tensor, the second feature tensor and the third feature tensor as input, and output a global feature vector.

[0094] The fully connected layer is used to receive the global feature vector as input and output the type of target ore.

[0095] In one implementation, each convolution branch is specialized for feature extraction of different images, which can fully capture the unique information contained in each image. For example, the RGB channel layer can capture the color and texture information of the ore, the near-infrared layer can reveal the internal characteristics of the ore, and the X-ray layer can provide the density and structure information of the ore.

[0096] In one implementation, the feature tensors from different convolution branches are effectively combined to form a global feature vector. This fusion method not only maintains the unique information of each channel, but also enhances the correlation between them, enabling the model to extract richer features from multimodal data and improve the ability to identify complex characteristics of ores.

[0097] In one implementation, by fusing image information from multiple bands, the model can effectively reduce the error caused by noise from a single data source and enhance the robustness of the model.

[0098] In one embodiment, the first convolution branch and the second convolution branch have the same convolution layer structure, but only have different weights; the feature fusion layer receives the first feature tensor, the second feature tensor, and the third feature tensor as input, and outputs the global feature vector including the following steps:

[0099] Step 1: Concatenate the first feature tensor and the second feature tensor in the channel dimension to obtain a fourth feature tensor.

[0100] Step 2: Use a preset number of 1×1 convolution kernels to perform a convolution operation on the fourth feature tensor to obtain a fifth feature tensor.

[0101] Step three, use the ReLu function to perform nonlinear transformation on the fifth eigentensor to obtain the sixth eigentensor.

[0102] Step 4: Expand the sixth eigentensor into a one-dimensional first eigenvector; expand the third eigentensor into a one-dimensional second eigenvector.

[0103] Step 5: Concatenate the first eigenvector and the second eigenvector to obtain a global eigenvector.

[0104] In one implementation, the visible light image and the near-infrared image correspond one-to-one at the pixel level, and the first convolution branch and the second convolution branch have the same convolution layer structure, which means that they are consistent in the way of feature extraction. This design ensures the uniformity of the feature extraction process and facilitates effective fusion.

[0105] In one implementation, two tensors are concatenated in the channel dimension, which means stacking of channels. For example, if the size of the first feature tensor is 64×64×16 and the size of the second feature tensor is 64×64×16, the size of the concatenated fourth feature tensor is 64×64×32. The concatenation of two vectors means an increase in the vector dimension. For example, if the first feature vector is a 32-dimensional column vector and the second feature vector is a 32-dimensional column vector, the concatenated global feature vector is a 64-dimensional column vector.

[0106] In one implementation, a 1×1 convolution kernel helps compress the feature dimension while fusing feature information from different modalities, avoiding the computational complexity problem caused by excessively high feature dimensions.

[0107] The embodiment of the present invention provides an artificial intelligence sorting device based on multi-image fusion. Figure 2 , Figure 2 The present invention provides a structural diagram of an artificial intelligence sorting device based on multi-image fusion. The device includes:

[0108] The image acquisition module is used to acquire the first band image, the second band image and the X-ray image of the target ore.

[0109] The image normalization module is used to scale the first band image, the second band image and the X-ray image to obtain a multi-channel feature map.

[0110] The ore classification module is used to use the multi-channel feature map as the input of the pre-trained classification model to obtain the type of target ore.

[0111] The separation module is used to transport the target ore to the corresponding receiving area according to the type of the target ore.

[0112] Among them, the first band image is a visible light image; the second band image is a near-infrared image.

[0113] An artificial intelligence sorting device based on multi-image fusion provided by an embodiment of the present invention can comprehensively utilize ore information in different bands by combining visible light images, near-infrared images and X-ray images, and then perform comprehensive analysis and processing through an artificial intelligence classification model. It can fuse multi-dimensional information, thereby accurately distinguishing slight differences in ores. Even when the ore compositions are highly similar, accurate sorting can be achieved, which significantly improves the accuracy and efficiency of ore sorting.

[0114] In one embodiment, the first band image and the second band image of the target ore are collected by the first collection point; the X-ray image is collected by the second collection point;

[0115] The image acquisition module includes:

[0116] The regional image acquisition module is used to acquire visible light images and near infrared images of the target area.

[0117] The first band image acquisition module is used to perform edge detection and image segmentation on the visible light image of the target area, determine the spatial position of the target ore, and obtain the first band image of the target ore.

[0118] The second band image acquisition module is used to extract the second band image of the target ore from the near infrared image of the target area according to the spatial position of the target ore.

[0119] The time difference determination module is used to obtain the acquisition time difference between the first band image and the X-ray image according to the position distance between the first acquisition point and the second acquisition point and the moving speed of the conveyor belt.

[0120] The X-ray image acquisition module is used to obtain an X-ray image of the target ore according to the acquisition time difference and the spatial position of the target ore.

[0121] In one embodiment, the image normalization module includes:

[0122] The image scaling module is used to scale the first band image, the second band image and the X-ray image to a preset size according to a bilinear interpolation algorithm.

[0123] The image stitching module is used to stitch the scaled first band image, the second band image and the X-ray image to obtain a multi-channel feature map.

[0124] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An artificial intelligence sorting method based on multi-image fusion, It is characterized in that The method comprises: Acquire a first band image, a second band image and an X-ray image of the target ore; the first band image is a visible light image; the second band image is a near infrared image; Scaling the first waveband image, the second waveband image and the X-ray image to obtain a multi-channel feature map; Using the multi-channel feature map as an input of a pre-trained classification model to obtain the type of target ore; According to the type of the target ore, the target ore is transported to a corresponding receiving area.

2. According to claim 1, an artificial intelligence sorting method based on multi-image fusion, It is characterized in that The first band image and the second band image of the target ore are collected by the first collection point; the X-ray image is collected by the second collection point; The obtaining of the first waveband image, the second waveband image and the X-ray image of the target ore comprises: Acquire visible light images and near infrared images of the target area; Performing edge detection and image segmentation on the visible light image of the target area to determine the spatial position of the target ore and obtain a first band image of the target ore; Extracting a second band image of the target ore from the near infrared image of the target area according to the spatial position of the target ore; Obtaining a time difference between the first band image and the X-ray image according to the position distance between the first acquisition point and the second acquisition point and the moving speed of the conveyor belt; An X-ray image of the target ore is obtained according to the acquisition time difference and the spatial position of the target ore.

3. According to claim 1, an artificial intelligence sorting method based on multi-image fusion, It is characterized in that Scaling the first band image, the second band image and the X-ray image to obtain a multi-channel feature map includes: According to a bilinear interpolation algorithm, scaling the first waveband image, the second waveband image and the X-ray image to a preset size; The scaled first band image, second band image and X-ray image are spliced ​​to obtain a multi-channel feature map.

4. According to claim 1, an artificial intelligence sorting method based on multi-image fusion, It is characterized in that The classification model includes a feature extraction layer, a feature fusion layer and a fully connected layer; the feature extraction layer includes a first convolution branch, a second convolution branch and a third convolution branch; the multi-channel feature map includes an RGB three-channel layer, a near-infrared single-channel layer and an X-ray single-channel layer; in: The first convolution branch is used to receive the RGB three-channel layer as input and output a first feature tensor; The second convolution branch is used to receive the near-infrared single-channel layer as input and output a second feature tensor; The third convolution branch is used to receive the X-ray single channel layer as input and output a third feature tensor; The feature fusion layer is used to receive the first feature tensor, the second feature tensor and the third feature tensor as input, and output a global feature vector; The fully connected layer is used to receive the global feature vector as input and output the type of target ore.

5. According to claim 4, an artificial intelligence sorting method based on multi-image fusion, It is characterized in that The feature fusion layer receives the first feature tensor, the second feature tensor, and the third feature tensor as input, and outputs a global feature vector including: Concatenate the first feature tensor and the second feature tensor in the channel dimension to obtain a fourth feature tensor; Using a preset number of 1×1 convolution kernels to perform convolution operations on the fourth feature tensor, respectively, to obtain a fifth feature tensor; Performing a nonlinear transformation on the fifth eigentensor using a ReLu function to obtain a sixth eigentensor, and expanding the sixth eigentensor into a one-dimensional first eigenvector; Opening the third eigentensor into a one-dimensional second eigenvector; The first feature vector and the second feature vector are concatenated to obtain a global feature vector.

6. An artificial intelligence sorting device based on multi-image fusion, It is characterized in that The device comprises: An image acquisition module, used to acquire a first band image, a second band image and an X-ray image of the target ore; the first band image is a visible light image; the second band image is a near infrared image; An image standardization module, used for scaling the first band image, the second band image and the X-ray image to obtain a multi-channel feature map; An ore classification module, used to use the multi-channel feature map as an input of a pre-trained classification model to obtain the type of target ore; The separation module is used to transport the target ore to a corresponding receiving area according to the type of the target ore.

7. The artificial intelligence sorting device based on multi-image fusion according to claim 6, It is characterized in that The first band image and the second band image of the target ore are collected by the first collection point; the X-ray image is collected by the second collection point; The image acquisition module comprises: A regional image acquisition module is used to acquire visible light images and near infrared images of the target area; A first band image acquisition module, used for performing edge detection and image segmentation on the visible light image of the target area, determining the spatial position of the target ore, and obtaining a first band image of the target ore; A second band image acquisition module, used to extract a second band image of the target ore from the near infrared image of the target area according to the spatial position of the target ore; A time difference determination module, used to obtain the acquisition time difference between the first band image and the X-ray image according to the position distance between the first acquisition point and the second acquisition point and the moving speed of the conveyor belt; The X-ray image acquisition module is used to obtain the X-ray image of the target ore according to the acquisition time difference and the spatial position of the target ore.

8. The artificial intelligence sorting device based on multi-image fusion according to claim 6, It is characterized in that The image standardization module comprises: An image scaling module, used for scaling the first waveband image, the second waveband image and the X-ray image to a preset size according to a bilinear interpolation algorithm; The image stitching module is used to stitch the scaled first band image, the second band image and the X-ray image to obtain a multi-channel feature map.

9. The artificial intelligence sorting device based on multi-image fusion according to claim 6, It is characterized in that The classification model includes a feature extraction layer, a feature fusion layer and a fully connected layer; the feature extraction layer includes a first convolution branch, a second convolution branch and a third convolution branch; the multi-channel feature map includes an RGB three-channel layer, a near-infrared single-channel layer and an X-ray single-channel layer; in: The first convolution branch is used to receive the RGB three-channel layer as input and output a first feature tensor; The second convolution branch is used to receive the near-infrared single-channel layer as input and output a second feature tensor; The third convolution branch is used to receive the X-ray single channel layer as input and output a third feature tensor; The feature fusion layer is used to receive the first feature tensor, the second feature tensor and the third feature tensor as input, and output a global feature vector; The fully connected layer is used to receive the global feature vector as input and output the type of target ore.

10. The artificial intelligence sorting device based on multi-image fusion according to claim 9, It is characterized in that The feature fusion layer is specifically used for: Concatenate the first feature tensor and the second feature tensor in the channel dimension to obtain a fourth feature tensor; Using a preset number of 1×1 convolution kernels to perform convolution operations on the fourth feature tensor, respectively, to obtain a fifth feature tensor; Performing a nonlinear transformation on the fifth eigentensor using a ReLu function to obtain a sixth eigentensor, and expanding the sixth eigentensor into a one-dimensional first eigenvector; Opening the third eigentensor into a one-dimensional second eigenvector; The first feature vector and the second feature vector are concatenated to obtain a global feature vector.

Images