A neural network-based ore image recognition method

Abstract

The application relates to a neural network-based ore image recognition method, which comprises the following steps: constructing and training a mine point recognition model, including an input module, a global feature extraction module, a multi-resolution feature fusion module and an output module connected in sequence; the input module is used for inputting an ore image; the global feature extraction module is used for extracting the global feature of the ore image; the multi-resolution feature fusion module is used for extracting the multi-resolution feature of the ore image based on the global feature and fusing to obtain a multi-dimensional feature map; the output module is used for outputting a mine point recognition result based on the multi-dimensional feature map; a current ore image is collected, input into the trained mine point recognition model, and a mine point recognition result is output. The complementary combination of the global feature and the multi-resolution feature can realize detail recognition under global constraint, and significantly reduce the missed detection rate and the false detection rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a mineral image recognition method based on a neural network. BACKGROUND

[0002] The mineral point recognition based on the neural network is a technical system in which the feature extraction and the pattern recognition capability of deep learning are deeply fused with multi-source data of mineral resource exploration. The core advantage lies in breaking through the efficiency bottleneck and subjective limitations of traditional manual interpretation, realizing end-to-end intelligent recognition of "macroscopic clues-microscopic features-mineralization regularity" of mineral points, and has become a core technical direction in the current mineral exploration and supervision field.

[0003] However, the distribution trend of mineral points is complex and difficult to identify. Although there are local textures similar to mineral points, the global view may not conform to the distribution rule of mineral points. If only global features or single resolution features are extracted, it is easy to cause the problems of missing small mineral points or misjudging impurities. How to improve the core network model to reduce the missing rate and the false detection rate is a technical problem to be solved in the field. SUMMARY

[0004] The application provides a mineral image recognition method based on a neural network, which comprises the following steps:

[0005] A mineral point recognition model is constructed and trained, which comprises an input module, a global feature extraction module, a multi-resolution feature fusion module and an output module connected in sequence;

[0006] The input module is used for inputting a mineral image.

[0007] The global feature extraction module is used for extracting global features of the mineral image.

[0008] The multi-resolution feature fusion module is used for extracting multi-resolution features of the mineral image based on the global features, and fusing to obtain a multi-dimensional feature map.

[0009] The output module is used for outputting a mineral point recognition result based on the multi-dimensional feature map.

[0010] A current mineral image is collected, input into the trained mineral point recognition model, and a mineral point recognition result is output.

[0011] Further, the global feature extraction module comprises a bottom feature extractor and a Transformer encoder connected in sequence.

[0012] The bottom feature extractor adopts a convolutional neural network and is used for converting the mineral image into a local feature map.

[0013] The Transformer encoder is used for extracting global features from the local feature map.

[0014] Further, the bottom layer feature extractor comprises, in sequence, a head convolution unit, a depth separable convolution unit, a down-sampling unit and a dilated convolution unit.

[0015] Further, the Transformer encoder comprises:

[0016] The input unit is configured to flatten the local feature map into a plurality of feature tokens and add a global token;

[0017] The plurality of encoder units each comprises:

[0018] The multi-head self-attention is configured to calculate the correlation between the feature tokens;

[0019] The feedforward network is configured to perform nonlinear transformation on the feature tokens to strengthen the feature expression;

[0020] The output layer is configured to concatenate the output vector of the global token and the feature map of each position token to obtain the overall feature;

[0021] The up-sampling layer is configured to up-sample the overall feature to the size of the ore image to obtain the global feature.

[0022] Further, the multi-resolution feature fusion module comprises a depth separable convolution layer, and a first multi-resolution feature extraction unit, a second multi-resolution feature extraction unit and a third multi-resolution feature extraction unit connected with the depth separable convolution layer and arranged side by side; and a feature fusion unit;

[0023] The depth separable convolution layer is connected with the global feature extraction module and is configured to extract the first resolution feature of the ore image based on the global feature;

[0024] The first multi-resolution feature extraction unit is configured to extract the fourth resolution feature and the fourth resolution enhanced feature of the ore image based on the first resolution feature;

[0025] The second multi-resolution feature extraction unit is configured to extract the second resolution feature, the fifth resolution feature and the fifth resolution enhanced feature of the ore image based on the first resolution feature;

[0026] The third multi-resolution feature extraction unit is configured to extract the third resolution feature, the sixth resolution feature and the sixth resolution enhanced feature of the ore image based on the first resolution feature;

[0027] The feature fusion unit is configured to fuse the resolution features and the enhanced features to obtain a multi-dimensional feature map of the ore image.

[0028] Further, the first multi-resolution feature extraction unit comprises:

[0029] The first eight times down-sampling layer is connected with the depth separable convolution layer, and is used for extracting fourth resolution features based on the first resolution features;

[0030] The first hole convolution layer is connected with the first eight times down-sampling layer, and is used for performing hole convolution on the fourth resolution features by using different hole rates to extract fourth resolution enhanced features.

[0031] Further, the second multi-resolution feature extraction unit comprises a two times down-sampling layer, a second eight times down-sampling layer, a second hole convolution layer and a two times up-sampling layer connected in sequence;

[0032] The two times down-sampling layer is connected with the depth separable convolution layer, and is used for extracting second resolution features based on the first resolution features;

[0033] The second eight times down-sampling layer is used for extracting fifth resolution features based on the second resolution features;

[0034] The second hole convolution layer is used for performing hole convolution on the fifth resolution features by using different hole rates to extract fifth resolution enhanced features;

[0035] The two times up-sampling layer is used for performing feature map size matching on the fifth resolution enhanced features to obtain the fifth resolution enhanced features after size matching.

[0036] Further, the third multi-resolution feature extraction unit comprises a four times down-sampling layer, a third eight times down-sampling layer, a third hole convolution layer and a four times up-sampling layer connected in sequence;

[0037] The four times down-sampling layer is connected with the depth separable convolution layer, and is used for extracting third resolution features based on the first resolution features;

[0038] The third eight times down-sampling layer is used for extracting sixth resolution features based on the third resolution features;

[0039] The third hole convolution layer is used for performing hole convolution on the sixth resolution features by using different hole rates to extract sixth resolution enhanced features;

[0040] The four times up-sampling layer is used for performing feature map size matching on the sixth resolution enhanced features to obtain the sixth resolution enhanced features after size matching.

[0041] Further, the feature fusion unit comprises a top-down fusion layer, a cross-layer connection layer, a superposition layer and a top-down sampling layer;

[0042] The top-down fusion layer is connected with the first multi-resolution feature extraction unit, and is used for top-down fusion of the first resolution feature, the fourth resolution feature and the fourth resolution enhanced feature to obtain a first multi-resolution feature set;

[0043] The cross-layer connection layer is connected with the multi-resolution feature fusion module, and is used for cross-layer connection of the first resolution feature, the fourth resolution feature, the fourth resolution enhanced feature, the second resolution feature, the fifth resolution feature, the fifth resolution enhanced feature matched in size, the third resolution feature, the sixth resolution feature and the sixth resolution enhanced feature matched in size to obtain a second multi-resolution feature set.

[0044] The superposition layer is used for superposition determination of the comprehensive feature based on the first multi-resolution feature set and the second multi-resolution feature set.

[0045] The top-down sampling layer is used for improving the resolution of the comprehensive feature to obtain a multi-dimensional feature map.

[0046] Further, the output module comprises a 1*1 convolution and a binary classification activation function; a 1*1 convolution is used to obtain an image with the same channel number as a label image, and a cross-entropy loss is calculated with the label image; finally, a Softmax is used to output the probability of "ore point" and "non-ore point", so as to realize the binary classification result of the ore point and the non-ore point.

[0047] The ore image recognition method based on the neural network provided by the application adopts the architecture of "input module-global feature extraction module-multi-resolution feature fusion module-output module", and is designed in view of the characteristics of the ore image (such as large difference in ore point size, strong correlation between local details and global structure, complex background interference and the like), and the core lies in: complementary of global features and multi-resolution features: the global feature extraction module focuses on global features (such as ore overall structure, ore point distribution trend and background matrix type), so as to avoid misjudgment caused by local isolated features (such as a single dark spot) (for example: although a certain area has local texture similar to the ore point, it does not conform to the ore point distribution rule in the global view, and can be excluded); the multi-resolution feature fusion module further supplements local details of different scales (such as the edge of a 0.4mm small ore point and the outline of a 10mm large ore point), so as to solve the problem of missing small ore points or misjudging impurities by using a single resolution feature; the combination of the two realizes "detail recognition under global constraint", and can significantly reduce the missing rate (especially for small ore points) and the false detection rate (such as impurities and noise). BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative effort.

[0049] Figure 1 Flowchart of an embodiment of the ore image recognition method based on neural network of the present application;

[0050] Figure 2 Structure diagram of an embodiment of the ore point recognition model of the present application;

[0051] Figure 3 Raw ore map of high and low energy channel merging of the present application; Figure 4 Recognition result diagram of the present application, white is the recognized ore point. DETAILED DESCRIPTION

[0052] In the following description, for the purpose of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0053] It should be understood that the term "comprising" as used in the specification and the appended claims indicates the presence of the recited features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0054] It should also be understood that the term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, and includes these combinations.

[0055] As used in the specification and the appended claims, the term "if' can be interpreted as meaning "when" or "once" or "in response to a determination" or "in response to detecting" depending on the context. Similarly, the phrase "if determined" or "if detected [the described condition or event]" can be interpreted to mean "once determined" or "in response to a determination" or "once detected [the described condition or event]" or "in response to detecting [the described condition or event]" depending on the context.

[0056] In addition, in the description of the present application and the appended claims, the terms "first", "second", "third", etc. are used only to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0057] In the present application, the reference to "one embodiment" or "some embodiments" means that the specific features, structures or characteristics described in connection with the embodiment are included in one or more embodiments of the present application. Therefore, the statements "in one embodiment", "in some embodiments", "in other some embodiments", "in further some embodiments" and the like appearing in different places in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "include", "contain", "have" and their variants mean "include but not limited to", unless otherwise specifically emphasized. For ease of understanding, the technical solutions of the present application will be described in detail below with reference to the drawings.

[0058] Figure 1 A neural network-based ore image recognition method is provided for an embodiment of the present application. For ease of illustration, only parts related to the present embodiment are shown. The method provided by the present embodiment includes the following steps:

[0059] S1: constructing and training a mine point recognition model, including an input module, a global feature extraction module, a multi-resolution feature fusion module and an output module connected in turn;

[0060] The input module is used for inputting an ore image.

[0061] The global feature extraction module is used for extracting global features of the ore image.

[0062] The multi-resolution feature fusion module is used for extracting multi-resolution features of the ore image based on the global features, and fusing to obtain a multi-dimensional feature map.

[0063] The output module is used for outputting a mine point recognition result based on the multi-dimensional feature map.

[0064] S2: collecting a current ore image, inputting the trained mine point recognition model, and outputting a mine point recognition result.

[0065] In this embodiment, an ore image recognition method of the present application is given, and the ore point recognition model adopts the architecture of "input module→global feature extraction module→multi-resolution feature fusion module→output module". The characteristics (such as large difference in ore point size, strong correlation between local details and global structure, complex background interference, etc.) of ore images (such as X-ray images, visible light images) are designed in a targeted manner, and the core lies in: complementary of global features and multi-resolution features: the global feature extraction module focuses on global features (such as overall structure of ore, ore point distribution trend, background matrix type), avoiding misjudgment caused by local isolated features (such as single dark spot) (for example: although a certain area has local texture similar to ore points, it does not conform to the ore point distribution rule in the global view, which can be excluded); the multi-resolution feature fusion module further supplements local details of different scales (such as the edge of 0.4mm small ore points and the outline of 10mm large ore points), solving the problem of "single resolution feature missing small ore points or misjudging impurities"; the combination of the two realizes "detail recognition under global constraint", which can significantly reduce the missing rate (especially for small ore points) and the false detection rate (such as impurities and noise).

[0066] (I) Structure of ore point recognition model:

[0067] 1. Input module:

[0068] Specifically, the ore image can be an X-ray image. Due to the characteristics of the device (such as fluctuation of X-ray intensity, noise of the detector) and the difference of the sample (such as uneven thickness of the ore), there are problems such as noise, uneven brightness, geometric shift, etc. The input module can optionally include several preprocessing units to perform preliminary preprocessing on the ore image.

[0069] Preferably, the input module can optionally include:

[0070] The denoising unit can use Gaussian filtering (to eliminate high-frequency noise), median filtering (to eliminate salt and pepper noise), or more optimal adaptive filtering (such as BM3D) to avoid misjudgment of noise as "small ore points";

[0071] The normalization unit can scale the pixel value to [0, 1] or [-1, 1] to eliminate the interference of image brightness difference on model training (such as different brightness of X-ray images taken in different batches);

[0072] The data augmentation unit expands the training data set by random rotation (±15°), translation (±5%), scaling (0.8-1.2 times), and horizontal flip to avoid model overfitting (especially suitable for small sample ore point data).

[0073] 2. Global feature extraction module

[0074] Preferably, the global feature extraction module comprises a bottom feature extractor and a Transformer encoder connected in sequence.

[0075] The bottom feature extractor adopts a convolutional neural network to convert the ore image into a local feature map.

[0076] Specifically, a lightweight convolutional neural network CNN (such as MobileNetV3) can be used as the bottom feature extractor, and a hollow convolution is introduced to enhance the position information, so as to convert the original ore image into a "local feature vector" (each vector contains texture, brightness and other information of a 32x32 region), and output a 16x16x512 feature map (compressing the 512x512 input to a spatial dimension of 16x16, and each position corresponds to a 32x32 pixel region of the original image), reducing the calculation amount of the subsequent Transformer encoder. More specifically, the bottom feature extractor comprises a head convolution unit, a depth separable convolution unit and a down-sampling unit (X32) connected in sequence.

[0077] The Transformer encoder is used to extract global features from the local feature map.

[0078] Preferably, the Transformer encoder comprises:

[0079] The input unit is used to flatten the local feature map into a plurality of feature tokens and add one global token; and is optionally connected with the above-mentioned down-sampling unit;

[0080] For example, the 16x16x512 feature map is flattened into 512 "feature tokens", each token having a dimension of 256; and one "global token" is added to aggregate overall information;

[0081] A plurality of encoder units, such as 6 layers of Transformer encoder, each encoder unit comprising:

[0082] Multi-Head Attention (4 heads) is used to calculate the correlation between each feature token, such as whether the ore point distribution of the "top-left token" is related to the "bottom-right token";

[0083] The feedforward network (FFN) is used to perform nonlinear transformation on each token to strengthen the feature expression;

[0084] The output layer is used to splice the output vector (256 dimensions) of the global token with the feature map of each position token (feature map of 512 position tokens (16x16x512)), to obtain the overall feature;

[0085] an up-sampling layer, configured to up-sample the overall feature to the size of the ore image to obtain a global feature.

[0086] In this embodiment, a preferred embodiment of the global feature extraction module is given, but it is not limited thereto. First, the local feature map is extracted through the bottom feature encoder; then, according to the ore point feature, the output vector of the global token and the feature map of each position token are extracted through the Transformer encoder, the former represents the global semantic, and the latter retains the spatial position information. Compared with the pure convolutional neural network CNN, it can capture the cross-regional ore point distribution rule (such as “ore point A and ore point B are separated by 100 pixels but belong to the same ore vein”), is suitable for complex ore samples (such as mixed ore and sparse distribution of ore points), and can further improve the accuracy of subsequent ore point recognition.

[0087] Preferably, the multi-resolution feature fusion module comprises a depth separable convolution layer, and a first multi-resolution feature extraction unit, a second multi-resolution feature extraction unit and a third multi-resolution feature extraction unit connected with the depth separable convolution layer and arranged side by side; and a feature fusion unit;

[0088] The depth separable convolution layer is connected with the global feature extraction module and is configured to extract the first resolution feature of the ore image based on the global feature.

[0089] The first multi-resolution feature extraction unit is configured to extract the fourth resolution feature and the fourth resolution enhanced feature of the ore image based on the first resolution feature.

[0090] The second multi-resolution feature extraction unit is configured to extract the second resolution feature, the fifth resolution feature and the fifth resolution enhanced feature of the ore image based on the first resolution feature.

[0091] The third multi-resolution feature extraction unit is configured to extract the third resolution feature, the sixth resolution feature and the sixth resolution enhanced feature of the ore image based on the first resolution feature.

[0092] Wherein, the first resolution > the second resolution > the third resolution > the fourth resolution > the fifth resolution > the sixth resolution.

[0093] The feature fusion unit is configured to fuse the resolution features and the enhanced features to obtain a multi-dimensional feature map of the ore image.

[0094] In this embodiment, a preferred embodiment of the multi-resolution feature fusion unit is given, which first extracts a first resolution feature based on global features, and then extracts resolution features and enhanced features of each dimension based on the first resolution feature. The essence is to solve the problems of "insufficient original data features" and "complex scene interference"; the resolution features of each dimension enable the model to "see clearly" the key information of ore spots at different scales, and the enhanced features enable the model to "focus and grasp" effective features to exclude interference. Both of them jointly support the model to break through the precision bottleneck of traditional recognition, directly determine the reliability of the whole process from "macroscopic screening" to "microscopic verification" of ore spot recognition, and improve the overall recognition accuracy of the model.

[0095] Notably, the specific resolution ranges of the first resolution, the second resolution, the third resolution, the fourth resolution, the fifth resolution, and the sixth resolution are not specifically limited, and are relative to the input ore image of the model. The resolutions of the extracted features are determined by different down-sampling layers, as long as the following conditions are met: first resolution > second resolution > third resolution > fourth resolution > fifth resolution > sixth resolution. Optionally, the first resolution is a high resolution, the first resolution feature extraction unit extracts lower resolution features and lower resolution enhanced features based on the high resolution; the second resolution feature extraction unit extracts medium resolution features, lower resolution features, and lower resolution enhanced features based on the high resolution features; the third resolution feature extraction unit extracts low resolution features, ultra-low resolution features, and ultra-low resolution enhanced features based on the high resolution features; wherein, high resolution > medium resolution > low resolution > lower resolution > lower resolution > ultra-low resolution.

[0096] As an example, the first multi-resolution feature extraction unit includes:

[0097] The first eight times down-sampling layer is connected with the depth separable convolution layer and is used to extract the fourth resolution feature based on the first resolution feature;

[0098] The first hollow convolution layer is connected with the first eight times down-sampling layer and is used to perform hollow convolution on the fourth resolution feature by using different hollow rates to extract the fourth resolution enhanced feature.

[0099] As an example, the second multi-resolution feature extraction unit includes a two times down-sampling layer, a second eight times down-sampling layer, a second hollow convolution layer, and a two times up-sampling layer connected in sequence;

[0100] The two times down-sampling layer is connected with the depth separable convolution layer and is used to extract the second resolution feature based on the first resolution feature;

[0101] The second eight times down-sampling layer is used to extract the fifth resolution feature based on the second resolution feature;

[0102] a second cavity convolutional layer, configured to perform cavity convolution on the fifth resolution feature with different cavity rates to extract a fifth resolution enhanced feature;

[0103] a two times up-sampling layer, configured to perform feature map size matching on the fifth resolution enhanced feature to obtain a size-matched fifth resolution enhanced feature.

[0104] As an example, the third multi-resolution feature extraction unit includes, in sequence, a four times down-sampling layer, a third eight times down-sampling layer, a third cavity convolutional layer, and a four times up-sampling layer;

[0105] the four times down-sampling layer is connected with the depth separable convolutional layer and is configured to extract a third resolution feature based on the first resolution feature;

[0106] the third eight times down-sampling layer is configured to extract a sixth resolution feature based on the third resolution feature;

[0107] the third cavity convolutional layer is configured to perform cavity convolution on the sixth resolution feature with different cavity rates to extract a sixth resolution enhanced feature;

[0108] the four times up-sampling layer is configured to perform feature map size matching on the sixth resolution enhanced feature to obtain a size-matched sixth resolution enhanced feature.

[0109] Preferably, the feature fusion unit includes, in sequence, a top-down fusion layer, a cross-layer connection layer, and a superposition layer;

[0110] the top-down fusion layer is connected with the first multi-resolution feature extraction unit and is configured to perform top-down fusion on the first resolution feature, the fourth resolution feature, and the fourth resolution enhanced feature to obtain a first multi-resolution feature set;

[0111] the cross-layer connection layer is connected with the multi-resolution feature fusion module and is configured to perform cross-layer connection on the first resolution feature, the fourth resolution feature, the fourth resolution enhanced feature, the second resolution feature, the fifth resolution feature, the size-matched fifth resolution enhanced feature, the third resolution feature, the sixth resolution feature, and the size-matched sixth resolution enhanced feature to obtain a second multi-resolution feature set;

[0112] the superposition layer is configured to superimpose a comprehensive feature based on the first multi-resolution feature set and the second multi-resolution feature set;

[0113] the top-down sampling layer is configured to improve the resolution of the comprehensive feature to obtain a multi-dimensional feature map. Specifically, the first resolution is extracted and the details are enriched by decoding from bottom to top, and a multi-dimensional feature map with the same size as the label image is output.

[0114] As an example, the cross-layer connection layer is specifically connected with a deep separable convolution layer, a first eight times down-sampling layer, a first hole convolution layer, a two times down-sampling layer, a second eight times down-sampling layer, a two times up-sampling layer, a four times down-sampling layer, a third eight times down-sampling layer and a four times up-sampling layer, for cross-layer connection of the first resolution feature, the fourth resolution feature, the fourth resolution enhanced feature, the second resolution feature, the fifth resolution feature, the fifth resolution enhanced feature after size matching, the third resolution feature, the sixth resolution feature and the sixth resolution enhanced feature after size matching, to obtain the second multi-resolution feature set.

[0115] Preferably, the output module includes a 1x1 convolution and a binary classification activation function; specifically, an image with the same number of channels as the label image is obtained through the 1x1 convolution, which can be exactly cross-entropy loss with the label image. Finally, the Softmax outputs the probability of "ore point" and "non-ore point", realizing the binary classification result of ore point and non-ore point.

[0116] (II) Training of the ore point recognition model

[0117] 1. Data set construction

[0118] Data collection: Collect ore images (such as X-ray images, visible light images) covering diversified scenes. Preferably, ore images of different ore types (such as copper ore, iron ore, gold ore), different ore point sizes (1mm-20mm), different forms (point, strip, aggregate), and different interference conditions (such as device noise, uneven sample thickness, background impurities) are collected to ensure data diversity. Preferably, at least 10,000 images are included, and if it is a small sample scene, it can be expanded to more than 5,000 by combining data enhancement.

[0119] Data labeling: Label the ore point; optionally labeled as 0 / 1.

[0120] Data division: Optionally divided into training set (model learning), validation set (evaluation during training, adjusting hyperparameters), and test set (final performance evaluation) in the ratio of 7:2:1. Preferably, ensure consistent distribution of the three types of data (such as same ore type ratio, same ore point size distribution) when dividing.

[0121] 2. Data preprocessing (input module integration), standardize the image before training, and keep consistent with the input module in the inference stage:

[0122] Denoising: Remove device noise using Gaussian filtering and median filtering;

[0123] Normalization: Scale the pixel value to [0, 1] or [-1, 1] (such as X-ray image grayscale value is usually 0-65535, which needs to be scaled to a unified range);

[0124] Data augmentation: Geometric transformations: random rotation (±30°), translation (±10%), scaling (0.7-1.3x), horizontal / vertical flipping; Pixel transformations: random adjustment of brightness (±20%) and contrast (±20%) to simulate different shooting conditions; Advanced enhancements: for small sample minerals, use StyleGAN to generate virtual samples, or enhance generalization ability through MixUp (image blending).

[0125] Training process

[0126] During training, the preferred loss function is cross-entropy loss, which optimizes the difference between the probability prediction and the true label.

[0127] The design of the global feature extraction module and the multi-resolution feature fusion module in the mineral deposit identification model provided in this application are used to extract global features and multi-resolution features of the ore image, respectively. This enables the mineral deposit identification model to extract rich feature information from the ore image, improves its mineral deposit recovery capability, and thus improves the accuracy of the mineral deposit recovery image.

[0128] On the other hand, this application also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above methods.

[0129] On the other hand, this application also provides an electronic device, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to perform any of the methods described above.

[0130] For example, the program code may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in an electronic device.

[0131] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the electronic device may also include input / output devices, network access devices, buses, etc.

[0132] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0133] The memory can be an internal storage unit of the electronic device, such as a hard disk or a memory. The memory can also be an external storage device of the electronic device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory can include both the internal storage unit and the external storage device of the electronic device. The memory is used to store the program code and other programs and data required by the electronic device. The memory can also be used to temporarily store data that has been output or will be output.

[0134] The computer storage medium and the electronic device based on the above method are created, and the technical effects and advantages are not repeated here. The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the description.

[0135] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A neural network-based ore image recognition method, characterized by, The method comprises: The method comprises: The method comprises: The input module is used for inputting the ore image; The global feature extraction module is used for extracting the global feature of the ore image; The multi-resolution feature fusion module is used for extracting the multi-resolution feature of the ore image based on the global feature and fusing to obtain a multi-dimensional feature map; the multi-resolution feature fusion module comprises a depth separable convolution layer, a first multi-resolution feature extraction unit, a second multi-resolution feature extraction unit and a third multi-resolution feature extraction unit which are connected with the depth separable convolution layer in parallel, and a feature fusion unit; the depth separable convolution layer is connected with the global feature extraction module and used for extracting the first resolution feature of the ore image based on the global feature; the first multi-resolution feature extraction unit is used for extracting the fourth resolution feature and the fourth resolution enhanced feature of the ore image based on the first resolution feature; the second multi-resolution feature extraction unit is used for extracting the second resolution feature, the fifth resolution feature and the fifth resolution enhanced feature of the ore image based on the first resolution feature; the third multi-resolution feature extraction unit is used for extracting the third resolution feature, the sixth resolution feature and the sixth resolution enhanced feature of the ore image based on the first resolution feature; wherein the first resolution > the second resolution > the third resolution > the fourth resolution > the fifth resolution > the sixth resolution; the feature fusion unit is used for fusing the resolution features and the enhanced features to obtain the multi-dimensional feature map of the ore image; The output module is used for outputting the ore spot recognition result based on the multi-dimensional feature map; 2. The method of claim 1, wherein, The current ore image is collected, input into the trained ore spot recognition model, and the ore spot recognition result is output. The global feature extraction module comprises a bottom feature extractor and a Transformer encoder which are connected in sequence. The bottom feature extractor adopts a convolutional neural network and is used for converting the ore image into a local feature map.

3. The method of claim 2, wherein, The Transformer encoder is used for extracting the global feature based on the local feature map.

4. The method of claim 2, wherein, The bottom feature extractor comprises a head convolution unit, a depth separable convolution unit, a down-sampling unit and a hole convolution unit which are connected in sequence. The Transformer encoder comprises: The input unit is used for flattening the local feature map into a plurality of feature tokens and adding one global token; Each encoder unit comprises: The multi-head self-attention is used for calculating the association between the feature tokens; The feedforward network is used for nonlinearly transforming the feature tokens to strengthen the feature expression; The output layer is used for splicing the output vector of the global token and the feature map of each position token to obtain the overall feature; 5. The method of claim 1, wherein, The up-sampling layer is used for up-sampling the overall feature to the size of the ore image to obtain the global feature. The first multi-resolution feature extraction unit comprises: The first eight times down-sampling layer is connected with the depth separable convolution layer and used for extracting the fourth resolution feature based on the first resolution feature; The first cavity convolutional layer is connected with the first eight times down-sampling layer and is configured to perform cavity convolution on the fourth resolution feature with different cavity rates to extract a fourth resolution enhanced feature.

6. The method of claim 1, wherein, The second multi-resolution feature extraction unit comprises, in sequence, a two times down-sampling layer, a second eight times down-sampling layer, a second cavity convolutional layer and a two times up-sampling layer; The two times down-sampling layer is connected with the depth separable convolutional layer and is configured to extract a second resolution feature based on the first resolution feature; The second eight times down-sampling layer is configured to extract a fifth resolution feature based on the second resolution feature; The second cavity convolutional layer is configured to perform cavity convolution on the fifth resolution feature with different cavity rates to extract a fifth resolution enhanced feature; The two times up-sampling layer is configured to perform feature map size matching on the fifth resolution enhanced feature to obtain a size-matched fifth resolution enhanced feature.

7. The method of claim 1, wherein, The third multi-resolution feature extraction unit comprises, in sequence, a four times down-sampling layer, a third eight times down-sampling layer, a third cavity convolutional layer and a four times up-sampling layer; The four times down-sampling layer is connected with the depth separable convolutional layer and is configured to extract a third resolution feature based on the first resolution feature; The third eight times down-sampling layer is configured to extract a sixth resolution feature based on the third resolution feature; The third cavity convolutional layer is configured to perform cavity convolution on the sixth resolution feature with different cavity rates to extract a sixth resolution enhanced feature; The four times up-sampling layer is configured to perform feature map size matching on the sixth resolution enhanced feature to obtain a size-matched sixth resolution enhanced feature.

8. The method of claim 1, wherein, The feature fusion unit comprises a top-down fusion layer, a cross-layer connection layer, a superposition layer and a bottom-up sampling layer; The top-down fusion layer is connected with the first multi-resolution feature extraction unit and is configured to perform top-down fusion on the first resolution feature, the fourth resolution feature and the fourth resolution enhanced feature to obtain a first multi-resolution feature set; The cross-layer connection layer is connected with the multi-resolution feature fusion module and is configured to perform cross-layer connection on the first resolution feature, the fourth resolution feature, the fourth resolution enhanced feature, the second resolution feature, the fifth resolution feature, the size-matched fifth resolution enhanced feature, the third resolution feature, the sixth resolution feature and the size-matched sixth resolution enhanced feature to obtain a second multi-resolution feature set; The superposition layer is configured to superimpose a comprehensive feature based on the first multi-resolution feature set and the second multi-resolution feature set; The bottom-up sampling layer is configured to improve the resolution of the comprehensive feature to obtain a multi-dimensional feature map.

9. The method according to any one of claims 1 to 8, characterized in that, The output module comprises a 1x1 convolution and a binary classification activation function; a 1x1 convolution is used to obtain an image with the same number of channels as a label image, and a cross-entropy loss is calculated between the image and the label image; finally, a Softmax output is used to output the probability of "mineral point" or "non-mineral point", so as to realize the binary classification result of the mineral point or the non-mineral point.

Images