Coal gangue image segmentation method and system based on frequency domain decomposition and convolutional neural network

Abstract

The invention discloses a coal gangue image segmentation method and system based on frequency domain decomposition and a convolutional neural network, and aims to improve the segmentation precision of a coal gangue image. The method comprises the following steps: preprocessing an input coal gangue image to obtain a preprocessed image; performing frequency domain decomposition on the preprocessed image, and decomposing the image into three frequency domain groups, namely a low frequency domain group, an intermediate frequency domain group and a high frequency domain group; thirdly, performing frequency domain internal interaction in the frequency domain group, modeling a relationship among different frequency domain components, and improving the representation capability of each piece of frequency domain information; through interaction between the frequency domains, inter-channel weights of all the frequency domain groups are calculated, and feature fusion between the frequency domains is enhanced; and finally, segmenting the image after frequency domain interaction by using a convolutional neural network to obtain a coal gangue segmented image. By combining the advantages of the frequency domain decomposition and the convolutional neural network, the method can effectively improve the segmentation precision of the coal gangue area, is suitable for coal gangue segmentation tasks in complex backgrounds and noise environments, and has wide application prospects, especially in the fields of coal mine resource recovery and environmental protection.

Description

Technical Field

[0001] The present invention relates to a coal gangue detection and segmentation technology in a coal mining process, and specifically to a coal gangue image segmentation method and system based on frequency domain decomposition and convolutional neural network, which is suitable for efficient and accurate segmentation of coal gangue images to support coal mine production process monitoring, coal gangue quality analysis and other aspects. Background Art

[0002] Gangue is a byproduct of coal mining, usually composed of useless rocks, impurity minerals and other incombustible substances. Accurate identification and segmentation of gangue is of great significance for coal mining production, environmental monitoring and resource recovery. Existing gangue detection methods usually rely on traditional image processing technology, but these methods are not effective when dealing with complex backgrounds, noise interference and lighting changes. Therefore, there is an urgent need for a technical method that can accurately segment the gangue area. Summary of the invention

[0003] The present invention provides a coal gangue image segmentation method based on frequency domain decomposition and convolutional neural network. Combining frequency domain decomposition and convolutional neural network technology, it can effectively extract important features from coal gangue images and achieve efficient segmentation. This method not only improves the accuracy of segmentation, but also can work stably under complex conditions.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A method for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network, comprising:

[0006] Preprocessing the input coal gangue image to obtain a preprocessed image;

[0007] Performing frequency domain decomposition on the preprocessed image to decompose the image into three frequency domain groups: low frequency, medium frequency and high frequency;

[0008] Within the frequency domain group, internal frequency domain interaction is performed to improve the representation capability of each frequency domain information by modeling the relationship between different frequency domain components;

[0009] In the frequency domain, the frequency domain interaction is carried out, and the inter-channel weights of each frequency domain group are calculated to enhance the feature fusion between frequency domains;

[0010] The convolutional neural network is used to segment the image after frequency domain interaction to obtain the gangue segmentation image.

[0011] Optionally, the preprocessing of the input coal gangue image to obtain the preprocessed image specifically includes:

[0012] Grayscale the input image to obtain a grayscale image;

[0013] Perform denoising on the grayscale image to remove the noise in the image;

[0014] Perform contrast enhancement processing on the denoised image to improve the visual contrast of coal gangue;

[0015] The enhanced image is normalized for subsequent model processing.

[0016] Optionally, performing frequency domain decomposition on the preprocessed image to decompose the image into three frequency domain groups of low frequency, medium frequency and high frequency specifically includes:

[0017] Use discrete cosine transform (DCT) to convert the image from the spatial domain to the frequency domain;

[0018] The frequency domain image is grouped into low-frequency, medium-frequency and high-frequency channels;

[0019] In each channel group, feature representations of different frequency domain components are extracted to provide a basis for subsequent intra-frequency domain interactions and inter-frequency domain interactions.

[0020] Optionally, the performing of internal frequency domain interaction within the frequency domain group to improve the representation capability of each frequency domain information by modeling the relationship between different frequency domain components specifically includes:

[0021] Interact the feature channels in the high-frequency group and the low-frequency group to model the correlation between high-frequency texture information and low-frequency color information;

[0022] In the intermediate frequency group, the features in the intermediate frequency components are interactively modeled across dimensions to remove redundant noise in the frequency domain information;

[0023] An adaptive attention mechanism is used to assign weights to features within the frequency domain, retaining important features and improving segmentation accuracy.

[0024] Optionally, the performing frequency domain interaction between frequency domains and enhancing the feature fusion between frequency domains by calculating the inter-channel weights of each frequency domain group specifically includes:

[0025] The feature channels of the low-frequency group, high-frequency group, and medium-frequency group are concatenated to form a comprehensive frequency domain representation;

[0026] Average pooling and maximum pooling are performed on the spliced ​​frequency domain images to extract significant features.

[0027] Optionally, the using of a convolutional neural network to segment the image after frequency domain interaction to obtain a gangue segmentation image specifically includes:

[0028] Use convolutional layers to extract features from the image after frequency domain interaction;

[0029] Use the pooling layer to reduce the dimension of features and retain key information;

[0030] Use the upsampling layer to restore the image to its original resolution;

[0031] The output layer is used to generate a binary segmentation image of coal gangue and determine the specific area of ​​coal gangue in the image.

[0032] A method for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network, characterized by comprising:

[0033] A preprocessing module is used to preprocess the input coal gangue image to obtain a preprocessed image;

[0034] A frequency domain decomposition module is used to decompose the preprocessed image into three frequency domain groups: low frequency, medium frequency and high frequency;

[0035] The frequency domain internal interaction module is used to interactively process the features within the frequency domain group and enhance the representation capability of each frequency domain information;

[0036] The frequency domain interaction module is used to fuse the features of each frequency domain group and enhance the interaction information between frequency domains;

[0037] The segmentation network module is used to segment the image after frequency domain interaction to obtain the gangue segmentation image.

[0038] Optionally, the preprocessing module includes:

[0039] A grayscale unit, used for grayscale processing of an input image;

[0040] A denoising unit, used to remove noise from the image;

[0041] Enhancement unit, used to enhance image contrast and details;

[0042] The normalization unit is used to normalize the image for subsequent processing.

[0043] Optionally, the frequency domain decomposition module includes:

[0044] Discrete Cosine Transform (DCT) unit, used to convert the image from the spatial domain to the frequency domain;

[0045] The frequency domain grouping unit is used to decompose the frequency domain image into low-frequency, medium-frequency and high-frequency channel groups.

[0046] Optionally, the frequency domain internal interaction module includes:

[0047] High-frequency and low-frequency interaction units, used to model the relationship between high-frequency and low-frequency features;

[0048] The intermediate frequency interaction unit is used to process the intermediate frequency information, remove noise and enhance texture features.

[0049] Optionally, the frequency domain interaction module includes:

[0050] A frequency domain fusion unit is used to combine the features of the low frequency group, the high frequency group, and the medium frequency group;

[0051] Pooling unit, used to extract salient features;

[0052] Convolutional unit, used to calculate the weights between feature channels.

[0053] Optionally, the network segmentation module includes:

[0054] Convolutional layer, used to extract features after frequency domain interaction;

[0055] Pooling layer, used to reduce the dimension of features;

[0056] Upsampling layer, used to restore image resolution;

[0057] The output layer is used to generate gangue segmentation images.

[0058] The present invention discloses a gangue image segmentation method and system based on frequency domain decomposition and convolutional neural network, aiming to improve the segmentation accuracy of gangue images. The method comprises preprocessing the input gangue image to obtain a preprocessed image; then performing frequency domain decomposition on the preprocessed image to decompose the image into three frequency domain groups of low frequency, medium frequency and high frequency; then, performing internal frequency domain interaction within the frequency domain group to model the relationship between different frequency domain components and improve the representation ability of each frequency domain information; then, through the interaction between frequency domains, calculating the channel weights of each frequency domain group to enhance the feature fusion between frequency domains; finally, using a convolutional neural network to segment the image after frequency domain interaction to obtain a gangue segmentation image. By combining the advantages of frequency domain decomposition and convolutional neural network, the present invention can effectively improve the segmentation accuracy of the gangue area, is suitable for gangue segmentation tasks under complex background and noise environment, and has broad application prospects, especially in the fields of coal mine resource recovery and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 A flow chart of a method for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network provided in Example 1 of the present invention;

[0061] Figure 2 A convolutional neural network structure diagram provided in Example 1 of the present invention;

[0062] Figure 3 A structural diagram of a frequency domain decomposition module provided in Embodiment 2 of the present invention;

[0063] Figure 4 A structural diagram of a frequency domain internal interaction module provided in Embodiment 2 of the present invention;

[0064] Figure 5 A structural diagram of a frequency domain interaction module provided in Embodiment 2 of the present invention; DETAILED DESCRIPTION

[0065] 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.

[0066] The purpose of the present invention is to provide a coal gangue image segmentation method and system based on frequency domain decomposition and convolutional neural network, which combines frequency domain decomposition and convolutional neural network, and performs multi-layer processing on the coal gangue image to accurately segment the coal gangue area.

[0067] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] Example 1

[0069] Figure 1 A flow chart of a method for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network provided in Example 1 of the present invention is as follows: Figure 1 As shown, the method includes:

[0070] Step 101: preprocessing the input coal gangue image to obtain a preprocessed image. Step 101 specifically includes:

[0071] Step 1011: grayscale the input image to obtain a grayscale image.

[0072] Step 1012: Perform denoising processing on the grayscale image to remove noise in the image.

[0073] Step 1013: performing contrast enhancement processing on the denoised image to improve the visual contrast of the coal gangue.

[0074] Step 1014: Normalize the enhanced image to facilitate subsequent model processing.

[0075] Step 102: Perform frequency domain decomposition on the preprocessed image to decompose the image into three frequency domain groups: low frequency, medium frequency and high frequency. Step 102 specifically includes:

[0076] Step 1021: Convert the image from the spatial domain to the frequency domain using discrete cosine transform (DCT).

[0077] Step 1022: Group the frequency domain image into low-frequency, medium-frequency and high-frequency channels.

[0078] Step 1023: In each channel group, feature representations of different frequency domain components are extracted to provide a basis for subsequent intra-frequency domain interaction and inter-frequency domain interaction.

[0079] Step 103: In the frequency domain group, internal frequency domain interaction is performed to improve the representation capability of each frequency domain information by modeling the relationship between different frequency domain components. Step 103 specifically includes:

[0080] Step 1031: interacting the feature channels in the high-frequency group and the low-frequency group to model the correlation between the high-frequency texture information and the low-frequency color information.

[0081] Step 1032: In the intermediate frequency group, cross-dimensional interaction modeling is performed on the features in the intermediate frequency components to remove redundant noise in the frequency domain information.

[0082] Step 1033: Use an adaptive attention mechanism to assign weights to features within the frequency domain, retain important features, and improve segmentation accuracy.

[0083] Step 104: Performing frequency domain interaction between frequency domains, and enhancing the feature fusion between frequency domains by calculating the inter-channel weights of each frequency domain group. Step 104 specifically includes:

[0084] Step 1041: Concatenate the feature channels of the low frequency group, the high frequency group, and the medium frequency group to form a comprehensive frequency domain representation.

[0085] Step 1042: Perform average pooling and maximum pooling on the spliced ​​frequency domain images to extract significant features.

[0086] Step 105: Use a convolutional neural network to segment the image after frequency domain interaction to obtain a gangue segmentation image. Step 105 specifically includes:

[0087] Step 1051: Use a convolutional layer to extract features from the image after frequency domain interaction.

[0088] Step 1052: Use the pooling layer to reduce the feature dimension and retain key information.

[0089] Step 1053: Use an upsampling layer to restore the image to its original resolution.

[0090] Step 1054: Use the output layer to generate a binary segmentation image of the gangue and determine the specific area of ​​the gangue in the image.

[0091] The principle of the present invention is specifically described below:

[0092] Step 1: Separate high-frequency, medium-frequency and low-frequency information in the gangue image based on discrete cosine transform (DCT)

[0093] In the end-to-end image deblurring task, the main purpose of the frequency domain decomposition module is to effectively restore the clarity of the image by performing frequency domain processing on the input features, separating and reconstructing low-frequency, medium-frequency and high-frequency information. We decompose the frequency domain components of image features based on discrete cosine transform (DCT), and group high-frequency, low-frequency and medium-frequency components by assigning corresponding frequency coefficients to each feature channel, providing a basis for subsequent intra-frequency domain interaction and inter-frequency domain interaction. The process is as follows:

[0094] We first perform layer normalization on each input feature F to improve stability during training. A simple gating mechanism is introduced after the convolution layer to divide the feature map into two parts in the channel dimension and multiply the two parts, thereby simplifying the computational complexity and achieving an effect similar to a nonlinear activation function to obtain the feature F'. In this process, the gating mechanism greatly reduces the computational burden brought by the nonlinear activation function. The process is as follows:

[0095] Assuming that the number of channels of the input feature map is C, we divide it into two parts F1 and F2, and the number of channels of each part is C / 2.

[0096] Then, by multiplying F1 and F2 channel by channel, we get F'=F1×F2. This step can be regarded as a weighted combination of features, which effectively reduces the complexity. Through this simple gating operation, we can retain the important information in the feature map and effectively suppress redundant feature information.

[0097] For each feature channel, we transform it to the frequency domain through 2D discrete cosine transform (2D-DCT) to obtain the frequency component corresponding to each channel where [u,v] is the two-dimensional index of the corresponding frequency component, is the frequency coefficient at that position. The high-frequency information of the frequency component is concentrated in the lower right corner of the component matrix, while the low-frequency information is concentrated in the upper left corner.

[0098] We group the frequency components into high frequency groups F H , low frequency group FL and IF group F M We define:

[0099] High frequency group F H : corresponds to the highest frequency in the frequency component, usually containing image details and texture.

[0100] Low frequency group F L : corresponds to the lowest frequency in the frequency component, reflecting the continuous color of the image and the color information of a large area.

[0101] Intermediate frequency group F M : Located between high frequency and low frequency, it contains a small amount of texture and color information, and often appears as a mixture of texture and color in a blurred image.

[0102] Using the inverse discrete cosine transform (2D-IDCT), the three frequency groups (high frequency group F H , low frequency group F L and IF group F M ) is restored to the spatial domain, which lays a solid foundation for the subsequent frequency domain internal interaction module and inter-frequency domain interaction module, and can further improve the clarity and detail reconstruction capability of coal gangue images.

[0103] Step 2: Use the feature correlation within the frequency domain grouping to achieve internal interaction in the frequency domain

[0104] The main purpose of the frequency domain internal interaction module is to utilize the feature correlation within the frequency domain grouping, establish the interaction between the frequency feature channels, remove redundant information, and maintain effective high-frequency and low-frequency features, so as to enhance the image deblurring effect. H , low frequency group F L and IF group F M By performing effective interactive processing, the model can better restore details, texture and color information. The process is as follows:

[0105] After the image is convolved to obtain multiple feature representations, the feature channels show high similarity and correlation. In order to remove redundant information and improve model performance, we consider modeling the correlation between feature channels in the frequency domain group. The features of the high-frequency and low-frequency groups can interact with each other to retain the effective information and eliminate redundant or interfering features. We first transform the high-frequency group F H and low frequency group F L Merge and filter out the intermediate frequency information. The specific operations are:

[0106] F KL =cat(F H ,F L )

[0107] Thus, after the merger, FHL It contains both high-frequency and low-frequency information without interfering with each other.

[0108] We measure F along the W axis HL Rotate to get Where W represents the spatial dimension. Then, we Perform the saliency information pooling (SIP) operation. By performing maximum pooling and average pooling in the W dimension, it ensures rich representation of features and extracts key information. The specific operations are as follows:

[0109]

[0110] Max pooling helps us capture the texture features of the image, while average pooling can obtain more global information. The pooled result is compressed in the W dimension, which simplifies the calculation while ensuring the richness of the feature representation.

[0111] The results after pooling of salient information We further extract the relationship between channels through convolution operations and use the sigmoid activation function to obtain the channel weight A on the W dimension. w , then, the obtained weight A w Apply to On the left, rotate along the W axis to restore the original shape. The specific operation is:

[0112]

[0113] Similarly, we will HL Rotate along the H axis to get The same saliency information pooling (SIP) method is used to perform maximum pooling and average pooling operations in the H dimension to obtain the pooling result. Then, the channel weight A of the H dimension is calculated through the convolution layer and the sigmoid activation function. h The obtained weight A h Apply to On the top, rotate along the H axis to restore the original shape. The specific operation is:

[0114]

[0115] For the intermediate frequency group F M , we no longer need rotation operations, but directly calculate channel weights in the spatial domain. We also perform maximum pooling and average pooling on the mid-frequency group through the saliency information pooling (SIP) operation to extract its features. Then, we calculate the weights in the H and W dimensions through convolution and activation functions, and finally get C out .

[0116] After the interactive operation of W dimension, H dimension and C dimension, the obtained groups of feature channels will be stacked together to form the output FI of the internal interaction module in the frequency domain out , the formula for this process is:

[0117] FI out =Avg(W out +H out +C out )

[0118] Step 3: Establish effective feature interactions between different frequency domain groups

[0119] The purpose of the frequency domain interaction module is to establish effective feature interactions between different frequency domain groups. H , low frequency group F L and IF group F M The interaction and fusion between them enhances the information flow and feature fusion between different frequency domains, thereby improving the deblurring effect and clarity of the image. Through this module, the model can more accurately utilize the feature information of different frequency levels, remove redundant or irrelevant information, and achieve better detail reconstruction in the final image. The process is as follows:

[0120] We will come from the high frequency group F H , low frequency group F L and IF group F M The features of the image are spliced ​​to perform interactive operations between frequency domains. The spliced ​​features contain information from different frequency domains and can provide more diverse image feature representations. The specific operations are:

[0121] F HLM =concat(F H ,F L ,F M )

[0122] Next, we concatenate the feature tensor F HLM Calculate the channel attention. Specifically, firstly, the features are pooled through the Adaptive Average Pooling operation, and then the pooled features are processed through a 1×1 convolution layer to calculate the attention weight A of each channel. fa , we have the concatenated feature F HLM Weighted, we get the feature representation F of the attention between frequency domains out , the specific formula is as follows:

[0123] F Aout =A fa ·F HLM

[0124] We will FI out and FA out Aggregate and introduce the original feature map. The calculation formula is as follows:

[0125] F out =Avg(FA out +FI out +F)

[0126] The features after interaction between frequency domains and the original features of the low-frequency group ensure that the expression of important information in the image is not lost.

[0127] Step 4: Process the features after frequency domain interaction and fusion to generate the gangue segmentation image

[0128] The segmentation network module includes several key operations: convolution layer, pooling layer, upsampling layer and output layer. Each layer plays a vital role in the process of image feature extraction, dimensionality reduction, restoration and output. The specific operations are as follows:

[0129] At the beginning of the segmentation network module, the convolutional layer is first used to further process the features after frequency domain interaction. The convolution operation extracts local patterns and information from the input features by sliding windows. The convolutional layer can effectively capture the detailed features of the image, such as the edge, texture, and structure, while gradually extracting high-level semantic information from the image. Specifically, by stacking multiple convolutional layers, the model can gradually build a deep feature representation from low-level features to high-level semantic features. The specific operations are as follows:

[0130] F conv =Conv(F input )

[0131] F input is the feature after frequency domain interaction, F conv is the extracted feature output of the convolutional layer.

[0132] The pooling layer usually follows the convolution layer. Its purpose is to reduce the dimensionality of features and reduce the spatial dimensions while retaining the most important feature information. The pooling operation can help the model reduce the amount of calculation, improve computational efficiency, and prevent overfitting. The pooling layer reduces the size of the feature map while retaining the importance of the features. The specific operation is as follows:

[0133] F pool =Pool(F conv )

[0134] F out It is the dimension reduction feature after the pooling operation. The model can compress the features through this stage while maintaining sensitivity to key features.

[0135] The function of the upsampling layer is to restore the resolution of the image and restore the feature map that has been reduced in dimension after the pooling operation to the size of the original image or close to the original size. Through upsampling, the model can reconstruct spatial information and produce a more detailed image structure. The specific operations are as follows:

[0136] F upsample =Upsample(F pool )

[0137] After convolution, pooling, and upsampling, the model needs to generate the final segmented image through the output layer. The output layer is a convolution layer with multiple channels, each channel corresponds to a category. The output layer calculates the category probability of each pixel and finally generates a binary image to represent the location of the gangue in the image. The shape of the output layer is the same as the shape of the input image, but the value of each pixel represents the category to which it belongs. The specific operation is as follows:

[0138] F out =Conv(F upsample )

[0139] F out It is the final output gangue segmentation image, in which the background pixel value is 0 and the gangue pixel value is 1.

[0140] Example 2

[0141] This embodiment provides a method for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network, including:

[0142] The preprocessing module is used to preprocess the input coal gangue image to obtain a preprocessed image.

[0143] The frequency domain decomposition module is used to decompose the preprocessed image into three frequency domain groups: low frequency, medium frequency and high frequency.

[0144] The frequency domain internal interaction module is used to interactively process the features within the frequency domain group and enhance the representation capability of each frequency domain information.

[0145] The frequency domain interaction module is used to fuse the features of each frequency domain group and enhance the interaction information between frequency domains.

[0146] The segmentation network module is used to segment the image after frequency domain interaction to obtain the gangue segmentation image.

[0147] In this embodiment, the preprocessing module specifically includes:

[0148] A grayscale unit, used for grayscale processing the input image to obtain a grayscale image;

[0149] A denoising unit, used for denoising the grayscale image to remove noise in the image;

[0150] An enhancement unit is used to perform contrast enhancement processing on the denoised image to improve the visual contrast of the coal gangue;

[0151] The normalization unit is used to normalize the enhanced image for subsequent model processing.

[0152] In this embodiment, the frequency domain decomposition module specifically includes:

[0153] Discrete Cosine Transform (DCT) unit, used to convert the image from the spatial domain to the frequency domain;

[0154] The frequency domain grouping unit is used to decompose the frequency domain image into low-frequency, medium-frequency and high-frequency channel groups.

[0155] In this embodiment, the frequency domain internal interaction module specifically includes:

[0156] High-frequency and low-frequency interaction units, used to model the relationship between high-frequency and low-frequency features;

[0157] The intermediate frequency interaction unit is used to process the intermediate frequency information, remove noise and enhance texture features.

[0158] In this embodiment, the frequency domain interaction module specifically includes:

[0159] A frequency domain fusion unit is used to combine the features of the low frequency group, the high frequency group, and the medium frequency group;

[0160] Pooling unit, used to extract salient features;

[0161] Convolutional unit, used to calculate the weights between feature channels.

[0162] In this embodiment, the network segmentation module specifically includes:

[0163] Convolutional layer, used to extract features after frequency domain interaction;

[0164] Pooling layer, used to reduce the dimension of features;

[0165] Upsampling layer, used to restore image resolution;

[0166] Output layer: used to generate gangue segmentation images.

[0167] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0168] (1) The present invention separates the different frequency information in the gangue image through frequency domain decomposition, which significantly improves the accuracy and efficiency of image segmentation. After adopting low-frequency, medium-frequency and high-frequency decomposition, the complexity of the gangue image is effectively reduced, and the noise interference and boundary blur problems in traditional image processing methods are reduced, thereby improving the robustness and stability of the segmentation model.

[0169] (2) The present invention converts the input image from the spatial domain to the frequency domain through discrete cosine transform (DCT), and divides it into three groups: low frequency, medium frequency and high frequency according to the characteristics of the frequency domain features. For each frequency domain group, the present invention uses 2D-IDCT to restore the decomposed frequency domain information to the spatial domain, providing rich feature expressions for subsequent segmentation operations. This frequency domain decomposition and restoration method effectively improves the accuracy of coal gangue image segmentation, especially in the case of complex background and more noise, and can significantly improve the reliability of the segmentation effect.

[0170] (3) The present invention optimizes the fusion and expression of frequency domain features by designing the internal frequency domain interaction module and the inter-frequency domain interaction module, and makes full use of the complementarity of different frequency domain features. The internal frequency domain interaction module enhances the retention of image details by modeling the feature relationship between different channels in the frequency domain, while the inter-frequency domain interaction module further enhances the feature fusion between low-frequency, high-frequency and medium-frequency information by calculating the channel weights of each frequency domain group, thereby improving the accuracy of image segmentation and adaptability to complex backgrounds.

[0171] (4) The method and system for gangue image segmentation based on frequency domain decomposition and convolutional neural network proposed in the present invention are not only suitable for automatic identification and segmentation of gangue in the coal mining field, but can also be widely used in other image segmentation tasks, especially in processing complex background, low contrast and high noise images, with strong adaptability. The combination of frequency domain feature decomposition and deep learning makes the method have good generalization ability and can achieve efficient and accurate image segmentation in different application scenarios.

[0172] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0173] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method and system for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network, characterized in that: include: Preprocessing the input coal gangue image to obtain a preprocessed image; Perform frequency domain decomposition on the preprocessed image and decompose the image into three frequency domain groups: low frequency, medium frequency and high frequency; Within the frequency domain group, internal frequency domain interaction is performed to improve the representation capability of each frequency domain information by modeling the relationship between different frequency domain components; In the frequency domain, the frequency domain interaction is carried out, and the inter-channel weights of each frequency domain group are calculated to enhance the feature fusion between frequency domains; The convolutional neural network is used to segment the image after frequency domain interaction to obtain the gangue segmentation image.

2. The method for segmenting coal gangue images according to claim 1, characterized in that: The step of preprocessing the input gangue image to obtain the preprocessed image specifically includes: Grayscale the input image to obtain a grayscale image; Perform denoising on the grayscale image to remove the noise in the image; Perform contrast enhancement processing on the denoised image to improve the visual contrast of coal gangue; The enhanced image is normalized for subsequent model processing.

3. The gangue image segmentation method according to claim 1, characterized in that: The frequency domain decomposition of the preprocessed image is performed to decompose the image into three frequency domain groups: low frequency, medium frequency and high frequency, specifically including: Use discrete cosine transform (DCT) to convert the image from the spatial domain to the frequency domain; The frequency domain image is grouped into low-frequency, medium-frequency and high-frequency channels; In each channel group, feature representations of different frequency domain components are extracted to provide a basis for subsequent intra-frequency domain interactions and inter-frequency domain interactions.

4. The method for segmenting coal gangue images according to claim 1, characterized in that: The frequency domain internal interaction is performed within the frequency domain group, and the representation capability of each frequency domain information is improved by modeling the relationship between different frequency domain components, specifically including: Interact the feature channels in the high-frequency group and the low-frequency group to model the correlation between high-frequency texture information and low-frequency color information; In the intermediate frequency group, the features in the intermediate frequency components are interactively modeled across dimensions to remove redundant noise in the frequency domain information; An adaptive attention mechanism is used to assign weights to features within the frequency domain, retaining important features and improving segmentation accuracy.

5. The method for segmenting coal gangue images according to claim 1, characterized in that: The frequency domain interaction is performed between frequency domains, and the feature fusion between frequency domains is enhanced by calculating the inter-channel weights of each frequency domain group, specifically including: The feature channels of the low-frequency group, high-frequency group, and medium-frequency group are concatenated to form a comprehensive frequency domain representation; Average pooling and maximum pooling are performed on the spliced ​​frequency domain images to extract significant features.

6. The method for segmenting coal gangue images according to claim 1, characterized in that: The method of using a convolutional neural network to segment the image after frequency domain interaction to obtain a gangue segmentation image specifically includes: Use convolutional layers to extract features from the image after frequency domain interaction; Use the pooling layer to reduce the dimension of features and retain key information; Use the upsampling layer to restore the image to its original resolution; The output layer is used to generate a binary segmentation image of coal gangue and determine the specific area of ​​coal gangue in the image.

7. The method for segmenting coal gangue images according to claim 6, characterized in that: The convolutional neural network adopts a residual network structure, which specifically includes: Use multiple residual blocks in the network to solve the gradient vanishing problem in deep network training; Each residual block consists of multiple convolutional layers and skip connections to optimize information flow and feature learning.

8. A method and system for segmenting coal gangue images based on frequency domain decomposition and convolutional neural network, characterized in that: include: A preprocessing module is used to preprocess the input coal gangue image to obtain a preprocessed image; A frequency domain decomposition module is used to decompose the preprocessed image into three frequency domain groups: low frequency, medium frequency and high frequency; The frequency domain internal interaction module is used to interactively process the features within the frequency domain group and enhance the representation capability of each frequency domain information; The frequency domain interaction module is used to fuse the features of each frequency domain group and enhance the interaction information between frequency domains; The segmentation network module is used to segment the image after frequency domain interaction to obtain the gangue segmentation image.

9. The method for segmenting coal gangue images according to claim 8, characterized in that: The preprocessing module comprises: A grayscale unit, used for grayscale processing of an input image; A denoising unit, used to remove noise from the image; Enhancement unit, used to enhance image contrast and details; The normalization unit is used to normalize the image for subsequent processing.

10. The method for segmenting coal gangue images according to claim 8, characterized in that: The frequency domain decomposition module comprises: Discrete Cosine Transform (DCT) unit, used to convert the image from the spatial domain to the frequency domain; The frequency domain grouping unit is used to decompose the frequency domain image into low-frequency, medium-frequency and high-frequency channel groups.

11. The method for segmenting coal gangue images according to claim 8, characterized in that: The frequency domain internal interaction module includes: High-frequency and low-frequency interaction units, used to model the relationship between high-frequency and low-frequency features; The intermediate frequency interaction unit is used to process the intermediate frequency information, remove noise and enhance texture features.

12. The method for segmenting coal gangue images according to claim 8, characterized in that: The frequency domain interaction module includes: a frequency domain fusion unit, which is used to splice the features of the low frequency group, the high frequency group and the medium frequency group; Pooling unit, used to extract salient features; Convolutional unit, used to calculate the weights between feature channels.

13. The method for segmenting coal gangue images according to claim 8, characterized in that: The segmentation network module includes: a convolution layer for extracting features after frequency domain interaction; Pooling layer, used to reduce the dimension of features; Upsampling layer, used to restore image resolution; The output layer is used to generate gangue segmentation images.

Images