Ore image segmentation method and system integrating channel space multi-scale features

Abstract

The present invention belongs to the technical field of ore image detection, and provides an ore image segmentation method and system that integrates multi-scale features of channel space. The technical solution is to add CSMA after combining CSA and DS‑ASPP to the ResUNet jump connection and bridge part to form a multi-scale attention mechanism ore segmentation network, which can more effectively detect boundaries in ore images, improve segmentation accuracy, and lay the foundation for subsequent ore particle size distribution detection. Ore images are usually taken in complex mineral processing scenes. CSA enhances key features through the attention mechanism, thereby effectively dealing with background interference and improving the robustness of segmentation, while DS‑ASPP can better extract global and local features. Through high-precision segmentation results, subsequent ore particle size detection will become more reliable and accurate, solving the problem of poor performance in complex ore image acquisition environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of ore image detection, and in particular relates to an ore image segmentation method and system integrating multi-scale features of channel space. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Ore particle size detection is crucial in the development and utilization of mineral resources. Its results can reflect the degree of crushing of the ore by the crusher, help determine the presence of oversized or strip-shaped ore in the ore, and thus prevent stone blockage. In addition, the results of ore particle size detection can also be used as a basis for reasonably setting the operating parameters of the crusher, such as power and belt speed. Therefore, accurate particle size detection helps to optimize process parameters and improve crusher efficiency. With the development of mining technology, especially the advancement of automation and intelligence, accurate particle size detection has become an important indicator of mining modernization, which can effectively support real-time monitoring and feedback adjustment. Ore particle size detection is not only related to the economic benefits of the mining industry, but also affects environmental protection and the sustainable use of resources. Therefore, it is particularly important to adopt efficient and accurate particle size detection technology.

[0004] Early manual physical screening is a traditional method of ore particle size detection and sorting, which mainly relies on manual operation and simple tools, usually using a metal or cloth screen, and the aperture is adjusted according to needs. This method is easy to operate, easy to understand, and suitable for small-scale operations, but the particle size detection accuracy is limited and it is easily affected by human factors. When processing a large number of samples, it is inefficient and time-consuming, and requires more manual participation, which increases the workload.

[0005] Traditional ore image segmentation methods mainly rely on digital image processing technology. First, after using a camera to obtain a high-definition image of the ore, image preprocessing operations such as denoising and contrast enhancement are performed to improve image quality. Then, the edges of ore particles are identified through algorithms (such as Canny edge detection) to segment the image into different particles. Commonly used methods include threshold segmentation and region growing. However, the image quality of traditional methods may be affected by changes in illumination, resulting in unstable processing results, and the analysis becomes more difficult when processing complex shapes or overlapping particles.

[0006] The deep learning-based segmentation method can reduce manual intervention, automate the segmentation process, and improve efficiency. The segmentation model can automatically extract the complex features of the ore and can usually achieve higher segmentation accuracy. Especially when processing complex images, it can automatically extract multi-level features and process diversified image data. However, the existing deep learning segmentation methods consume a lot of computing resources, require high-performance computing equipment for training, and the model is prone to overfitting when there are insufficient samples or uneven data distribution. Due to the complex environment of ore image acquisition, simple deep learning models may not perform well in complex ore image acquisition environments and are easily affected by background interference, resulting in reduced segmentation accuracy. Summary of the invention

[0007] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides an ore image segmentation method and system integrating channel space multi-scale features, which aims to improve the comprehensive efficiency of ore particle size detection, and designs a more reliable and accurate ore particle size detection method based on the design and construction effect of the image segmentation model, so as to achieve effective control of the ore size and promote the development of ore particle size detection.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] The first aspect of the present invention provides a method for ore image segmentation integrating channel space multi-scale features, comprising the following steps:

[0010] Obtaining an ore image dataset;

[0011] The ore image segmentation model is trained based on the ore image data set to obtain a trained ore image segmentation model; wherein the construction process of the ore image segmentation model includes:

[0012] The encoder extracts the initial feature map of the ore image; the CSMA module is added to the jump connection and bridge between the encoder and the decoder; the low-level feature information is fused into the high-level feature map through the CSMA module, including:

[0013] Based on the initial feature map of the ore image, extracting the channel attention feature map of the ore image;

[0014] The initial feature map of the ore image and the channel attention feature map are fused to obtain the channel fused attention feature map;

[0015] Based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map;

[0016] The channel space fusion attention feature map is subjected to multiple depth-separable dilated convolutions with different dilation rates to obtain the multi-scale feature map of the ore;

[0017] The multi-scale feature map of the ore is weight-mapped and fused with the channel fusion attention feature map to obtain the ore edge contour feature map;

[0018] The ore edge contour feature map is input into the decoder for decoding to obtain the segmentation result.

[0019] Furthermore, after obtaining the ore image dataset, the ore images are preprocessed, including square map equalization and low-pass filtering image enhancement operations on the ore images. Then, the enhanced images are flipped, translated and rotated for data enhancement, and Labelme is used to annotate the ore images to form a new ore image dataset.

[0020] Furthermore, based on the initial feature map of the ore image, a channel attention feature map of the ore image is extracted, including:

[0021] The initial feature map of the ore image is subjected to maximum pooling and average pooling, and the results of maximum pooling and average pooling are input into a shared multi-layer perceptron. After two convolution operations, the maximum pooling and average pooling results are connected together, and then the output results of the shared multi-layer perceptron are converted into weight values ​​between 0 and 1 through the Sigmoid activation function, and the channel attention feature map is obtained through channel attention.

[0022] Furthermore, the channel space fusion attention feature map is subjected to multiple depth-separable dilated convolutions with different dilation rates to obtain the ore multi-scale feature map, which is expressed as:

[0023] ,

[0024] in, is the output multi-scale feature map of the ore, It is a depth-separable dilated convolution using different dilation rates. represents the global average pooling module, The ASPP Pooling module is concatenated with the output of each depth-separable hole convolutional layer to introduce global context information. Indicates that a 1x1 convolutional layer is used for channel integration after concatenation.

[0025] Furthermore, when the initial feature map and the channel attention feature map of the ore image are fused to obtain the channel fused attention feature map, the channel fused attention feature map is obtained by element-by-element multiplication of the initial feature map and the channel attention feature map.

[0026] Furthermore, when extracting the initial feature map of the ore image, the original ore image is subjected to convolution and residual block operations to obtain the initial feature map.

[0027] Furthermore, based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map, including: performing global average pooling and maximum pooling on the channel fusion attention feature map in the spatial dimension, splicing the pooled features in the channel dimension, and performing a convolution operation to obtain the channel space fusion attention feature map of the ore image.

[0028] Furthermore, the method also includes using OpenCV to perform contour detection on the ore segmentation result to obtain the ore particle size detection result.

[0029] The second aspect of the present invention provides an ore image segmentation system integrating channel space multi-scale features, comprising:

[0030] An image acquisition module, which is used to acquire an ore image data set;

[0031] The image segmentation model training module is used to train the ore image segmentation model based on the ore image data set to obtain the trained ore image segmentation model; wherein the construction process of the ore image segmentation model includes:

[0032] The encoder extracts the initial feature map of the ore image; the CSMA module is added to the jump connection and bridge between the encoder and the decoder; the low-level feature information is fused into the high-level feature map through the CSMA module, including:

[0033] Based on the initial feature map of the ore image, extracting the channel attention feature map of the ore image;

[0034] The initial feature map of the ore image and the channel attention feature map are fused to obtain the channel fusion attention feature map; based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map; the channel space fusion attention feature map is subjected to a plurality of depth-separable dilated convolutions with different void rates to obtain the ore multi-scale feature map; the ore multi-scale feature map is weight-mapped and fused with the channel fusion attention feature map to obtain the ore edge contour feature map;

[0035] The image segmentation module is used for the ore image segmentation model to segment the ore image to obtain the segmentation result.

[0036] Furthermore, in the image segmentation model training module, when the initial feature map and the channel attention feature map of the ore image are fused to obtain the channel fused attention feature map, the channel fused attention feature map is obtained by element-by-element multiplication of the initial feature map and the channel attention feature map.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] Ore images are usually taken in complex mineral processing scenarios. The present invention combines CSA with DS-ASPP to form a channel space multi-scale attention mechanism (CSMA). CSA enhances key features through the attention mechanism, thereby effectively coping with background interference and improving the robustness of segmentation, while DS-ASPP can better extract global and local features. Through high-precision segmentation results, subsequent ore particle size detection will become more reliable and accurate, which can bring significant economic benefits in the fields of resource development, geological exploration, etc., can more effectively detect boundaries in ore images, improve segmentation accuracy, and lay the foundation for subsequent ore particle size distribution detection.

[0039] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0041] Figure 1 It is a flow chart of a method for segmenting an ore image by integrating multi-scale features of channel space provided by an embodiment of the present invention;

[0042] Figure 2 is an image of ore on a conveyor belt in a real scene provided by an embodiment of the present invention;

[0043] Figure 3 is an enhanced ore image provided by an embodiment of the present invention;

[0044] Figure 4 is an ore image label provided by an embodiment of the present invention;

[0045] Figure 5 It is the overall structure of the ResUNet-CSMA model provided by the embodiment of the present invention;

[0046] Figure 6 This is the data processing process of the CSMA model provided by the embodiment of the present invention;

[0047] Figure 7 1 is a comparison result diagram of the segmentation effects of three networks provided in an embodiment of the present invention, wherein (a) is the U-net segmentation result, (b) is the ResUNet segmentation result, and (c) is the ResUNet-CSMA segmentation result. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0050] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0051] In order to allow the segmentation model to obtain information in different receptive fields, the network can better handle ores of different shapes and sizes and complex scenes, and ensure that the model focuses on more important ore boundary features. In the present invention, a new channel spatial fusion attention mechanism (Channel and Spatial Attention, CSA) is designed, and the improved depthwise separable atrous spatial pyramid pooling (Depthwise Separable Atrous Spatial Pyramid Pooling, DS-ASPP) is added to form a new module - Channel and Spatial Multi-scale Attention (CSMA), which is introduced into the jump connection and bridge part of ResUNet to form a new network model to segment ore images.

[0052] The overall process of the present invention is as follows:

[0053] Firstly, ore images on a conveyor belt are collected from a real scene, and image enhancement operations such as histogram equalization and low-pass filtering are performed on the ore images. Then, data enhancement such as flipping, translation and rotation are performed on the enhanced images, and Labelme is used to annotate the ore images to form a new ore image data set; secondly, the ore image segmentation network of the present invention is built, and ResUNet is used as the basic network, and CSMA is introduced into its bridge and jump connection positions to form a new segmentation network; then, the built segmentation network is used to train the newly created ore image data set for multiple times, and the actual training parameters with the best effect are saved and retained; finally, the trained segmentation network model is used to predict the ore image that needs to be segmented, and the segmentation result is used to use the contour detection function in Opencv to realize the detection of ore particle size.

[0054] Embodiment 1

[0055] like Figure 1As shown, this embodiment provides a method for ore image segmentation integrating channel space multi-scale features, comprising the following steps:

[0056] Step 1: Obtain an image of ore on a conveyor belt in a real scene;

[0057] Step 2: Preprocess the ore image to obtain a preprocessed ore image data set;

[0058] In this embodiment, the ore image may usually contain noise or background interference. The key features of the ore (such as the texture, color, structure, etc. of the mineral) can be highlighted through image enhancement operations, making it easier to identify and extract useful information in the subsequent analysis process. The image enhancement operations include grayscale, histogram equalization, Gaussian filtering, etc. At the same time, annotation is performed on the basis of the original ore image to form a corresponding ore image label. Figure 2 , Figure 3 , Figure 4 The original ore images, ore images after image enhancement and corresponding labels of the dataset are shown respectively.

[0059] Step 3: Using the preprocessed ore image data set to train the ore image segmentation model to obtain a trained ore image segmentation model;

[0060] Traditional convolutional neural networks may not be able to pay full attention to all edge features when extracting mineral image features, resulting in the neglect of some important mineral detail information.

[0061] In order to allow the ore image segmentation model to obtain information in different receptive fields, the network can better handle ores of different shapes and sizes and complex scenes, and ensure that the model focuses on more important ore boundary features. In the present invention, a new channel and spatial fusion attention mechanism (CSA) is designed, and the improved depthwise separable atrous spatial pyramid pooling (DS-ASPP) is added to form a new module - Channel and Spatial Multi-scale Attention (CSMA), which is introduced into the jump connection and bridge part of ResUNet to form a new network model ResUNet-CSMA to segment ore images.

[0062] like Figure 5The figure shows the overall structure of the ResUNet-CSMA model. The specific structure of the model is as follows: The ResUNet-CSMA structure is based on the ResUNet architecture, but a CSMA module is added to ResUNet. It is mainly composed of an encoder, a decoder, a bridge, a jump link, a CSMA module, and an output layer.

[0063] The encoder is used to gradually extract the features of the input image. It consists of multiple convolutional layers, batch normalization layers (BatchNormalization) and activation functions (ReLU). The decoder is used to restore the spatial resolution of the image and generate the final segmentation result. The decoder of ResUNet uses upsampling to restore the size of the feature map to the size of the input image. In order to avoid the problem of gradient disappearance or gradient explosion, each module in the encoder and decoder contains a residual connection, which can alleviate the gradient disappearance problem of the deep network and better extract low-frequency information. The bridge is located between the encoder and the decoder, aiming to further extract features and is responsible for processing the feature map output by the encoder. The output of each decoding module will be spliced ​​with the feature map of the corresponding encoding module by skip connection, which can better preserve spatial information. ResUNet retains the symmetric structure and adds a CSMA structure to each skip connection between the encoder and the decoder to enhance the transmission of information and the fusion and learning of features. Finally, ResUNet maps the feature map output by the decoder to the number of categories of the target segmentation (ore image segmentation is binary classification) through a 1x1 convolution layer.

[0064] ResUNet data processing flow: First, the original image is preprocessed before training, including normalization, cropping or scaling, and data augmentation. During training, the data is input into the network in mini-batch mode. The data is processed in batches, and each batch of data contains multiple images and their corresponding labels.

[0065] The input image is extracted through the ResUNet encoder to obtain the initial feature map , after each layer of convolution and residual block, the spatial resolution of the image gradually decreases and the feature map gradually becomes richer. The CSMA module is added to the bridge part to further extract features. After the bridge, the decoder begins to gradually restore the spatial resolution of the image and fuses the low-level feature information into the high-level feature map through jump connections and CSMA modules. Finally, the feature map generated by the output layer is used to predict the segmentation result of the image. The cross entropy loss function is used to measure the difference between the model prediction and the true label. After the loss is calculated, the gradient is calculated by the back propagation algorithm, and the optimization algorithm is used to update the parameters of the network to reduce the loss function value. During the training process, the validation set is used to evaluate the performance of the model, and the hyperparameters (learning rate, batch size, etc.) are adjusted according to the results on the validation set.

[0066] like Figure 6 As shown, the data processing process of the CSMA module specifically includes the following steps:

[0067] Step 301, extracting a channel attention feature map of the ore image;

[0068] The initial feature map Perform maximum pooling and average pooling to generate channel descriptors, which are input into a shared multi-layer perceptron (MLP) respectively. The results of the two pooling are connected together and processed through a small fully connected network to generate the weight of each channel.

[0069] After the Sigmoid activation function, the output of the MLP is converted into a weight value between 0 and 1, and the channel attention output feature map is recorded as , the formula is as follows:

[0070] ,

[0071] in, Representative ore characteristic diagram, ,in C , H , W Represent the channel, depth and width of the input feature map respectively; , Express Perform maximum pooling and average pooling respectively. and Represent the maximum pooling feature and average pooling feature of each channel C of the ore image, represents a multi-layer perceptron, represents the Sigmoid activation function, and Represent two convolution operations respectively.

[0072] Step 302: Fusion of feature maps of ore images FAnd channel attention feature map Get channel fusion attention feature map :

[0073] ,

[0074] in, The feature map and Element-wise multiplication result.

[0075] Step 303: Based on the channel fusion attention feature map, global average pooling and maximum pooling are performed in the spatial dimension, and after the pooled features are spliced ​​in the channel dimension, a convolution operation is performed to obtain a channel space fusion attention feature map of the ore image;

[0076] Usually, in order to perceive the spatial feature information of the ore image, we first Global average pooling and maximum pooling are performed in the spatial dimension, and then the two pooled feature maps are concatenated in the channel dimension to form a new feature map. This process combines the features of average and maximum pooling to capture richer spatial information.

[0077] Perform convolution operation on the concatenated feature map, usually using The convolution kernel is used to learn the spatial attention map and form the final channel space fusion attention mechanism CSA. The formula is as follows:

[0078] ,

[0079] In the formula, Indicates that the convolution kernel size is The convolution operation, That is the overall output feature map of the CSA structure.

[0080] CSA improves the overall performance of the model by dynamically adjusting the weights of feature maps to enhance the representation of important features. Channel attention aims to capture the importance of different channels in feature maps. It enhances the network's attention to important features by learning the weights of each channel. Spatial attention focuses on the spatial information in the feature map and aims to assign higher weights to important spatial positions to enhance the model's sensitivity to feature positions. In the CSA module, it helps the model focus on significant areas in the input feature map, extracts significant ore edge features, and emphasizes those visually more prominent ore edge areas, thereby improving the model's representation ability. The combination of channel and spatial attention mechanisms provides a more comprehensive feature attention mechanism, which enables the model to consider the combination of local features and global information when judging feature importance, thereby improving the performance of ore segmentation tasks.

[0081] Step 304: The obtained channel space fusion attention feature map is subjected to a plurality of depth-separable dilated convolutions with different dilation rates to obtain a multi-scale feature map of the ore;

[0082] In order to integrate multi-scale information, DS-ASPP uses four The deep separable dilated convolution, APooling module (including adaptive mean pooling layer, 1×1 convolution layer and upsampling) and channel fusion module are used to enhance the ability of ore image feature extraction.

[0083] DS-ASPP can effectively improve computational efficiency, improve context capture capabilities and enhance model robustness. A smaller expansion ratio can reduce model complexity and reduce the risk of overfitting. Its formula is:

[0084] ,

[0085] in, is the output multi-scale feature map of the ore, It is a depth-separable dilated convolution using different dilation rates. represents the global average pooling module, The ASPP Pooling module is concatenated and fused with the output of each depth-separable atrous convolutional layer to introduce global context information. Indicates that a 1x1 convolutional layer is used for channel integration after concatenation.

[0086] Step 305: weight mapping the ore multi-scale feature map, and fusing it with the channel feature map to obtain a channel space multi-scale attention feature map, and obtain the ore edge contour feature;

[0087] Output feature map The output is normalized to the range of [0, 1] through the Sigmoid activation function, and we get ; The formula is as follows:

[0088] ,

[0089] in, Represents the Sigmoid activation function. Finally, the multi-scale weight coefficient of the output Fusion feature map with channels Perform element-by-element multiplication and assign different weights to the feature map to strengthen or suppress the features at the corresponding positions in the feature map:

[0090] ,

[0091] In the formula, represents element-wise multiplication; is the final output feature map.

[0092] In order to accurately segment the ore image, this embodiment proposes a densely connected U-type network based on the CSMA channel space multi-scale attention mechanism. The model adopts the encoder-decoder structure of ResUNet, which is mainly divided into an encoder module, a context extraction module and a decoder module. ResUNet uses residual blocks in the encoder and decoder, thereby improving the learning ability and training efficiency of the model. This embodiment adds the improved CSMA module to the ResUNet jump connection and bridge segment, and the improved structure can achieve higher accuracy in the ore image segmentation task.

[0093] Step 4: Perform contour detection on the ore image segmentation result to obtain the ore particle size detection result.

[0094] In this embodiment, contour detection is performed on the ore image segmentation result to obtain the ore particle size detection result, including: using OpenCV to perform contour detection on the ore segmentation result. The specific process includes image preprocessing, contour extraction, contour drawing and analysis. Contour detection can help us extract the shape characteristics of the ore and provide useful information for subsequent analysis and processing. The process of using OpenCV to perform contour detection on the ore segmentation result mainly includes image preprocessing, contour extraction and visualization contour and analysis. These steps help us extract the contour of the ore from the binary image and then analyze the morphological characteristics of the ore. First, load the image of the ore segmentation result. Usually, the ore segmentation result is a binary image with a foreground (ore part) of white and a background of black. The core step of contour detection is to extract contour information from the binary image, output the coordinate points of the contour, and the extracted contour is usually a list containing multiple coordinate points. Then the detected contour is drawn to the original image or a blank image to visualize the contour detection result. After the contour is detected, the area, perimeter, shape characteristics, etc. of the contour are further analyzed and calculated. These characteristics help to further analyze the shape of the ore.

[0095] By detecting the contours of each ore area in the image, their particle size characteristics can be further calculated. Particle size detection mainly includes the size of individual particles in the ore and the size distribution of the ore in the entire sample, which is quantified by the area, perimeter, diameter, aspect ratio, etc. of the ore. Finally, the particle size distribution of the particles is statistically divided into multiple particle size ranges, such as: less than 10 pixels, 10-50 pixels, 50-100 pixels, etc., and the particle size detection results are finally obtained.

[0096] The present invention selects U-net network and ResUNet network for comparative experiments. The segmentation comparison results of U-net, ResUNet and ResUNet-CSMA are shown in the figure. Figure 7 As shown, Figure 7 (a) is the U-net segmentation result. Figure 7 (b) is the ResUNet segmentation result. Figure 7 (c) is the ResUNet-CSMA segmentation result, which proves that the improved ResUNet-CSMA network solves the weak boundary segmentation problem. The obtained image segmentation result is closer to the label and is effective for ore segmentation.

[0097] Embodiment 2

[0098] This embodiment provides an ore image segmentation system integrating channel space multi-scale features, including:

[0099] An image acquisition module, which is used to acquire an ore image data set;

[0100] The image segmentation model training module is used to train the ore image segmentation model based on the ore image data set to obtain the trained ore image segmentation model; wherein the construction process of the ore image segmentation model includes:

[0101] The encoder extracts the initial feature map of the ore image; the CSMA module is added to the jump connection and bridge between the encoder and the decoder; the low-level feature information is fused into the high-level feature map through the CSMA module, including:

[0102] Based on the initial feature map of the ore image, extracting the channel attention feature map of the ore image;

[0103] The initial feature map of the ore image and the channel attention feature map are fused to obtain the channel fusion attention feature map; based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map; the channel space fusion attention feature map is subjected to a plurality of depth-separable dilated convolutions with different void rates to obtain the ore multi-scale feature map; the ore multi-scale feature map is weight-mapped and fused with the channel fusion attention feature map to obtain the ore edge contour feature map;

[0104] The image segmentation module is used for the ore image segmentation model to segment the ore image to obtain the segmentation result.

[0105] Embodiment 3

[0106] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the ore image segmentation method integrating channel space multi-scale features as described above are implemented.

[0107] Embodiment 4

[0108] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the ore image segmentation method integrating channel space multi-scale features as described above are implemented.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for ore image segmentation integrating multi-scale features of channel space, characterized in that: The steps include: Obtaining an ore image dataset; The ore image segmentation model is trained based on the ore image data set to obtain a trained ore image segmentation model; wherein the construction process of the ore image segmentation model includes: The encoder extracts the initial feature map of the ore image; the CSMA module is added to the jump connection and bridge between the encoder and the decoder; the low-level feature information is fused into the high-level feature map through the CSMA module, including: Based on the initial feature map of the ore image, extracting the channel attention feature map of the ore image; The initial feature map of the ore image and the channel attention feature map are fused to obtain the channel fused attention feature map; Based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map; The channel space fusion attention feature map is subjected to multiple depth-separable dilated convolutions with different dilation rates to obtain the multi-scale feature map of the ore; The multi-scale feature map of the ore is weight-mapped and fused with the channel fusion attention feature map to obtain the ore edge contour feature map; The ore image is segmented based on the trained ore image segmentation model to obtain the segmentation result.

2. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1 is characterized in that: After obtaining the ore image dataset, the ore images are preprocessed, including square image equalization and low-pass filtering image enhancement operations. Then, the enhanced images are flipped, translated and rotated for data enhancement, and Labelme is used to annotate the ore images to form a new ore image dataset.

3. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1, characterized in that: Based on the initial feature map of the ore image, the channel attention feature map of the ore image is extracted, including: The initial feature map of the ore image is subjected to maximum pooling and average pooling, and the results of maximum pooling and average pooling are input into a shared multi-layer perceptron. After two convolution operations, the maximum pooling and average pooling results are connected together, and then the output results of the shared multi-layer perceptron are converted into weight values ​​between 0 and 1 through the Sigmoid activation function, and the channel attention feature map is obtained through channel attention.

4. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1, characterized in that: The channel space fusion attention feature map is subjected to multiple depth-separable dilated convolutions with different dilation rates to obtain the multi-scale feature map of the ore, which is expressed as: ， in, is the output multi-scale feature map of the ore, It is a depth-separable dilated convolution using different dilation rates. represents the global average pooling module, The ASPP Pooling module is concatenated with the output of each depth-separable hole convolutional layer to introduce global context information. Indicates that a 1x1 convolutional layer is used for channel integration after concatenation.

5. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1, characterized in that: When the initial feature map and the channel attention feature map of the ore image are fused to obtain the channel fused attention feature map, the channel fused attention feature map is obtained by element-by-element multiplication of the initial feature map and the channel attention feature map.

6. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1, characterized in that: When extracting the initial feature map of the ore image, the original ore image is subjected to convolution and residual block operations to obtain the initial feature map.

7. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1, characterized in that: Based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map, including: performing global average pooling and maximum pooling on the channel fusion attention feature map in the spatial dimension, splicing the pooled features in the channel dimension, and performing a convolution operation to obtain the channel space fusion attention feature map of the ore image.

8. The ore image segmentation method integrating channel space multi-scale features as claimed in claim 1, characterized in that: The method further comprises using OpenCV to perform contour detection on the ore segmentation result to obtain the ore particle size detection result.

9. The ore image segmentation system integrating channel space multi-scale features is characterized by: include: An image acquisition module, which is used to acquire an ore image data set; The image segmentation model training module is used to train the ore image segmentation model based on the ore image data set to obtain the trained ore image segmentation model; wherein the construction process of the ore image segmentation model includes: The encoder extracts the initial feature map of the ore image; the CSMA module is added to the jump connection and bridge between the encoder and the decoder; the low-level feature information is fused into the high-level feature map through the CSMA module, including: Based on the initial feature map of the ore image, extracting the channel attention feature map of the ore image; The initial feature map of the ore image and the channel attention feature map are fused to obtain the channel fusion attention feature map; based on the channel fusion attention feature map, the spatial dimension information is captured to obtain the channel space fusion attention feature map; the channel space fusion attention feature map is subjected to a plurality of depth-separable dilated convolutions with different void rates to obtain the ore multi-scale feature map; the ore multi-scale feature map is weight-mapped and fused with the channel fusion attention feature map to obtain the ore edge contour feature map; The image segmentation module is used to segment the ore image based on the trained ore image segmentation model to obtain a segmentation result.

10. The ore image segmentation system integrating channel space multi-scale features as claimed in claim 9, characterized in that: In the image segmentation model training module, when the initial feature map and the channel attention feature map of the ore image are fused to obtain the channel fused attention feature map, the channel fused attention feature map is obtained by element-by-element multiplication of the initial feature map and the channel attention feature map.

Images