Flotation bubble size measurement method integrating control mark watershed and U-Net
By combining the method of controlling the marking watershed and U-Net, combining Euclidean distance transformation and OTSU automatic threshold segmentation, the problem of undersegment and non-closed boundary in flotation bubble size measurement is solved, and high-precision bubble size measurement is achieved.
Patent Information
- Application Number
- CN202510398553.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-01
- Publication Date
- 2025-07-04
AI Technical Summary
Among the existing flotation bubble size measurement methods, there are problems of undersegment or oversegment based on the watershed algorithm, while the U-Net method is prone to generate non-closed boundaries of bubbles, resulting in inaccurate measurements.
The method of fusion control marking watershed and U-Net is used to obtain the bubble connection area through preliminary segmentation of U-Net convolutional neural network, combined with Euclidean distance transformation and OTSU automatic threshold segmentation, extract the first and second types of control marks, and then perform fusion and use the watershed algorithm to perform the final segmentation.
It significantly improves the accuracy and reliability of flotation bubble size measurement, solves the shortcomings in the traditional methods, and realizes the closed and accurate measurement of bubble boundaries.
Smart Images
Figure CN120259403A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing, and in particular to a method for measuring the size of flotation bubbles by fusing control marker watershed and U-Net. Background Art
[0002] At present, more than 90% of non-ferrous metal ores are processed by foam flotation. Flotation is a physico-chemical process that utilizes the differences in the surface properties of minerals to separate valuable minerals from valueless minerals by means of bubble attachment. During the flotation process, bubbles are the carriers of mineral particles, and the mineral particles will adhere to the surface of the bubbles. The bubble size is a typical visual feature of the surface foam in flotation and plays a crucial role in the flotation process, directly affecting the collection and separation efficiency of mineral particles. In order to improve the flotation efficiency, the bubble size must be appropriate. If the bubble radius is too large, the number of mineral particles that the same volume of air can carry will decrease, thereby reducing the flotation efficiency; if the bubble radius is too small, the surface of the bubble is likely to carry too many mineral particles, resulting in an excessive average density of the mineralized bubbles and inability to float, which will also affect the flotation recovery rate and even make it impossible to achieve flotation in severe cases. Therefore, accurately obtaining the size of flotation bubbles is the key to ensuring that the bubble size is within the optimal range and achieving an increase in flotation efficiency and mineral recovery rate.
[0003] At present, the measurement of flotation bubble size is usually carried out by means of image segmentation. The bubble boundary is extracted through computer vision technology to obtain the size of each bubble. The main methods include the image segmentation method based on the watershed algorithm and the image segmentation method based on U-Net. However, both of these methods have corresponding problems, which are specifically as follows:
[0004] 1. To reduce the over-segmentation problem caused by noise, the image segmentation method based on the watershed usually adopts the watershed algorithm based on control markers. However, the working conditions of foam flotation are variable, the size distribution of flotation bubbles is uneven, the image gray-scale distribution is complex, the bubble shapes are randomly shaped, adhered to each other, numerous and the boundaries are blurred, and there is no obvious background. In this case, the extraction of control markers mainly relies on empirical selection according to the characteristics of the foam, and it is difficult to accurately extract the control markers corresponding to each flotation bubble, inevitably resulting in under-segmentation or over-segmentation, and ultimately affecting the accurate measurement of the flotation bubble size.
[0005] 2. The image segmentation method based on U-Net constructs a dataset of flotation bubble boundaries, adopts the U-Net architecture, and constructs a relationship model from the foam image to the bubble boundary from end to end to achieve the segmentation of the boundary region and non-boundary region in the flotation image. However, this method often produces the problem of non-closed boundaries of bubbles in practical applications, resulting in multiple bubbles being mis-merged and recognized as one bubble, thus affecting the accurate measurement of the flotation bubble size. Summary of the Invention
[0006] In view of the deficiencies existing in the prior art, the present invention provides the following technical solutions:
[0007] A method for measuring the size of flotation bubbles by fusing control markers with watershed and U-Net includes the following steps:
[0008] S10: Use the U-Net convolutional neural network to preliminarily segment the foam image, obtain the initial bubble segmentation map, and extract the connected regions of each flotation bubble according to the initial bubble segmentation map;
[0009] S20: According to the preliminary segmentation result, obtain the reconstructed landform map, and extract the local area minima from the reconstructed landform map as the first type of control markers;
[0010] S30: Denoise the foam image, and use the OTSU automatic threshold segmentation method based on the connected region mask to obtain the highlighted region of the flotation bubbles as the second type of control markers;
[0011] S40: Analyze the connected regions of the bubbles, fuse the first type of control markers and the second type of control markers to obtain the fused control markers;
[0012] S50: According to the fused control markers extracted in S40 and the reconstructed landform map extracted in S20, use the control marker watershed algorithm for segmentation to obtain the final boundary of the flotation bubbles, and realize the measurement of the size of the flotation bubbles.
[0013] As an improvement of the above technical solution, the steps of obtaining the initial bubble segmentation map and the connected regions of each flotation bubble in step S10 include the following steps:
[0014] S11: Construct a dataset of flotation bubble boundaries, and establish a flotation bubble boundary segmentation model based on U-Net. This model takes the foam image as the input and the flotation bubble boundary marker map as the output.
[0015] S12: Use the U-Net convolutional neural network to preliminarily segment the input foam image to obtain the initial bubble segmentation map.
[0016] S13: According to the initial bubble segmentation map, use the image connected region tracking algorithm to extract the connected regions of each flotation bubble.
[0017] As an improvement of the above technical solution, the extraction of the first type of control markers in step S20 includes the following steps:
[0018] S21: Perform Euclidean distance transformation on the initial bubble segmentation map, so that the pixel points closer to the preliminary predicted boundary points have lower numerical values, and obtain the distance-transformed image.
[0019] S22: Process the image after distance transformation using an exponential decay function based on the Euler number to achieve non-linear distance mapping and normalization, and make the pixel values higher for those closer to the preliminary predicted boundary points of the distance.
[0020] S23: Expand the gray value range of the image processed in S22, and perform smoothing processing on the image using Gaussian filtering to obtain the final reconstructed landform map.
[0021] As an improvement of the above technical solution, the Euclidean transformation depends on the following formula:
[0022] D(p) = min q∈F ||p - q||2
[0023] where D(p) is the value of pixel p in the distance transformation result image, F is the set of all boundary pixels, that is, non-zero values, and ||p - q||2 represents the Euclidean distance between pixels p and q.
[0024] As an improvement of the above technical solution, the exponential decay function based on the Euler number is:
[0025]
[0026] where D max is the maximum value in D(x, y), which is the value of pixel p in the distance transformation result image.
[0027] As an improvement of the above technical solution, the method for extracting local region minima from the reconstructed landform map in step S20 depends on the following steps:
[0028] S24: Adopt the method of extracting eight-neighborhood pixels to obtain the local region minima in the reconstructed landform map and use them as the first type of control markers.
[0029] As an improvement of the above technical solution, the denoising process of the input image in step S30 includes the following steps:
[0030] S31: At least use the Gaussian filtering method and the mean filtering method to remove noise from the image.
[0031] S32: Enhance the local contrast of the image by comparing the limited adaptive histogram equalization.
[0032] As an improvement of the above technical solution, the method of using the OTSU automatic threshold segmentation method based on the connected region mask to obtain the highlighted region of the flotation bubbles as the second type of control marker in step S30 includes the following steps:
[0033] S33: In each flotation bubble connection region obtained in S10, the OTSU method is respectively used for automatic threshold segmentation to obtain the highlighted region of each flotation bubble.
[0034] S34: Merge the highlighted regions of each flotation bubble connection region into one graph as the second type of control marker.
[0035] As an improvement of the above technical solution, the acquisition of the fusion control marker in step S40 depends on the following steps for analysis:
[0036] S41: If there is only one first type of control marker in the flotation bubble connection region, select the first type of control marker as the fusion control marker;
[0037] S42: If there are two or more first type of control markers in the flotation bubble connection region, extract the corresponding second type of control marker of this connection region as the fusion control marker.
[0038] Advantages of the present invention:
[0039] The present invention innovatively combines the Unet convolutional neural network and the watershed algorithm to construct an efficient bubble boundary detection method. It fully integrates the advantages of the watershed algorithm in forming a closed bubble boundary and the ability of the U-Net model to efficiently extract boundary features from data samples, effectively solving the limitations of the traditional watershed algorithm's over-reliance on domain knowledge and the technical problem of unclosed bubble boundaries in the existing U-Net methods, thus significantly improving the accuracy and reliability of bubble size measurement. Description of the drawings
[0040] Figure 1 Is the basic framework for the implementation of the present invention;
[0041] Figure 2 Is the preliminary result of the flotation bubble boundary segmented by the U-Net convolutional neural network of the present invention;
[0042] Figure 3 Is the final reconstructed landform map obtained by the present invention;
[0043] Figure 4 Is the schematic diagram after local region minimum value marking of the reconstructed landform map of the present invention;
[0044] Figure 5 Is the schematic diagram of the highlighted region in the flotation bubble image of the present invention;
[0045] Figure 6 Is the preliminary boundary prediction result obtained by the U-Net of the present invention and each bubble connection region;
[0046] Figure 7The first type of control marker result diagram extracted for the present invention;
[0047] Figure 8 The second type of control marker result diagram extracted for the present invention;
[0048] Figure 9 The comparative analysis diagram of the first type of control marker and the second type of control marker for the present invention;
[0049] Figure 10 The final result diagram after watershed processing for the present invention. Detailed implementation manners
[0050] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.
[0051] Currently, the measurement of flotation bubble size usually adopts the method of image segmentation. The bubble boundary is extracted through computer vision technology to obtain the size of each bubble. The main methods include the image segmentation method based on the watershed algorithm and the image segmentation method based on U-Net. However, both of these methods have corresponding problems, which are specifically as follows:
[0052] 1. To reduce the over-segmentation problem caused by noise, the image segmentation method based on the watershed usually adopts the watershed algorithm based on control markers. However, in the flotation bubble condition, the working conditions are variable, the size distribution of flotation bubbles is uneven, the image gray distribution is complex, the bubble shape is random, sticky, numerous and the boundary is blurred, and there is no obvious background. In this case, the extraction of control markers mainly relies on empirical selection according to the characteristics of the foam, and it is difficult to accurately extract the control markers corresponding to each flotation bubble, inevitably resulting in under-segmentation or over-segmentation, and ultimately affecting the accurate measurement of the flotation bubble size.
[0053] 2. The image segmentation method based on U-Net constructs a flotation bubble boundary data set, adopts the U-Net architecture, and constructs a relationship model from the foam image to the bubble boundary from end to end to realize the segmentation of the boundary area and non-boundary area in the flotation image. However, this method often produces the problem of non-closed boundaries of bubbles in actual applications, resulting in multiple bubbles being mis-merged and recognized as one bubble, thus affecting the accurate measurement of the flotation bubble size.
[0054] To solve the above problems, a method for measuring the flotation bubble size by fusing the control marker watershed and U-Net is provided. The implementation steps are as Figure 1 shown, including the following steps:
[0055] S10: Use the U-Net convolutional neural network to perform preliminary segmentation on the foam image, obtain the initial bubble segmentation map, and extract the connected regions of each flotation bubble according to the initial bubble segmentation map.
[0056] Specifically, the obtaining of the initial bubble segmentation map and the connected regions of each flotation bubble in step S10 includes the following steps:
[0057] S11: Construct a flotation bubble boundary dataset, establish a U-Net-based flotation bubble boundary segmentation model, which takes the foam image as the input and the flotation bubble boundary labeling map as the output.
[0058] Train the U-Net convolutional neural network on the flotation bubble boundary dataset, and then perform step S12 for transformation.
[0059] S12: Perform preliminary segmentation on the input foam image through the U-Net convolutional neural network to obtain the initial bubble segmentation map, as Figure 2 shown.
[0060] S13: According to the initial bubble segmentation map, use the image connected region tracking algorithm to extract the connected regions of each flotation bubble, as Figure 3 shown.
[0061] S20: Obtain the reconstructed geomorphic map according to the preliminary segmentation result, and extract the local regional minimum value from the reconstructed geomorphic map as the first type of control marker.
[0062] Specifically, step S20 includes the following steps:
[0063] S21: Perform Euclidean distance transformation on the initial bubble segmentation map, so that the pixel points closer to the preliminary predicted boundary points have lower numerical values, and obtain the distance-transformed image.
[0064] The calculation method for the Euclidean transformation depends on the following formula:
[0065] D(p) = min q∈F ||p - q||2
[0066] where D(p) is the value of pixel p in the distance-transformed result image, F is the set of all boundary pixels, that is, non-zero values, and ||p - q||2 represents the Euclidean distance between pixels p and q.
[0067] where ||p - q||2 means
[0068] To create a non - linear distance mapping \(W(x,y)\), an exponential decay function is adopted in this embodiment. In other words, this function assigns a relatively high value to the pixels near the boundary and decreases as the distance from the boundary increases. Specifically, as shown in step S22:
[0069] S22: Process the image after distance transformation using an exponential decay function based on the Euler number to achieve non - linear distance mapping and normalization, and make the pixel values higher for the pixels closer to the preliminary predicted boundary points.
[0070] Among them, the acquisition of the non - linear distance mapping and normalization depends on the following formula:
[0071]
[0072] Among them, \(D\) max is the maximum value of the value \(D(x,y)\) of the pixel \(p\) in the distance transformation result image, which can be expressed as:
[0073] \(D\) max \(=\max(D(x,y))\)
[0074] After completing step S22, the image also needs to be smoothed, specifically as shown in step S23.
[0075] S23: Expand the gray - value range of the image processed in S22, and use Gaussian filtering to smooth the image to obtain the final reconstructed landform map.
[0076] Among them, the final reconstructed landform map is as Figure 4 shown. After completing the reconstruction of the landform map, local area minima need to be extracted from it. Specifically, the following steps are executed:
[0077] S24: Adopt the method of extracting eight - neighborhood pixels to obtain the local area minima in the reconstructed landform map and use them as the first - type control markers.
[0078] The first - type markers extracted through step S24 are as Figure 5 shown. Since the extraction of local minima is affected by the topographic map, although the landform map has undergone a series of filtering processes, it is still inevitable that there are two or more local minima in the same basin. For this problem, corresponding processing methods are proposed in the subsequent part of this embodiment.
[0079] First of all, complex foam images usually have the following characteristics: Minerals are enriched at the top of the bubbles in the foam layer. Under direct sunlight, the center of the bubble shows high reflectivity, which is manifested as local bright reflection areas on the image. The pixel values of each bubble center are the highest, and the pixel values gradually decrease towards the four - side boundaries of the bubble, making the gray - values of the image boundary points relatively low. AsFigure 6 As shown, the two high points of the bubble are outlined by the red rectangle, and the size of the highlighted area is proportional to the size of the bubble, and the number of highlighted areas is also proportional to the number of bubbles, that is, each highlighted area can correspond to a bubble.
[0080] The existing image segmentation methods based on convolutional neural networks have not been able to effectively incorporate the feature knowledge of the above-mentioned foam image boundary extraction, resulting in a large dependence on a large amount of data and low accuracy of image segmentation. Therefore, in this embodiment, the highlighted area of the bubble is extracted as the second type of marker to incorporate domain knowledge into the algorithm. Specifically, as shown in step S30.
[0081] S30: Denoise the foam image, and use the OTSU automatic threshold segmentation method based on the connected region mask to obtain the highlighted area of the flotation bubble as the second type of control marker.
[0082] Specifically, the denoising process of the input image in step S30 includes the following steps:
[0083] S31: At least use the Gaussian filtering method and the mean filtering method to remove noise from the image.
[0084] Step S31 is to make the image smoother and reduce the impact of noise on subsequent image processing.
[0085] S32: Enhance the local contrast of the image by comparing the limited adaptive histogram equalization.
[0086] The contrast-limited adaptive histogram equalization method in step S32 can improve the visibility of the highlighted area of the bubble in the image. Since the highlighted area of the bubble usually has obvious brightness or color differences from other areas of the bubble, by enhancing these differences, it can be more easily separated from the bubble.
[0087] Based on step S32, the method for fusing the highlighted areas to obtain the second type of marker includes the following steps:
[0088] S33: In each connected region of the flotation bubble obtained in S10, the OTSU method is respectively used for automatic threshold segmentation to obtain the highlighted area of each flotation bubble.
[0089] The Otsu automatic threshold segmentation method selects the optimal threshold by maximizing the between-class variance, thereby realizing the binarization of the image. In this way, the highlighted area of the bubble can be extracted.
[0090] S34: Merge the highlighted areas of each connected region of the flotation bubble into one image as the second type of control marker.
[0091] After the extraction of both the first - type control markers and the second - type control markers is completed, step S40 is executed to analyze the two types of extracted control markers.
[0092] S40: Analyze the connected region of the bubbles, fuse the first - type control markers and the second - type control markers, and obtain the fused control markers.
[0093] When there is one, two, or more than two situations in the same region, it is necessary to make a judgment according to the specific situation, which specifically includes the following steps:
[0094] S41: If there is only one first - type control marker in the connected region of the flotation bubbles, select the first - type control marker as the fused control marker.
[0095] As Figure 7 shown, it shows that the segmentation effect of the Unet model is good and can be directly used as a control marker for watershed algorithm segmentation.
[0096] S42: If there are two or more first - type control markers in the connected region of the flotation bubbles, extract the corresponding second - type control marker of this connected region as the fused control marker.
[0097] There are mainly the following two situations where there are two or more first - type control markers in the connected region of the flotation bubbles:
[0098] The first situation is that the predicted result output has an unclosed result line, that is, there is a break point. As Figure 8 shown, in the Unet prediction result, it can be clearly seen that there is a discontinuous boundary, which will cause two bubbles to be connected into one region. There are two first - type control markers and two second - type control markers in this connected region, indicating that this connected region is composed of two bubbles and there is an unclosed line. Then, select the second type as the control marker to guide the watershed algorithm segmentation.
[0099] The second situation is that the first - type markers are extracted incorrectly due to the interference of the terrain boundary shape. As Figure 9 shown, because the first - type control markers are affected by the boundary shape, even if the Unet model is segmented correctly, there may be a situation where the first - type control markers are extracted incorrectly, resulting in two or more first - type control markers in a connected region of a bubble. At this time, the second - type control marker (highlighted area) should be selected to guide the watershed segmentation.
[0100] The first type of control markers and the second type of control markers can complement each other. The first type of markers can provide accurate enough identification so that even smaller or dimmer bubbles can be effectively marked. The second type of markers is used to control the accuracy of the first type of markers: even if there are multiple bubbles in the same area, or the same bubble is wrongly marked as multiple first type of control markers, the second type of control markers can be adjusted through quantity constraints to ensure the accuracy of the control markers. After the extraction of the fused control markers in S42 is completed, step S50 is executed.
[0101] S50: According to the fused control markers extracted in S40 and the reconstructed geomorphic map extracted in S20, the control marker watershed algorithm is used for segmentation to obtain the final flotation bubble boundary and realize the measurement of the flotation bubble size.
[0102] Among them, the segmentation operation of the watershed algorithm is a technique widely used in the prior art, so the scheme thereof will not be elaborated in this embodiment. The result finally output through this embodiment is as Figure 10 shown.
[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes completed by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.
Claims
1. A method for measuring the size of flotation bubbles by fusing control markers with watershed and U-Net, characterized in that, It includes the following steps: S10: Use a U-Net convolutional neural network to perform preliminary segmentation on the foam image, obtain an initial bubble segmentation map, and extract the connected regions of each flotation bubble according to the initial bubble segmentation map; S20: Obtain a reconstructed geomorphic map according to the preliminary segmentation result, and extract the local area minimum value from the reconstructed geomorphic map as the first type of control marker; S30: Denoise the foam image, and use the OTSU automatic threshold segmentation method based on the connected region mask to obtain the highlighted area of the flotation bubble as the second type of control marker; S40: Analyze the connected regions of the bubbles, fuse the first type of control marker and the second type of control marker to obtain a fused control marker; S50: According to the fused control marker extracted in S40 and the reconstructed geomorphic map extracted in S20, use the control marker watershed algorithm for segmentation to obtain the final flotation bubble boundary and realize the measurement of the flotation bubble size.
2. The method for measuring the flotation bubble size by fusing control marker watershed and U-Net according to claim 1, wherein: The steps of obtaining the initial bubble segmentation map and the connected regions of each flotation bubble in step S10 include the following steps: S11: Construct a flotation bubble boundary dataset, and establish a flotation bubble boundary segmentation model based on U-Net. This model takes the foam image as the input and the flotation bubble boundary marker map as the output; S12: Perform preliminary segmentation on the input foam image through the U-Net convolutional neural network to obtain an initial bubble segmentation map; S13: According to the initial bubble segmentation map, use the image connected region tracking algorithm to extract the connected regions of each flotation bubble.
3. The method for measuring the flotation bubble size by fusing the control marker watershed and U-Net according to claim 1, wherein: The extraction of the first type of control marker in step S20 includes the following steps: S21: Perform Euclidean distance transformation on the initial bubble segmentation map, so that the pixel points closer to the preliminary predicted boundary points have lower values, and obtain the distance-transformed image; S22: Use an exponential decay function based on the Euler number to process the distance-transformed image, realize non-linear distance mapping and normalization processing, and make the pixel points closer to the preliminary predicted boundary points have higher values; S23: Perform gray value range expansion on the image processed in S22, and use Gaussian filtering to smooth the image to obtain the final reconstructed geomorphic map.
4. The method for measuring the flotation bubble size by fusing control marker watershed and U-Net according to claim 3, wherein: The Euclidean transformation depends on the following formula: D(p) = min q∈F ||p - q||₂ where D(p) is the value of pixel p in the distance transformation result image, F is the set of all boundary pixels, that is, non-zero values, and ||p - q||2 represents the Euclidean distance between pixel p and q.
5. The method for measuring the flotation bubble size by fusing control markers of watershed and U-Net according to claim 3, wherein: The acquisition of the non-linear distance mapping depends on the following formula: where D max is the maximum value among the values D(x, y) of the pixels p in the distance transformation result image.
6. The method for measuring the flotation bubble size by fusing control markers with watershed and U-Net according to claim 1, characterized in that: The method of extracting the local area minimum value from the reconstructed geomorphic map in step S20 depends on the following steps: S24: Use the method of extracting eight-neighborhood pixels to obtain the local area minimum value in the reconstructed geomorphic map and use it as the first type of control marker.
7. The method for measuring the flotation bubble size by fusing control marker watershed and U-Net according to claim 1, characterized in that: The denoising process of the input image in step S30 includes the following steps: S31: At least use the Gaussian filtering method and the mean filtering method to remove noise from the image; S32: Enhance the local contrast of the image by comparing the contrast limited adaptive histogram equalization.
8. The method for measuring the flotation bubble size by fusing control markers with watershed and U-Net according to claim 1, characterized in that: In step S30, the OTSU automatic threshold segmentation method based on the connected region mask is used to obtain the highlighted region of the flotation bubbles as the second type of control marker, including the following steps: S33: In each connected region of the flotation bubbles obtained in S10, the OTSU method is respectively used for automatic threshold segmentation to obtain the highlighted region of each flotation bubble; S34: The highlighted regions of each connected region of the flotation bubbles are merged into one graph as the second type of control marker.
9. The method for measuring the flotation bubble size by fusing control marker watershed and U-Net according to any one of claims 1-8, characterized in that: In step S40, obtaining the fusion control marker depends on the following steps for analysis: S41: If there is only one first type of control marker in the connected region of the flotation bubbles, select the first type of control marker as the fusion control marker; S42: If there are two or more first type of control markers in the connected region of the flotation bubbles, extract the corresponding second type of control marker of this connected region as the fusion control marker.