A static feature extraction method for froth images of coal slime flotation

Abstract

The invention discloses a static feature extraction method for a coal slime flotation foam image, comprising: obtaining a coal slime flotation foam video, and manually marking a plurality of frames of images to generate a two-dimensional matrix with the same resolution as the image. Convolution smoothing is performed on a Gaussian kernel point set to generate a foam density map, and the map is standardized. A density regression network is designed and trained to predict the density of the coal slime flotation foam and realize regional division. The binary edge map of significant bubbles is converted into an edge confidence map, an edge regression network is designed and trained, and an adaptive breakpoint continuation algorithm is used to optimize the edge prediction result to obtain complete significant bubble edge information. Finally, the density map and the edge regression map are combined to extract and analyze the density, boundary and spatial distribution characteristics of the foam. The invention can comprehensively extract the static features of the coal slime flotation foam image, and provide effective support for monitoring and optimization of the flotation process.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a static feature extraction method for coal slime flotation foam images. Background Art

[0002] In the context of the country's vigorous development of new quality productivity, intelligent monitoring and control of flotation has become one of the important scenarios for the intelligentization of coal washing and processing. Specifically, the state of coal slime flotation foam can directly reflect the stability of the flotation process and product quality. Therefore, analyzing the flotation foam state based on images is an important part of flotation intelligent technology.

[0003] Since the flotation foam of coal slime is highly dynamic and has complex structural characteristics, especially in complex scenarios such as high-concentration coal slime water environment and different foam formation conditions, the static characteristics of the flotation foam of coal slime are more complex and changeable, which further increases the difficulty of processing. Although the traditional method can achieve a certain degree of feature extraction under laboratory conditions, in actual complex working conditions, due to changes in ambient lighting, irregular foam morphology, color diversity and other issues, the accuracy and stability of foam image segmentation and feature extraction are greatly reduced.

[0004] In recent years, with the development of image processing technology and artificial intelligence, coal slime flotation foam feature extraction methods based on image analysis have gradually become a research hotspot. Through image analysis technology, these methods can automatically extract and analyze the morphology, distribution, size and other characteristics of coal slime foam, and provide quantitative data support. However, existing research mainly focuses on laboratories or relatively simple environments, and there are still relatively few studies on coal slime foam feature extraction under complex working conditions. Especially in actual production, factors such as changes in lighting conditions, diverse foam morphology, and complex colors have put forward higher requirements on the stability and robustness of image processing algorithms. Therefore, there is an urgent need for a method that can stably and accurately extract the static characteristics of coal slime foam in complex scenarios, which can provide real-time and accurate data support for foam control, thereby optimizing the foam control strategy and accelerating the realization of intelligent coal slime flotation systems. Summary of the invention

[0005] The present invention aims at the defects of the prior art and provides a static feature extraction method of coal slime flotation foam image.

[0006] In order to achieve the above invention object, the technical solution adopted by the present invention is as follows:

[0007] A static feature extraction method for coal slime flotation foam image comprises the following steps:

[0008] S1. Obtain image data: Obtain a video of coal slime flotation foam under normal production conditions in a coal preparation plant; Take several frames of images from the video for manual point annotation, abstract the continuous foam group into a discrete point set, and generate a two-dimensional matrix of the same size as the image resolution;

[0009] S2. Generate a foam density map: Generate a sparse density matrix according to the relative positions of bubbles in the map; Use a Gaussian kernel to perform convolution smoothing on the sparse matrix to obtain a foam density map;

[0010] S3. Deep learning network training: Standardize the foam density map, including density value standardization and map size standardization; design a density regression network with an encoder-decoder structure, and introduce a loss selective discarding mechanism; use the processed foam density map as a supervised learning target, train the density regression network, and realize the prediction of the density of coal slime flotation foam and regional division;

[0011] S4. Segmentation of significant bubbles: According to the prediction results of the density of coal slime flotation foam, the significant bubbles in the non-dense area are manually annotated at the pixel level, and the binary edge annotation results of the significant bubbles are converted into edge confidence maps through the expansion of boundary pixels and the calculation of probability values;

[0012] Design an edge regression network with an encoder-decoder structure and auxiliary decoding branch capable of multi-resolution feature extraction, and train it with edge confidence maps as supervision targets;

[0013] According to the predicted edge probability value, a binary edge map is generated through a threshold, and the breakpoints are connected using the skeleton extraction, eight-neighborhood method breakpoint search and point-bubble matching methods to obtain complete significant bubble edge information;

[0014] S5. Feature extraction and analysis:

[0015] Based on the output of the density regression network, the features of the dense foam area are extracted to analyze the density and distribution characteristics of the foam;

[0016] Based on the output of the edge regression network, the spatial distribution, quantity, and size characteristics of significant bubbles are extracted;

[0017] The density map and the marginal regression map are combined to form a static feature description of the foam image, including the statistical values ​​of foam shape, surface texture, loading rate, and the cumulative distribution data of foam size and distribution.

[0018] Furthermore, the density regression network includes:

[0019] Encoder: Uses HRNet module to maintain multi-resolution feature maps and fuse semantic information of different resolutions;

[0020] Decoder: Use the deconvolution module to enlarge the resolution of the network output and extract the details of the bubble boundaries;

[0021] Loss selective discarding mechanism: remove abnormal loss values ​​and optimize network training.

[0022] Furthermore, the density regression network is trained as follows: the foam image and the generated foam density map are input into the density regression network for training; the density regression network outputs a predicted density map, and by summing all density values, the number of bubbles in the image, the area ratio between large and small bubbles, and their spatial distribution are obtained.

[0023] Furthermore, significant bubbles need to be detected in the significant bubble segmentation, and the significant bubbles in the non-dense area of ​​the image need to be detected using an object detector; the bounding box of each significant bubble is marked, and the image area of ​​the significant bubble is cropped along the bounding box;

[0024] Perform binary edge annotation at the semantic segmentation level for each salient bubble; calculate the edge confidence map through Gaussian distribution and normalize it to the range of [0,1].

[0025] Furthermore, calculating the edge confidence map specifically includes:

[0026] Choose mean μ and standard deviation σ to define the Gaussian kernel;

[0027] Calculate the Gaussian distribution for each pixel point x in the vertical direction of the edge center i :

[0028]

[0029] Among them, E(x i ) represents pixel x i The edge confidence of x i The value of is the distance of the pixel from the center pixel on the edge;

[0030] Find the edge confidence map E(x i ) in the maximum value E max ;

[0031] Normalize the values ​​in the edge confidence map to the range [0,1]:

[0032]

[0033] The normalized edge confidence map E norm (x) Save as image.

[0034] Furthermore, the edge regression network includes: based on the HRNet-V2 module, using the channel attention module and the spatial attention module to extract and fuse features; using the boundary confidence map to train the network and output the predicted edge confidence map.

[0035] Furthermore, the value of the points marked in the two-dimensional matrix is ​​1, and the values ​​of the other positions are 0; if at pixel x i If there are bubbles at , the matrix annotated with N bubbles is expressed as:

[0036]

[0037] Where N is the total number of bubbles detected in the image, x is a certain position in the image, and x i is the center position of the ith bubble, δ(xx i ) is the indicator function, when x=x i The value is 1 when the condition is true, otherwise it is 0.

[0038] Furthermore, the generation formula of the foam density map is:

[0039] D′(x)=D(x)*G(x)

[0040] Among them, G(x) is the Gaussian kernel function and * represents the convolution operation.

[0041] Furthermore, the deconvolution module consists of a 4×4 deconvolution layer, a batch normalization layer, a ReLU activation function, and a residual block.

[0042] Furthermore, the steps of the breakpoint continuation algorithm include: detecting the breakpoints in the edge graph, finding the eight neighborhoods of the breakpoints, matching the breakpoints with the connected areas to which they belong, connecting the breakpoints, and obtaining complete edge information.

[0043] Compared with the prior art, the advantages of the present invention are:

[0044] 1. Accurate processing for complex production scenarios: The present invention takes into account the complex situation of coexistence of dense foam and significant bubbles in the coal slime flotation foam image, adopts a divide-and-conquer strategy to accurately separate dense foam from significant bubbles, and provides a complete feature extraction idea to improve the accuracy and reliability of feature extraction.

[0045] 2. Application of density regression method: In view of the difficulty in accurately extracting the characteristics of each small bubble in the dense foam area, the present invention introduces a density regression method to uniformly analyze the characteristics of the dense foam area, avoiding the difficulties of traditional algorithms in extracting small bubble features. This method uses a density regression network to perform high-resolution feature extraction, achieving accurate estimation of the number and distribution of bubbles in the dense foam area.

[0046] 3. Improved edge regression method: The present invention transforms the extraction of significant bubbles from the traditional "region segmentation" method to the "edge regression" method. The edge regression network solves the problems of low boundary segmentation accuracy, easy adhesion of connected regions, and unstable segmentation effect in traditional methods by generating and training edge confidence maps, and improves the accuracy and integrity of significant bubble boundaries.

[0047] 4. Multi-scale information characterization: The present invention combines the results of density regression and edge regression, and adopts a multi-scale information characterization method to calculate and combine large-scale statistical characteristic values ​​(such as the color, shape, surface texture and load rate of foam groups and single bubbles), medium-scale cumulative distribution, and small-scale density maps and edge regression maps. This method provides static feature descriptions at three scales of "statistics + cumulative distribution + regression map", and comprehensively characterizes the characteristics of flotation foam images.

[0048] 5. Adaptive breakpoint continuation algorithm: To address the problem of abnormal breakpoints that occur during the significant bubble segmentation process, the present invention adopts an adaptive breakpoint continuation algorithm to search for breakpoints and their possible continuity directions through the eight-neighborhood method, match the breakpoints and their connected area relationships, achieve the connection of the breakpoints, and ensure the integrity of the significant bubble boundaries.

[0049] 6. Improved network structure: The present invention adopts HRNet module, deconvolution module, channel attention module and spatial attention module in the density regression network and edge regression network, optimizes the extraction and fusion of high-resolution features, and improves the prediction accuracy and training stability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of a static feature extraction method of coal slime flotation foam image according to an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of a foam density map generation process provided by an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of an encoder portion in a density regression network provided by an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of a decoder deconvolution module in a density regression network provided in an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of Gaussian kernel function operation provided by an embodiment of the present invention;

[0055] Figure 6 A schematic diagram of the edge confidence map generation process provided by an embodiment of the present invention;

[0056] Figure 7 A schematic diagram of the froth edge confidence of coal slime flotation provided by an embodiment of the present invention;

[0057] Figure 8 A schematic diagram of a boundary regression network provided by an embodiment of the present invention;

[0058] Fig. 9 A schematic diagram of the multi-scale information fusion characterization process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0060] A static feature extraction method for coal slime flotation foam image includes three parts: a foam density area division unit, a significant bubble segmentation unit and a multi-scale information representation unit.

[0061] Firstly, the froth image of coal slime flotation in complex production scenarios must be a mixed state of dense foam and significant bubbles. The foam density area division unit is used to divide the foam dense area and the non-dense area according to the density of the foam in the picture.

[0062] The core of the foam density area division unit is the foam density map and the density regression network. The foam density map is used as supervised learning information to train the density regression network.

[0063] The position information of the bubbles in the image is obtained by including but not limited to manual annotation (such as point annotation) or automatic generation (such as corner point detection), and the relative position relationship between the bubbles is quantitatively expressed by the number of bubbles in a unit area to obtain a foam density map.

[0064] The density regression network needs to include three links: encoder, decoder and loss selective discarding mechanism, which can be used to predict the density distribution of the newly input coal slime flotation foam image. The encoder is used to maintain multi-resolution feature maps in parallel throughout the network and fuse semantic information of different resolutions to achieve high-resolution feature extraction of the foam density map; the decoder extracts more details of the bubble boundary by enlarging the resolution of the network output, making the predicted density map smoother; in this process, inevitable labeling errors and other reasons will produce abnormal density, which in turn affects the training of the density regression network. Therefore, it is necessary to selectively discard the abnormal loss caused by abnormal density to optimize the network training process.

[0065] Finally, after the density map predicted by the network is normalized by labels, the total number of bubbles in the foam image can be obtained by adding up all the pixel values ​​(i.e., density values). The area ratio and spatial distribution of dense and non-dense foam areas can be obtained by dividing based on the density threshold.

[0066] Secondly, the significant bubble segmentation unit needs to rely on the results of the foam density area division unit, that is, for the significant bubbles in the non-dense area, the problem of unstable effect of the traditional foam segmentation method is solved based on the edge regression concept.

[0067] The core of the salient bubble segmentation unit is the edge confidence map and the edge regression network. The edge confidence map is used as supervised learning information to train the edge regression network.

[0068] Manual annotation is performed based on the idea of ​​semantic segmentation to obtain the binary edges of significant bubbles in non-dense areas, and the binary edges are converted into a probability map of the surrounding area belonging to the edge through kernel functions including but not limited to Gaussian kernel function and uniform kernel function, thereby obtaining the edge confidence map of the significant bubble group.

[0069] The edge regression network contains an Encoder-Decoder structure with multi-resolution feature extraction capabilities and an auxiliary decoding branch. During the feature generation and fusion process, the channel attention module is used to coordinate the high / low resolution features of the encoder part, and the weights of features of different resolutions can be calculated by any channel attention method; similarly, in the decoder part, the spatial attention module is used to adaptively fuse high / low resolution features, making its output smoother than connecting features of different resolutions; considering that low-resolution features have reliable semantic information, a gating signal is designed in the low-resolution features to achieve the regression of high-resolution boundary confidence maps; the auxiliary branch helps the network find a more robust feature representation by introducing auxiliary learning, which is beneficial to the target task.

[0070] Finally, the edge confidence map predicted by the network is processed into a binary map of the edge of significant bubbles through adaptive threshold screening and skeleton extraction. However, abnormal breakpoints will inevitably appear in the binary edges of the result, affecting the division and extraction of significant bubbles. Therefore, the breakpoints and their possible continuous directions are found based on the eight-neighborhood method, and the breakpoints are connected to obtain complete edge information by matching the relationship between the breakpoints and their connected areas, thereby obtaining accurate and complete significant bubbles.

[0071] Combining the first two aspects, the multi-scale information characterization unit calculates the statistical characteristic values ​​(large scale) and their cumulative distribution (medium scale) including but not limited to the color, shape, surface texture and carrying capacity of foam groups and single bubbles based on the results of dense foam density regression and significant bubble edge regression, and combines the density map and edge regression map (small scale) to form a static feature of the foam image using three scales of "statistical value + cumulative distribution + regression map".

[0072] like Figure 1 As shown, in order to achieve the above purpose, the present invention also adopts the following steps to analyze and detect the foam state of any frame in a section of coal slime flotation foam video:

[0073] Step S1: taking a coal slime flotation foam image of any single frame in a coal slime flotation foam video, and generating a foam density map according to the relative positions of bubbles in the image;

[0074] Step S2: inputting the processed foam density map into a density regression network for training to obtain a predicted density map, thereby achieving the characterization of the positional relationship between bubbles and the division of the foam density area in the map;

[0075] Step S3: for the significant bubbles in the non-dense area, convert the binary edges of the significant bubbles into a probability map to generate an edge confidence map of the significant bubble group;

[0076] Step S4: input the edge confidence map into the edge regression network for training to obtain a predicted edge map, and use an adaptive breakpoint continuation algorithm to obtain a high-confidence significant bubble segmentation result;

[0077] Step S5: According to the characteristics of dense / significant areas, the relevant feature extraction operators are optimized respectively, the global statistical values ​​(several scalars) and cumulative distributions (several groups of one-dimensional vectors) of all areas are calculated, and the density map and edge regression map (two-dimensional matrix) are combined to finally obtain the static characteristics of coal slime foam under three description scales of "large, medium and small".

[0078] Under normal production conditions in a coal preparation plant, a video of coal flotation foam was obtained. Several frames of images were taken for manual point annotation, and a two-dimensional matrix with the same size as the image resolution was generated. The value of the annotated points in the matrix is ​​1, and the values ​​of the remaining positions are 0. If the pixel x i If there are bubbles at , the matrix annotated with N bubbles will be expressed as:

[0079]

[0080] Where [a,b] represents the vector space of the annotated binary image. The above equation is used to generate the density map, and a very sparse matrix will be obtained. To solve the problem of sparse image matrix, a Gaussian kernel is introduced to convolve and smooth the density map, and the final foam density map is as follows: Figure 2 shown.

[0081] The density regression network can be improved by any convolutional neural network with an encoder-decoder structure. In a preferred embodiment, the density regression network consists of an encoder, a decoder and a loss selective discarding mechanism, such as Figure 3 The encoder uses the HRNet module to maintain multi-resolution feature maps in parallel throughout the network, and fuses semantic information of different resolutions to achieve high-resolution feature extraction of the foam density map; the decoder uses the deconvolution module to amplify the resolution of the network output and extract more details of the bubble boundary, making the predicted density map smoother; the loss selective discarding mechanism can remove abnormal loss values ​​from the network training process, reducing the negative impact of abnormal losses.

[0082] In order to make the details of the density map regression as accurate as possible, the encoder needs to achieve high-resolution feature output. Figure 3 In a preferred embodiment, the HRNet module starts with a backbone and reduces the resolution to 1 / 4 of the input image. From the first stage to the third stage, a new branch is forked from the current branch in parallel at 1 / 2 of the lowest resolution. After four stages of feature extraction, the three branches are enlarged to the resolution of the high-resolution branch through bilinear upsampling, and the representations of all branches are concatenated and used as the output of the encoder.

[0083] In order to extract more details of dense foam and better characterize the spatial distribution of bubbles, it is recommended to enlarge the resolution of the network output. The bilinear upsampling strategy is too coarse for the regression task, resulting in tearing of bubble boundaries in the predicted density map and biasing the estimated spatial density of bubbles. In a preferred embodiment, a deconvolution module is designed as a decoder to make the predicted density map smoother. Figure 3 As shown in the figure, the deconvolution module consists of a 1×1 convolution, a deconvolution part, a residual block, and a bilinear upsampling operation. The input of the deconvolution module is the output feature map of the encoder part, and its output is a single-channel image (density map) with the size of the input image. Specifically, a 4×4 deconvolution is used, followed by batch normalization and ReLU, the resolution of the feature map is enlarged from 1 / 4 to 1 / 2 (the size of the input image), four residual blocks are stacked to refine the enlarged feature map, 1×1 convolution is added to adjust the number of channels, and bilinear upsampling is used to enlarge the resolution to the size of the input image, making the predicted density map smoother.

[0084] The foam image and density map are input into the density regression network for training to obtain the predicted density map; all predicted density values ​​are added together to obtain the number of bubbles in the foam image, the area ratio between large and small bubbles, and their spatial distribution.

[0085] For the segmentation of significant bubbles in non-dense areas, the image can be processed globally at the same time, or independent significant bubbles can be detected first and then processed one by one. Figure 4 As shown, in a preferred embodiment, the coal slime foam image is detected by an object detector (including but not limited to the YOLO series and the Mask-RCNN series) to detect significant bubbles in the non-dense area, and the area and position of each bubble are marked by a bounding box. After the bubble image is cropped along the bounding box, the binary edge annotation at the semantic segmentation level is performed, and then the pixels on the edge are regarded as the center, and the Gaussian distribution is calculated along its vertical direction. Then, the pixel values ​​on the vertical line are normalized to the range of [0,1] using Min-Max Normalization, and finally a global boundary confidence map of the image is generated as shown in Figure 5 shown.

[0086] The boundary confidence map is used to train the boundary regression network, which can be improved from the convolutional neural network of any encoder-decoder structure. Figure 6 As shown, in a preferred embodiment, a boundary regression network is designed based on HRN et-V2, and the generation / decoding of multi-resolution blocks in HRNet-V2 is improved to adapt to the boundary generation task, including an Encoder-Decoder structure and an auxiliary branch; in the Encoder-Decoder structure, a channel attention module is used to coordinate the high / low resolution features of the encoder part during feature generation and fusion, and the weights of features of different resolutions can be calculated by any channel attention method; in the same way, the spatial attention module is used to adaptively fuse the high / low resolution features of the decoder part, so that its output is smoother than the output of cascading features of different resolutions; considering that low-resolution features have reliable semantic information, a gating signal is designed in the low-resolution features to realize the regression of high-resolution boundary confidence maps;

[0087] Compared with the segmentation task, the global regression task of the entire image means that the network must solve a wider solution space, which increases the difficulty of convergence. Therefore, auxiliary learning is introduced to help the network find a more robust feature representation and facilitate the target task. In a preferred embodiment, the feature X∈R 56×56×32 , which consists of four different resolution features and is enlarged to R by bilinear upsampling 112×112×32 Then, it is calculated by 1×1 convolution (filter = 2), and the above result is calculated by batch normalization and Softmax to obtain the binary boundary segmentation result, which is called the auxiliary branch task. At the same time, it is calculated by 1×1 convolution (filter = 1), and the above result is calculated by batch normalization and ReLU to obtain the boundary regression result, which is called the target branch task.

[0088] For the edge confidence map predicted by the edge regression network, after threshold screening and skeleton extraction, a preliminary flotation foam binary edge map can be obtained, but abnormal breakpoints will inevitably appear in the map, affecting the division and extraction of significant bubbles. In a preferred embodiment, the eight-neighborhood method and point-bubble matching are used to achieve adaptive breakpoint continuation. The process and schematic diagram are as follows: Figure 8 After connecting the breakpoints, we can get the complete edge information, and then get accurate and complete significant bubbles.

[0089] Relying on the objective laws of coal slime flotation theory, aiming at the characteristics of dense / significant areas, the characteristics of foam are summarized and described at different scales. In a preferred embodiment, the relevant feature extraction operators are optimized respectively (including but not limited to calculating the eigenvalues ​​after removing the shadow and bright spot influence for each bubble based on the segmentation results); the statistical values ​​(several scalars) including but not limited to the foam shape, surface texture and load capacity and the cumulative distribution of its size, speed and load capacity (several groups of one-dimensional vectors) are calculated, and the density map and the edge regression map (two-dimensional matrix) are combined to form the static characteristics of the coal slime flotation foam image in complex scenes described in three scales of "statistical value + cumulative distribution + regression map" ( Fig. 9 ).

[0090] The method according to the present invention described above can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-transitory machine-readable medium downloaded through a network and will be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor or programmable or dedicated hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the static feature extraction method of the froth image of coal slime flotation described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the processing shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the processing shown herein.

[0091] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A static feature extraction method for coal slime flotation foam image, characterized in that: The following steps are involved: S1. Obtain image data: Obtain a video of coal slime flotation foam under normal production conditions in a coal preparation plant; Take several frames of images from the video for manual point annotation, abstract the continuous foam group into a discrete point set, and generate a two-dimensional matrix of the same size as the image resolution; S2. Generate a foam density map: generate a sparse density matrix according to the relative positions of bubbles in the map; The sparse matrix is ​​convoluted and smoothed using a Gaussian kernel to obtain a foam density map; S3. Deep learning network training: Standardize the foam density map, including density value standardization and map size standardization; design a density regression network with an encoder-decoder structure, and introduce a loss selective discarding mechanism; use the processed foam density map as a supervised learning target, train the density regression network, and realize the prediction of the density of coal slime flotation foam and regional division; S4. Segmentation of significant bubbles: According to the prediction results of the density of coal slime flotation foam, the significant bubbles in the non-dense area are manually annotated at the pixel level, and the binary edge annotation results of the significant bubbles are converted into edge confidence maps through the expansion of boundary pixels and the calculation of probability values; Design an edge regression network with an encoder-decoder structure and auxiliary decoding branch capable of multi-resolution feature extraction, and train it with edge confidence maps as supervision targets; According to the predicted edge probability value, a binary edge map is generated through a threshold, and the breakpoints are connected using the skeleton extraction, eight-neighborhood method breakpoint search and point-bubble matching methods to obtain complete significant bubble edge information; S5. Feature extraction and analysis: Based on the output of the density regression network, the features of the dense foam area are extracted to analyze the density and distribution characteristics of the foam; Based on the output of the edge regression network, the spatial distribution, quantity, and size characteristics of significant bubbles are extracted; The density map and the marginal regression map are combined to form a static feature description of the foam image, including the statistical values ​​of foam shape, surface texture, loading rate, and the cumulative distribution data of foam size and distribution.

2. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: Density regression network, including: Encoder: Adopts HRNet module to maintain multi-resolution feature maps and fuse semantic information of different resolutions; Decoder: Use the deconvolution module to enlarge the resolution of the network output and extract the details of the bubble boundaries; Loss selective discarding mechanism: remove abnormal loss values ​​and optimize network training.

3. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: The specific steps of training the density regression network are as follows: the foam image and the generated foam density map are input into the density regression network for training; the density regression network outputs a predicted density map, and by summing all density values, the number of bubbles in the image, the area ratio between large and small bubbles, and their spatial distribution are obtained.

4. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: In the significant bubble segmentation, it is necessary to detect significant bubbles. The target detector is used to detect significant bubbles in the non-dense area of ​​the image. The bounding box of each significant bubble is marked, and the image area of ​​the significant bubble is cropped along the bounding box. Each salient bubble is annotated with binary edges at the semantic segmentation level; the edge confidence map is calculated using Gaussian distribution and normalized to the range [0,1].

5. The static feature extraction method of coal slime flotation foam image according to claim 4, characterized in that: Calculating the edge confidence map specifically includes: Choose mean μ and standard deviation σ to define the Gaussian kernel; Calculate the Gaussian distribution for each pixel point x in the vertical direction of the edge center i : Among them, E(x i ) represents pixel x i The edge confidence of x i The value of is the distance of the pixel from the center pixel on the edge; Find the edge confidence map E(x i ) in the maximum value E max ; Normalize the values ​​in the edge confidence map to the range [0,1]: The normalized edge confidence map E norm (x) Save as image.

6. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: The edge regression network includes: based on the HRNet-V2 module, using the channel attention module and the spatial attention module to extract and fuse features; using the boundary confidence map to train the network and output the predicted edge confidence map.

7. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: The value of the marked point in the two-dimensional matrix is ​​1, and the value of the other positions is 0; if at pixel x i If there are bubbles at , the matrix annotated with N bubbles is expressed as: Where N is the total number of bubbles detected in the image, x is a certain position in the image, and x i is the center position of the ith bubble, δ(xx i ) is the indicator function, when x=x i The value is 1 when the condition is true, otherwise it is 0.

8. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: The formula for generating the foam density map is: D′(x)=D(x)*G(x) Among them, G(x) is the Gaussian kernel function and * represents the convolution operation.

9. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: The deconvolution module consists of a 4×4 deconvolution layer, a batch normalization layer, a ReLU activation function, and a residual block.

10. The static feature extraction method of coal slime flotation foam image according to claim 1, characterized in that: The steps of the breakpoint continuation algorithm include: detecting breakpoints in the edge graph, finding the eight neighborhoods of the breakpoints, matching the breakpoints with the connected areas to which they belong, connecting the breakpoints, and obtaining complete edge information.

Images