Improved coal slime flotation froth image segmentation method based on mask coverage

Abstract

The invention relates to an improved coal slime flotation froth image segmentation method based on mask coverage, belongs to the technical field of coal slime flotation, and solves the problem of low segmentation precision of a coal slime flotation froth image in the prior art. The method comprises the following specific steps: marking flotation bubble boundaries in a collected historical coal slime flotation bubble image instance to obtain an image training set; based on the image training set and the loss function, training a flotation froth image mask instance segmentation model; inputting a flotation froth image instance to be segmented into the trained flotation froth image mask instance segmentation model to obtain a predicted segmentation result; based on the segmentation result, marking a position corresponding to the input flotation froth image to obtain a segmentation mask image of the coal slime flotation froth; and performing boundary fitting on the flotation froth segmentation mask image by using a least square method to obtain the actual boundary of the flotation froth, thereby improving the segmentation precision of the coal slime flotation froth image.

Description

Technical Field

[0001] The invention relates to the technical field of coal slime flotation, and in particular to an improved coal slime flotation foam image segmentation method based on mask covering. Background Art

[0002] Flotation is the main method for sorting fine coal slime and is widely used in the sorting of fine coal slime in coal preparation plants. Flotation foam is the main carrier of clean coal aggregation in the flotation chamber, and can reflect a lot of control information in the flotation condition of coal slime, including clean coal ash content, clean coal recovery rate, reagent addition dosage, aeration volume and flotation machine working status. In the existing technical methods, these feature information are extracted from flotation foam images based on computer vision technology. Among them, morphological features are one of the important visual features of flotation foam images and are widely used in flotation modeling and prediction research. The necessary technical process for extracting the morphological features of flotation foam images is image segmentation. At present, there are some segmentation methods applied to the segmentation of metal ore flotation foam images. In the application of coal slime flotation foam images, due to the blurred bubble boundaries and high viscosity, the current technologies and methods have large errors in dealing with this problem.

[0003] In existing technologies or methods, the segmentation of flotation foam images relies on an image threshold parameter, which depends on the visual attributes of the flotation foam (foam texture, color and bubble size distribution) and lighting conditions. The image threshold varies with different ore types or flotation conditions, thus affecting the accuracy of traditional segmentation methods. In addition, existing froth image segmentation techniques are greatly affected by image quality, lighting environment and the merging of smaller bubbles by larger bubbles. Adaptive labeled watershed transform is another method used to measure the bubble size distribution in froth images, in which adaptive threshold processing is combined with mean-related image shift to form relatively flat areas near the bubble top area and bubble boundaries. Applying an adaptive median filter to the generated image can smooth all small intensity changes in each flat area to prevent over-segmentation. However, there are obvious visual differences in flotation foams of different mineral types. Similar methods have large segmentation errors in coal slime flotation foam images, and there are practical problems such as over-segmentation or under-segmentation. Summary of the invention

[0004] In view of the above analysis, the embodiment of the present invention aims to provide an improved coal slime flotation foam image segmentation method based on mask covering, so as to solve the problem of low accuracy of coal slime flotation foam image segmentation in the prior art.

[0005] The purpose of the present invention is mainly achieved through the following technical solutions:

[0006] The embodiment of the present invention provides an improved coal slime flotation foam image segmentation method based on mask coverage, comprising the following steps:

[0007] Annotate the flotation bubble boundaries in the collected historical coal slime flotation foam image instances to obtain an image training set;

[0008] Based on the image training set and the loss function, training a flotation foam image mask instance segmentation model;

[0009] Inputting the flotation foam image instance to be segmented into the trained flotation foam image mask instance segmentation model to obtain a predicted segmentation result;

[0010] Based on the segmentation result, the position corresponding to the input flotation foam image is marked to obtain a segmentation mask image of the coal slime flotation foam; the flotation foam segmentation mask image is subjected to boundary fitting using the least squares method to obtain the actual boundary of the flotation bubble.

[0011] Furthermore, the least square method performs boundary fitting based on the following formula:

[0012] F(a,b,c)=CΣ(x i 2 +y i 2 +ax i +by i +c)C T ,

[0013] Where F represents the geometric distance of any point (x, y) in the bubble box in the segmentation mask image; C is the regularization matrix; a, b, c are the constraints that satisfy 4ac-b 2 =1 linear parameter.

[0014] Furthermore, the flotation foam image mask instance segmentation model includes:

[0015] The backbone network is used to extract the basic features of the flotation bubbles in the image and obtain feature maps of different sizes;

[0016] A pyramid feature extraction layer, used for performing feature fusion on the feature maps of different sizes to obtain fused feature maps of the same size;

[0017] A receptive field Align module, used to generate multiple query frame diagrams on the fused feature map, and extract flotation bubble features in the query frame diagram from the feature map using the receptive field Align;

[0018] The prediction module includes a category prediction module and a mask prediction module; the category prediction module is used to predict the position of the bubble by classifying the bubble and the background based on the flotation bubble characteristics; the mask prediction module is used to predict the bubble mask according to the bubble position.

[0019] Furthermore, the calculation formula of the category prediction module is as follows:

[0020]

[0021] Among them, x is the vector representation of the input feature; i is the encoding representation of the classification category; Softmax represents the probability of floating bubbles in the query box.

[0022] Furthermore, the loss function is calculated based on the following formula:

[0023]

[0024] Among them, -a(1-p)m*logn is the category prediction loss, a is the adjustment factor, m is the weight distribution parameter, p is the probability of flotation bubble; n is a fixed value; is the mask prediction loss; X is the pixel category of the true segmentation image; Y is the pixel category of the predicted segmentation mask image.

[0025] Furthermore, the receptive field Align is calculated based on the following formula:

[0026]

[0027] Among them, Ls is the loss function; p i Represents the pixel point on the feature map; p rj represents the jth pixel of the rth receptive field of the feature map; i*(r,j) represents the maximum pixel of the query frame diagram.

[0028] Furthermore, a self-attention mechanism is introduced into the pyramid feature extraction layer. After the input feature maps of different sizes are transformed using the self-attention mechanism, they are fused through an upsampling operation to obtain the fused feature maps of the same size.

[0029] Furthermore, based on the convolutional network as the backbone network, basic features of flotation bubbles of different sizes in the flotation foam image are extracted, including: using 5 layers of 3*3 convolutional layers to obtain 5 feature maps of different sizes.

[0030] Furthermore, before labeling the flotation foam image instance, double wave filtering is performed based on the following formula:

[0031]

[0032] and

[0033] in, is the filtering weight; f(i,j) represents the pixel point output after filtering; K and L are the horizontal and vertical coordinates of the (2n+1) pixel distance centered at (i,j); g(i,j) represents the input pixel point before filtering; g avg is the average grayscale of the input pixel g(i,j), g set is the set grayscale value, d is the filter radius, and d1 <d2。

[0034] Furthermore, the segmentation performance of the flotation foam image mask instance segmentation model was evaluated during training using the intersection-over-union (IOU) and F1-score based on the following formula:

[0035]

[0036] Among them, precision represents the accuracy of the corresponding pixel mask prediction; recall represents the recall rate of the corresponding pixel mask prediction; M is the pixel of the segmented area predicted by the flotation foam image mask instance segmentation model, and G is the pixel of the real instance segmentation area.

[0037] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0038] 1. The present invention proposes to accurately capture the real boundary of the coal slime flotation foam image by manually annotating the image, and obtain the mask image of the coal slime flotation foam segmentation by combining the mask instance segmentation model. The least squares fitting method can be used to obtain a more realistic bubble morphology boundary, effectively improving the segmentation accuracy of the coal slime flotation foam image;

[0039] 2. By improving the least square method to fit the actual and merged boundaries of bubbles, the morphological characteristics of flotation foam images are obtained, and the segmentation accuracy of flotation foam images is improved;

[0040] 3. The introduction of the self-attention mechanism can enhance the recognition probability of a single flotation bubble area and improve the segmentation accuracy;

[0041] 4. By marking the real boundaries of bubbles in the image based on the improved bilateral filtering, a more accurate data set can be obtained, which helps to improve the accuracy of model training.

[0042] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0044] Figure 1 It is a flow chart of an improved method for instance segmentation of coal slime flotation foam image based on mask covering according to an embodiment of the present invention;

[0045] Figure 2 It is a flowchart of the implementation of the improved mask-covering-based coal slime flotation foam image instance segmentation according to an embodiment of the present invention;

[0046] Figure 3 This is an effect diagram of bilateral filtering of froth image of coal slime flotation according to an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of an improved mask coverage-based instance segmentation model structure according to an embodiment of the present invention;

[0048] Figure 5 It is a schematic diagram of the segmentation effect of the coal slime flotation foam image mask example according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0050] A specific embodiment of the present invention discloses an improved method for segmenting coal slime flotation foam images based on mask coverage, such as Figure 1 As shown, the following steps are included:

[0051] Step S1, marking the flotation bubble boundaries in the collected historical coal slime flotation foam image instances to obtain an image training set;

[0052] Step S2, training a flotation foam image mask instance segmentation model based on the image training set and the loss function;

[0053] Step S3, inputting the flotation foam image instance to be segmented into the trained flotation foam image mask instance segmentation model to obtain a predicted segmentation result;

[0054] Step S4: based on the segmentation result, mark the position corresponding to the input flotation foam image to obtain a segmentation mask image of the coal slime flotation foam; perform boundary fitting on the flotation foam segmentation mask image using the least squares method to obtain the actual boundary of the flotation bubble.

[0055] Through the above method, the flotation foam image mask instance segmentation model is trained based on the flotation foam image with manually annotated bubbles, the trained model is used to perform mask segmentation on the flotation foam image, and the predicted mask segmentation map is fitted with the least squares method to obtain a more realistic bubble morphology boundary, thereby improving the segmentation accuracy of the coal slime flotation foam image.

[0056] Specifically, in step S1, a coal slime flotation foam image visual system is designed and built, and a coal slime flotation foam image dataset is collected on an industrial flotation machine in a coal preparation plant.

[0057] Exemplarily, the visual system includes: an industrial camera, an LED diffuse reflection light source, a sunshade, and a windows server. The visual system is arranged in a stable area just above the foam layer of the industrial flotation machine, the collection time is set, and the collected samples try to fully cover the entire working condition of the coal slime flotation process system, and the collected flotation foam images are saved locally. According to the above operation, it took one week on site to collect a total of 545 foam images.

[0058] Furthermore, the noise existing in the original foam image will affect the subsequent segmentation effect. According to the bright spot and shadow properties between the top and the boundary of the foam image, an improved bilateral filtering method is designed. The grayscale difference between the top and the boundary of the foam image is used as the basis. The filtering radius is determined by the grayscale mean of the image, and the filtering weight is set to be only related to the difference in pixel values ​​to improve the calculation efficiency. The filtering formula for the flotation foam image is as follows:

[0059]

[0060] and

[0061] in, is the filter weight; f(i,j) represents the pixel point output after filtering; K and L are the horizontal and vertical coordinates of the (2n+1) pixel distance centered at (i,j); g(i,j) represents the input pixel point of the filter; g avg is the average grayscale of the input pixel g(i,j), g set is the set grayscale value, d is the filter radius, and d1 <d2。

[0062] For example, the size of the input image is scaled to 224*224 pixels, and the diameters d1 and d2 of the filter are set to 5 and 9 pixels respectively. If the average grayscale of the input foam image is calculated to be 43, which is less than the grayscale cutoff value of 68, a filter diameter of 5 pixels is selected to preserve the features of the light and dark areas on the foam image. The foam image before and after filtering is shown in Figure 2. Figure 3 shown.

[0063] The filtered images are saved in a fixed dataset, and each bubble in the coal slime flotation foam image is annotated based on the Labelme tool. Then the annotated files are converted into an image recognition dataset (Common Objects in Context, COCO dataset) as the input data for training the flotation foam image mask instance segmentation model, i.e., the image training set.

[0064] Specifically, in step 2, a flotation foam image mask instance segmentation model is constructed based on the pyramid network. The network structure flow chart of the segmentation model is as follows: Figure 4 As shown, specifically including:

[0065] Step S21, based on the backbone network (Backbone), extracting the basic features of the flotation foam image to obtain multiple feature maps;

[0066] For example, based on a 6-layer convolutional network as the backbone, the basic features of the flotation foam image are extracted to obtain 5 feature maps; the model is written based on the PyTorch1.4 architecture, and the coal slime flotation foam image in the input data set is input into the backbone network. The pixel size of the input image is 224*224 pixels, as shown in Table 1, which is the data composition of the backbone structure:

[0067] Table 1

[0068] Number of layers Convolution kernel size Feature map size C1 7*7 64*64 C2 3*3 128*128 C3 3*3 128*128 C4 3*3 256*256 C5 3*3 256*256 C6 3*3 512*512

[0069] The foam image passes through the L2-L6 layers of the backbone network, and each layer outputs a feature map. The feature map of each layer can ensure the identification of all bubbles of different sizes on the foam image.

[0070] Step S22, the extracted feature map is passed into four pyramid feature extraction layers F1-F4 (including F1 texture feature, F2 size feature, F3 foam layer thickness feature and F4 dynamic feature), and after the four 3*3 convolution layer feature fusion, a new feature map of the same size is generated, i.e., a fused feature map;

[0071] Furthermore, a self-attention module is introduced in the pyramid feature extraction layer, such as passing the five feature maps output in step S21 through a multi-scale self-attention module. The module is different from the existing self-attention mechanism module in that the output feature maps are fused into five 512*512 fused feature maps after upsampling. The formula for transforming the input foam image feature matrix by the self-attention module is as follows:

[0072]

[0073] Among them, Q, K, and V are linear transformations of the input feature matrix X; d is the dimension of the feature matrix, which is used for normalization to prevent the value from being too large; and F is the output feature matrix of the self-attention module.

[0074] By introducing a dynamically improved self-attention module, the input foam image feature matrix is ​​transformed to enhance the feature representation capability. At the same time, upsampling is used to find the area where the bubbles are located so that the bubbles are scattered in each feature map (that is, the number of bubbles in each feature map is as small as possible).

[0075] Step S23, the fused feature map uses receptive field Align (ROIAlign) to capture and process the features of the flotation bubbles to prevent some bubbles from being lost in the feature map;

[0076] For example, in order to search the entire feature map, k query blocks Qb = {b i |i-1,2,...,k}, fill the entire feature map, and then use k receptive field Align on the feature map to search in turn from the feature map to determine the location of the query frame map. The frame map area with the smallest partial derivative of the loss function L is the query frame where the final bubble is located. The corresponding flotation bubble features in the query frame area are extracted to obtain the feature map for subsequent bubble position and mask prediction. The receptive field Align calculation formula is as follows:

[0077]

[0078] Among them, Ls represents the loss function; p i Represents the pixel point on the feature map, p rj It represents the jth pixel of the rth receptive field after upsampling and pooling of the feature map, and i*(r,j) represents the maximum pixel of the candidate area (i.e., the query frame). The k receptive field sizes are completed by recursive calculation

[0079] Step S24, connecting the feature map output by the receptive field Align to the prediction module of the coal slime flotation foam image, predicting the bubble position and mask of the features extracted by ROIAlign, and obtaining the predicted segmentation result.

[0080] Specifically, the prediction module consists of two parts, the category prediction module and the mask prediction module. The category prediction module predicts the position of the floating bubbles by classifying the bubbles and the image background into two categories, and obtains the segmented area image. The query border in the image can be selected as needed. The input features are represented in vector form, and the module calculation formula is as follows:

[0081]

[0082] Among them, x represents the input feature vector; i represents the encoding representation of the classification category, such as 0 represents background and 1 represents bubble; Softmax is the probability that the calculation area is a flotation bubble, and the pixel area with the maximum probability value in the calculated area, that is, the query frame diagram, is taken as a single bubble area.

[0083] Correspondingly, the mask prediction module is responsible for predicting the mask at the bubble prediction position to obtain the bubble mask image. The calculation formula of the mask prediction module is as follows:

[0084]

[0085] Among them, x j With y i They are the feature map of the j-th channel of the output feature tensor X and the feature map of the i-th channel of the output feature tensor Y, respectively.

[0086] During the training process, based on the loss function of the model, by observing the loss value of the model, the model hyperparameters are continuously optimized, including the number of iterations (Epoch), initial learning rate (Lr), and category (num class).

[0087] Specifically, according to the category output and mask output of the flotation foam image mask instance segmentation model, the loss function is composed of the sum of the category prediction loss and the mask prediction loss, and the calculation formula is as follows:

[0088]

[0089] Among them, the first half of the right side of the formula is the category prediction loss, a is the adjustment factor, m is the weight distribution parameter, p is the probability of floating bubbles, which is calculated by the Softmax of the category prediction module, and n is a fixed value; the second half of the right side of the formula is the mask prediction loss, X is the pixel category of the real segmented image, and Y is the pixel category of the predicted segmented image.

[0090] In addition, based on the training loss curve of the model, the F1-score and intersection-over-union (IOU) parameters are selected as evaluation indicators of the model segmentation performance to evaluate the model performance. The calculation formulas of F1-score and IOU are as follows:

[0091]

[0092] Among them, precision represents the accuracy of the corresponding pixel mask prediction, recall represents the recall rate of the corresponding pixel mask prediction, M represents the pixels of the segmented area predicted by the model, and G represents the pixels of the real instance segmented area. When the evaluation index error is large, the model hyperparameters are adjusted and the model is retrained until the segmentation accuracy meets the specific application requirements.

[0093] For example, 545 coal slime flotation foam images collected on an industrial flotation machine are randomly divided into training set, validation set and test set in a ratio of 7:2:1. The number of iterations (Epoch) is set to 500, the initial learning rate (Lr) is set to 0.01, and the class (num_class) is set to 2. The IOU of the flotation foam image dataset training is 0.75, and the F1-score (IOU=0.75) is 82.46 under the condition of IOU of 0.75, indicating that the model has achieved a relatively reliable segmentation accuracy.

[0094] Furthermore, a high-precision coal slime flotation foam image instance segmentation model was obtained on the training set and validation set, and the generalization performance of the model was verified on the test set.

[0095] Specifically, in step S4, the prediction results (bubble mask image and segmented area image) output by the model are used to mark the positions corresponding to the flotation foam original image, and the segmented mask image of the coal slime flotation foam image is obtained, that is, the coal slime flotation foam segmented image covered with the mask, such as Figure 5 At the same time, the bubble mask image and the segmented region image can be visualized and output.

[0096] It should be noted that the segmented coal slime flotation foam segmentation mask image is not the real morphological boundary of the flotation foam. In order to further obtain a more realistic coal slime flotation foam image boundary, such as Figure 2 As shown in the figure, the improved least square method is used to fit the actual and merged boundaries of the flotation bubbles. The difference is that the regularization matrix C is added to make the bubble fitting boundary smoother. The calculation formula of the fitting boundary is as follows:

[0097] F(a,b,c)=C∑(x i 2 +y i 2 +ax i +by i +c)C T ,

[0098] And 4ac-b 2 =1,

[0099] Where F is the geometric distance of any point (x, y) in the bubble box in the segmentation mask image, which can be calculated by Euclidean distance; a, b, c are the linear parameters of the equation, 4ac-b 2 =1 is the constraint condition for the ellipse shape.

[0100] Furthermore, the morphological boundaries of the coal slime flotation bubbles are depicted according to the fitted shapes, and the morphological characteristics of the flotation foam image are calculated.

[0101] Exemplarily, the least squares fitting is implemented based on optimize.leastsq, and Dfun=None, full_output=0, a=0, b=1.49012e-08, c=1.49012e-08, d=0.0, f=0, epsfcn=None, factor=96, diag=None is set. The outer ellipse of the fitted mask is the real bubble morphology. The pixel area of ​​the bubble mask area, the major and minor axes of the fitted ellipse, the perimeter and the area are calculated as the morphological characteristics of the coal slime flotation bubble.

[0102] Compared with the prior art, the present embodiment provides an improved coal slime flotation foam image segmentation method based on mask coverage, which inputs the filtered and annotated flotation foam image into a mask instance segmentation model, performs real boundary fitting on the obtained segmented mask image based on the least squares method, and finally realizes accurate segmentation of individual flotation bubbles on the coal slime flotation foam image; at the same time, a self-attention mechanism is introduced into the mask instance segmentation model based on the pyramid network structure to improve the segmentation accuracy.

[0103] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0104] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An improved mask-covering-based coal slime flotation foam image instance segmentation method, characterized in that: The steps include: Annotate the flotation bubble boundaries in the collected historical coal slime flotation foam image instances to obtain an image training set; Based on the image training set and the loss function, training a flotation foam image mask instance segmentation model; Inputting the flotation foam image instance to be segmented into the trained flotation foam image mask instance segmentation model to obtain a predicted segmentation result; Based on the segmentation result, the position corresponding to the input flotation foam image is marked to obtain a segmentation mask image of the coal slime flotation foam; the flotation foam segmentation mask image is subjected to boundary fitting using the least squares method to obtain the actual boundary of the flotation bubble.

2. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 1, characterized in that: The least squares method performs boundary fitting based on the following formula: F(a,b,c)=C∑(x i 2 +y i 2 +ax i +by i +c)C T , Where F represents the geometric distance of any point (x, y) in the bubble box in the segmentation mask image; C is the regularization matrix; a, b, c are the constraints that satisfy 4ac-b 2 =1 linear parameter.

3. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 1, characterized in that: The flotation foam image mask instance segmentation model comprises: The backbone network is used to extract the basic features of the flotation bubbles in the image and obtain feature maps of different sizes; A pyramid feature extraction layer, used for performing feature fusion on the feature maps of different sizes to obtain fused feature maps of the same size; A receptive field Align module, used to generate multiple query frame diagrams on the fused feature map, and extract flotation bubble features in the query frame diagram from the feature map using the receptive field Align; The prediction module includes a category prediction module and a mask prediction module; the category prediction module is used to predict the position of the bubble by classifying the bubble and the background based on the flotation bubble characteristics; the mask prediction module is used to predict the bubble mask according to the bubble position.

4. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 3, characterized in that: The calculation formula of the category prediction module is as follows: Among them, x is the vector representation of the input feature; i is the encoding representation of the classification category; Softmax represents the probability of floating bubbles in the query box.

5. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 4, characterized in that: The loss function is calculated based on the following formula: Among them, -a(1-p) m *logn is the category prediction loss, a is the adjustment factor, m is the weight distribution parameter, p is the probability of flotation bubble; n is a fixed value; is the mask prediction loss; X is the pixel category of the true segmentation image; Y is the pixel category of the predicted segmentation mask image.

6. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 5, characterized in that: The receptive field Align is calculated based on the following formula: Among them, Ls is the loss function; p i Represents the pixel point on the feature map; p rj represents the jth pixel of the rth receptive field of the feature map; i*(r,j) represents the maximum pixel of the query frame diagram.

7. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 3, characterized in that: A self-attention mechanism is introduced into the pyramid feature extraction layer. After the input feature maps of different sizes are transformed using the self-attention mechanism, they are fused through an upsampling operation to obtain the fused feature maps of the same size.

8. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 3, characterized in that: Based on the convolutional network as the backbone network, the basic features of flotation bubbles of different sizes in the flotation foam image are extracted, including: using 5 layers of 3*3 convolutional layers to obtain 5 feature maps of different sizes.

9. An improved mask-covering-based coal slime flotation foam image instance segmentation method according to any one of claims 1 to 8, characterized in that: Before annotating the flotation foam image example, double wave filtering is performed based on the following formula: and in, is the filtering weight; f(i,j) represents the pixel point output after filtering; K and L are the horizontal and vertical coordinates of the (2n+1) pixel distance centered at (i,j), respectively; g(i,j) represents the input pixel before filtering; g avg is the average grayscale of the input pixel g(i,j), g set is the set grayscale value, d is the filter radius, and d1 <d2。 10. The improved mask-covering-based coal slime flotation foam image instance segmentation method according to claim 1, characterized in that: During training, the segmentation performance of the flotation foam image mask instance segmentation model was evaluated using the intersection-over-union (IOU) and F1-score based on the following formula: Among them, precision represents the accuracy of the corresponding pixel mask prediction; recall represents the recall rate of the corresponding pixel mask prediction; M is the pixel of the segmented area predicted by the flotation foam image mask instance segmentation model, and G is the pixel of the real instance segmentation area.

Images