Mta-net deposition simulation experiment image particle size distribution prediction method and electronic equipment

Abstract

The application discloses a kind of MTA-Net deposition simulation experiment image particle size distribution prediction method and electronic equipment, belong to image processing technical field.The present application can not rely on single particle contour accurate segmentation, using the explicit multi-scale texture feature extracted by transverse multi-scale local binary pattern middle layer texture guide module to guide middle layer convolution feature, and by high layer channel-space joint attention enhancement module, enhance the discriminant feature related to particle size composition, to improve the feature expression ability under the condition of particle overlap, boundary fuzzy and multi-scale texture mixture.The MTA-Net particle size composition prediction model constructed by the present application can directly output the composition ratio of multiple preset particle size intervals, after Softmax normalization, each component is non-negative and the sum is 1, can represent the mixed composition of multiple particle levels in the same sample.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and electronic device for predicting particle size distribution in MTA-Net deposition simulation experimental images. Background Technology

[0002] Sedimentary simulation images record the spatial texture features formed by sedimentary particles under transport, deposition, and sorting processes, serving as a crucial data source for quantitative analysis of sedimentary processes. Existing grain size analysis methods mainly include sieving, sedimentation, and laser grain size analysis. While these methods offer high measurement reliability, they require point-by-point sampling and laboratory testing, resulting in long processing cycles and difficulty in directly reflecting the continuous spatial variations of sedimentary bodies. Sedimentary simulation images, however, can record grain arrangement, bedding distribution, grayscale fluctuations, and local texture structures in a non-contact manner, providing a foundation for rapid estimation of grain size distribution. Therefore, establishing an efficient and stable method for predicting grain size distribution from sedimentary simulation images is a critical issue that urgently needs to be addressed for the quantitative analysis of sedimentary simulation experiments.

[0003] Existing image-based particle size analysis methods mainly include particle contour segmentation methods and artificial texture feature methods. Particle contour segmentation methods typically use thresholding, watershed, or edge detection to identify particle boundaries and estimate particle size based on the particle contour. However, sedimentation simulation experimental images often contain particles that are in contact or overlapping, have blurred boundaries, local occlusion, and varying illumination, making it difficult to accurately and stably separate single particle contours. While artificial texture feature methods can describe local gray-level relationships and texture statistics through gray-level co-occurrence matrices, local binary patterns, or wavelet features, they struggle to simultaneously characterize particle alignment orientation, local gray-level fluctuations, and multi-scale texture differences corresponding to different particle size components. They also find it difficult to fully establish the nonlinear mapping relationship between complex texture features and particle size distribution.

[0004] Convolutional neural networks (CNNs) can automatically extract multi-level image features, providing a new technical approach for image-based grain size prediction. However, grain size information in sedimentation simulation images mainly manifests as grain arrangement, lateral distribution, local undulations, and multi-scale texture structures. Given a limited number of sedimentation experiment samples, relying solely on CNNs is insufficient to reliably utilize the explicit statistical regularities that signify sedimentary texture. Furthermore, the same sedimentation image typically contains a mixture of particles from multiple grain size ranges, with different grain sizes collectively forming the image texture. Therefore, this task is not about determining a single grain size category or predicting a single grain size value, but rather about simultaneously predicting the continuous proportions of multiple grain size ranges. Summary of the Invention

[0005] The purpose of this invention is to overcome the technical problems existing in the prior art, such as unstable contour segmentation caused by particle overlap and boundary blurring, the single scale and limited expressive power of traditional artificial texture features, the difficulty of pure convolutional networks to stably learn explicit texture rules under limited sample conditions, and the inability of classification or single particle size regression to represent mixed particle size composition. The invention provides an MTA-Net deposition simulation experimental image particle size distribution prediction method and electronic device.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a method for predicting particle size distribution in MTA-Net deposition simulation experimental images is provided, including the following steps: S1: Acquire images of sedimentation simulation experiments and surface sediment samples, and preprocess them to obtain an image-grain size composition label dataset; S2: Divide the image-particle size composition label dataset into a training set, a validation set, and a test set, and train the constructed MTA-Net particle size composition prediction model. The MTA-Net particle size distribution prediction model includes a convolutional feature extraction backbone, a mid-layer texture guidance module for horizontal multi-scale local binary patterns, a high-level channel-space joint attention enhancement module, and a particle size distribution prediction module. The convolutional feature extraction backbone is used to extract multi-level convolutional features from the input sedimentation simulation experimental image, which evolve from shallow local features to high-level semantic features. The mid-layer texture guidance module of the horizontal multi-scale local binary mode is used to extract explicit multi-scale texture features and perform residual channel-wise modulation on the mid-layer convolution features output by the convolution feature extraction backbone. The high-level channel-spatial joint attention enhancement module is used to adaptively filter the high-level semantic features output by the convolutional feature extraction backbone, and enhance the feature channels and spatial regions related to particle size composition prediction. The particle size distribution prediction module is used to convert the attention-enhanced high-level semantic features into the composition ratio of multiple preset particle size intervals. S3: Input the sedimentation simulation experiment image to be tested into the trained MTA-Net particle size composition prediction model, and output the corresponding particle size composition ratio.

[0007] As a preferred embodiment of the present invention, step S1, which involves obtaining sedimentation simulation experimental images and grain size composition data, includes: S11: Determine the cutting location, direction, and route based on the morphology of the sedimentary body and the requirements for the layout of the experimental profile, and cut the sedimentary body accordingly; S12: Within the same target area, acquire images of sedimentation simulation experiments and extract regions of interest to form image samples; S13: Collect surface sediment samples corresponding to the space of the region of interest, and perform grain size testing to obtain the grain size test results; S14: Establish a one-to-one correspondence between the image sample and the particle size test result based on the image sampling location and sample number.

[0008] As a preferred embodiment of the present invention, the preprocessing in step S1 includes: unifying the size, tensorizing and standardizing the values ​​of the image samples; summarizing the particle size test results into multiple preset particle size intervals, calculating the volume fraction of each particle size interval and normalizing it to form a particle size composition label with each component being non-negative and the sum being 1.

[0009] As a preferred embodiment of the present invention, the convolutional feature extraction backbone in step S2 includes the following process: the preprocessed deposition simulation experimental image is sequentially passed through the initial convolutional layer, the first residual module, and the second residual module to obtain the mid-level convolutional features containing information on particle edges, local grayscale changes, and spatial structure. The output mid-level convolutional features are first input into the mid-level texture guidance module of the lateral multi-scale local binary mode for texture modulation. The modulated mid-level convolutional features are then sequentially input into the third residual module and the fourth residual module to extract high-level semantic features.

[0010] As a preferred embodiment of the present invention, the transverse multi-scale local binary mode mid-layer texture guidance module in step S2 includes the following process: The deposition simulation experiment image was converted into a grayscale image, and lateral enhancement sampling points were set within a neighborhood radius r. The set of lateral enhancement sampling points is as follows: ; Among them, the first four lateral enhancement sampling points are used to characterize the lateral arrangement of particles and the continuity of layering, while the remaining lateral enhancement sampling points are used to supplement vertical grayscale changes and local structural information. The grayscale values ​​of each lateral enhancement sampling point are compared with the center pixel to form a local binary code, and mapped to 59 classes using the uniform-u2 mode. The effective coding area is divided into 4×4 sub-regions, and 59-dimensional histograms are statistically analyzed and spliced ​​in a fixed order to obtain a 944-dimensional multi-scale texture vector. The 944-dimensional multi-scale texture vector is sequentially passed through layer normalization, a first fully connected layer, batch normalization, an activation layer, and a second fully connected layer, mapped to a modulation vector with the same number of output channels as the second residual module, and reconstructed into a C×1×1 channel modulation tensor M. The mid-layer texture modulation is represented as: ; in, The mid-level convolutional features output by the second residual module. The channel modulation tensor is obtained by explicit multi-scale texture feature mapping and reconstruction. This indicates element-wise multiplication broadcast via a channel. This is the learnable scaling factor.

[0011] As a preferred embodiment of the present invention, the high-level channel-space joint attention enhancement module in step S2 includes the following process: the high-level channel-space joint attention enhancement module adopts a convolutional block attention module and is set after the complete output of the fourth residual module. The convolutional block attention module first weights different feature channels through channel attention, and then weights different spatial positions through spatial attention to obtain the high-level semantic features after attention enhancement.

[0012] As a preferred embodiment of the present invention, the particle size distribution prediction module in step S2 includes a global average pooling layer, a fully connected prediction head, and a Softmax composition normalization layer. First, the spatial dimension of high-level semantic features is compressed by the global average pooling layer to obtain image-level global features. Then, the image-level global features are input into the fully connected prediction head, and the unnormalized prediction values ​​are output in numbers equal to the preset particle size intervals. Finally, the composition is normalized by the Softmax composition normalization layer to obtain a particle size composition ratio vector in which each component is non-negative and the sum is 1.

[0013] As a preferred embodiment of the present invention, the fully connected prediction head includes a fully connected layer, a batch normalization layer, an activation layer, a dropout layer, and an output layer connected in sequence, ultimately outputting five unnormalized prediction values. 9. An electronic device, characterized in that it includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the MTA-Net deposition simulation experimental image particle size distribution prediction method as described above.

[0014] On the other hand, an electronic device is provided, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the MTA-Net deposition simulation experimental image particle size distribution prediction method described above.

[0015] Compared with the prior art, the advantages of the present invention are as follows: The present invention can guide the mid-level convolution features by using the explicit multi-scale texture features extracted by the mid-level texture guidance module of the lateral multi-scale local binary mode without relying on the accurate segmentation of single particle contours, and enhance the discriminative features related to particle size composition through the high-level channel-space joint attention enhancement module, thereby improving the feature expression ability under the conditions of particle overlap, boundary ambiguity and multi-scale texture mixing.

[0016] The MTA-Net particle size composition prediction model constructed in this invention can directly output the composition ratio of multiple preset particle size intervals. After Softmax normalization, each component is non-negative and the sum is 1, which can characterize the mixed composition of multiple particle sizes in the same sample. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a flowchart of a method for predicting particle size distribution in MTA-Net deposition simulation experimental images as described in Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating the construction of a sedimentation simulation experiment image and a grain size composition label dataset for a method to predict grain size distribution in MTA-Net sedimentation simulation experiment images, as described in Embodiment 1 of the present invention. Figure 3 This is an internal structure diagram of the MTA-Net particle size composition prediction model of the MTA-Net deposition simulation experimental image particle size distribution prediction method described in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the explicit multi-scale (HMS-LBP) texture feature extraction process of the MTA-Net deposition simulation experimental image particle size distribution prediction method described in Embodiment 1 of the present invention; Figure 5 This is a flowchart of the overall method for predicting the particle size composition of sedimentation images according to the MTA-Net sedimentation simulation experimental image particle size distribution prediction method described in Embodiment 1 of the present invention. Figure 6 This is a schematic diagram of the structure of an electronic device according to Embodiment 2 of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] Example 1 A method for predicting particle size distribution in MTA-Net deposition simulation experimental images, such as Figure 1 As shown, it includes the following steps: S1: Acquire images of sedimentation simulation experiments and surface sediment samples, and preprocess them to obtain an image-grain size composition label dataset; Preferably, the acquisition of sedimentation simulation experiment images and grain size composition data in step S1 includes, for example: Figure 2 As shown: S11: Determine the cutting location, direction, and route based on the morphology of the sedimentary body and the requirements for the layout of the experimental profile, and cut the sedimentary body. This includes determining the cutting location, which may include: determining the cross section, determining the longitudinal section, and other types. S12: Within the same target area, acquire images of sedimentation simulation experiments and extract regions of interest (ROI) to form image samples; S13: Collect surface sediment samples corresponding to the space of the region of interest, and perform grain size testing to obtain the grain size test results; S14: Establish a one-to-one correspondence between the image sample and the particle size test result based on the image sampling location and sample number.

[0020] Preferably, the preprocessing in step S1 includes: unifying the size, tensorizing, and standardizing the values ​​of the image samples. In this embodiment, the image is scaled to maintain its original aspect ratio and filled to 224×224 pixels, and the pixel values ​​are scaled to [0,1] before standardization. During the training phase, only the brightness, contrast, and saturation of the convolution branch input are perturbed; the texture guidance module in the horizontal multi-scale local binary mode always extracts texture features based on the corresponding grayscale image with unified size and without random color perturbation, and no random enhancement is performed during the verification, testing, and inference phases.

[0021] The particle size test results are summarized into multiple preset particle size intervals. The volume fraction of each particle size interval is calculated and normalized to form a particle size composition label with non-negative components and a sum of 1. Based on the sampling location and sample number, image samples are matched one-to-one with particle size composition labels to construct an image-particle size composition label dataset. In this embodiment, the particle size intervals include less than 100 μm, 100–250 μm, 250–500 μm, 500–1000 μm, and greater than 1000 μm.

[0022] S2: Divide the image-particle size composition label dataset into a training set, a validation set, and a test set, and train the constructed MTA-Net particle size composition prediction model. The training set is used for learning the parameters of the MTA-Net particle size composition prediction model, the validation set is used for monitoring the training process, model selection, and parameter adjustment, and the test set is only used for evaluating the performance of the MTA-Net particle size composition prediction model. Specifically, in this embodiment, a total of 202 sets of valid image-granularity label samples were obtained, which were divided into 145 training samples, 26 validation samples, and 31 test samples. The granularity range with the largest proportion in the sample labels was used as the main stratification basis, and the experimental round information was also taken into account to maintain the consistency of granularity composition and experimental source distribution of different data subsets as much as possible.

[0023] The MTA-Net particle size composition prediction model is as follows: Figure 3 As shown, it includes a convolutional feature extraction backbone, a mid-layer texture guidance module for Horizontal Multi-Scale Local Binary Pattern (HMS-LBP), a high-level channel-space joint attention enhancement module (CBAM), and a particle size distribution prediction module. The convolutional feature extraction backbone is used to extract multi-level convolutional features from the input sedimentation simulation experimental image, which evolve from shallow local features to high-level semantic features. Preferably, the convolutional feature extraction backbone in step S2 includes the following process: the preprocessed deposition simulation experiment image is sequentially passed through the initial convolutional layer, the first residual module (stage1), and the second residual module (stage2) to obtain mid-level convolutional features containing information on particle edges, local grayscale changes, and spatial structure. The output mid-level convolutional features are first input into the mid-level texture guidance module of the lateral multi-scale local binary mode for texture modulation. The modulated mid-level convolutional features are then sequentially input into the third residual module (stage3) and the fourth residual module (stage4) to extract high-level semantic features.

[0024] The mid-layer texture guidance module of the horizontal multi-scale local binary mode is used to extract explicit multi-scale texture features and perform residual channel-wise modulation on the mid-layer convolution features output by the convolution feature extraction backbone. Preferably, the transverse multi-scale local binary mode mid-layer texture guidance module in step S2 includes the following process, such as... Figure 4 As shown: The deposition simulation experiment image was converted into a grayscale image, and lateral enhancement sampling points were set within a neighborhood radius r. The set of lateral enhancement sampling points is as follows: ; The first four lateral enhancement sampling points are used to characterize the lateral arrangement of particles and the continuity of stratification. The remaining lateral enhancement sampling points are used to supplement the vertical grayscale changes and local structural information. The grayscale values ​​of each lateral enhancement sampling point are compared with the center pixel to form a local binary code, and mapped to 59 classes using the uniform-u2 mode. The effective coding area is divided into 4×4 sub-regions. The 59-dimensional histograms are statistically analyzed and spliced ​​in a fixed order to obtain a 944-dimensional single-scale texture vector. In this embodiment, texture features are extracted at r=1 and r=2 respectively, and element-wise averaging fusion is performed to obtain a 944-dimensional HMS-LBP multi-scale texture vector. Specifically, the corresponding single-scale texture vectors are obtained under two different neighborhood radii (r={1,2} in this paper), and the single-scale texture vectors are fused element-wise to obtain the HMS-LBP multi-scale fused texture vector with a dimension of 944.

[0025] The 944-dimensional HMS-LBP multi-scale fused texture vector is sequentially passed through layer normalization, a first fully connected layer, batch normalization, an activation layer, and a second fully connected layer, mapped to a modulation vector with the same number of output channels as the second residual module, and reconstructed into a C×1×1 channel modulation tensor M. The mid-layer texture modulation is represented as: ; in, The mid-level convolutional features output by the second residual module. The channel modulation tensor is obtained by explicit multi-scale texture feature mapping and reconstruction. This indicates element-wise multiplication broadcast via a channel. This is the learnable scaling factor.

[0026] The high-level channel-spatial joint attention enhancement module is used to adaptively filter the high-level semantic features output by the convolutional feature extraction backbone, and enhance the feature channels and spatial regions related to particle size composition prediction. Preferably, the high-level channel-space joint attention enhancement module in step S2 includes the following process: the high-level channel-space joint attention enhancement module adopts a convolutional block attention module and is set after the complete output of the fourth residual module. The convolutional block attention module first weights different feature channels through channel attention, and then weights different spatial positions through spatial attention to obtain the attention-enhanced high-level semantic features.

[0027] The mid-level texture guidance module of the horizontal multi-scale local binary mode is responsible for introducing explicit texture priors in the mid-level, while the high-level channel-space joint attention enhancement module is responsible for filtering discriminative semantic features in the high-level. The two work together in the order of mid-level texture guidance first and high-level attention enhancement last.

[0028] The particle size distribution prediction module is used to convert the attention-enhanced high-level semantic features into the composition ratio of multiple preset particle size intervals. Preferably, the particle size distribution prediction module in step S2 includes a global average pooling layer, a fully connected prediction head, and a Softmax composition normalization layer, such as... Figure 5 The flowchart of the overall method for predicting the particle size composition of sedimentation images is shown. First, the spatial dimension of high-level semantic features is compressed through a global average pooling layer to obtain image-level global features. Then, the image-level global features are input into a fully connected prediction head, which outputs unnormalized prediction values ​​equal to the number of preset particle size intervals. Finally, the composition is normalized through a Softmax composition normalization layer to obtain a particle size composition ratio vector in which each component is non-negative and the sum is 1.

[0029] Furthermore, the fully connected prediction head includes a fully connected layer, a batch normalization layer, an activation layer, a Dropout layer, and an output layer connected in sequence, ultimately outputting five unnormalized prediction values.

[0030] Furthermore, the five unnormalized predicted values ​​are normalized by a Softmax layer to obtain five components, expressed as follows: , ;

[0031] The five components correspond to: <100μm; 100~250μm; 250~500μm; 500~1000μm; >1000μm.

[0032] Specifically, during the training of the MTA-Net particle size distribution prediction model, the actual particle size distribution labels corresponding to the training samples are compared with the predicted composition proportions output by the MTA-Net particle size distribution prediction model via Softmax. A loss function is calculated based on the composition distribution error between the two, and the model parameters are updated through backpropagation. The process involves using a composite distribution loss consisting of cumulative distribution loss, absolute error loss, and cosine loss to constrain the ordered cumulative deviation between adjacent particle size intervals, the proportional deviation between each particle size interval, and the overall difference in particle size distribution morphology, respectively. The trained model parameters are selected based on the loss or evaluation metrics on the validation set.

[0033] S3: Input the sedimentation simulation experiment image to be tested into the trained MTA-Net particle size composition prediction model, and output the corresponding particle size composition ratio.

[0034] Further, the sedimentation simulation experiment images to be predicted are acquired, and the RGB images are processed according to the same size unification and numerical standardization method as in the training phase. Explicit multi-scale texture features are extracted from the corresponding grayscale images. Both are input into the trained MTA-Net particle size composition prediction model, which undergoes mid-layer texture modulation, high-layer attention enhancement, global average pooling, prediction head, and Softmax processing to output the composition ratios of multiple preset particle size intervals. In this embodiment, a five-dimensional vector is output, corresponding to less than 100μm, 100–250μm, 250–500μm, 500–1000μm, and greater than 1000μm, respectively, with each component being non-negative and summing to 1.

[0035] This invention addresses the instability of contour segmentation caused by particle overlap and blurred boundaries. Instead of relying on single-particle contours, it constructs explicit multi-scale texture features from image grayscale information and uses these features to guide mid-level convolutional features. To address the limitations of traditional artificial texture features, which have a single scale and limited expressive power, it enhances particle arrangement, grayscale fluctuations, and texture differences at different scales through lateral enhancement sampling and multi-scale local statistical characterization. A high-level channel-space joint attention enhancement module strengthens discriminative features related to particle size composition, thereby improving feature representation under conditions of particle overlap, blurred boundaries, and mixed multi-scale textures. To address the difficulty of pure convolutional networks stably learning explicit texture patterns under limited sample conditions, it maps explicit multi-scale texture features to channel modulation information and performs residual-based channel-wise modulation on mid-level convolutional features. To address the issue of effective responses being mixed with irrelevant information in high-level features, channel attention and spatial attention enhancement are performed sequentially after the high-level semantic feature output. Finally, to address the inability of classification or single-particle-size regression to represent mixed particle size composition, a prediction head and Softmax jointly output the composition ratios of multiple preset particle size intervals. This leads to a hierarchical collaborative technical solution of "explicit multi-scale texture construction - mid-level texture modulation - high-level attention enhancement - multi-granularity composition ratio joint prediction".

[0036] The MTA-Net particle size distribution prediction model can directly output the composition ratios of multiple preset particle size intervals. After Softmax normalization, each component is non-negative and its sum is 1, which can characterize the mixed composition of multiple particle sizes in the same sample. The results of this embodiment show that the χ² distance, MAE, and NRMSE of MTA-Net are 0.0125±0.0023, 0.0310±0.0024, and 16.53±0.36%, respectively, which are reduced by 45.41%, 29.71%, and 12.63% compared with ResNet34. This can serve as a representative embodiment of the technical solution that can improve the accuracy of particle size distribution prediction and improve the stability of multiple training results.

[0037] Example 2 like Figure 6 As shown, an electronic device includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to perform a particle size distribution prediction method for MTA-Net deposition simulation experimental images as described in the foregoing embodiments. The input / output interface may include a display, keyboard, mouse, and USB interface for inputting and outputting data; a power supply provides electrical power to the electronic device.

[0038] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks. When the integrated units of the present invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0039] The preferred embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings. However, the present disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0040] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0041] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for predicting the particle size distribution of a MTA-Net deposition simulation experiment image, characterized in that, Includes the following steps: S1: Acquire images of sedimentation simulation experiments and surface sediment samples, and preprocess them to obtain an image-grain size composition label dataset; S2: Divide the image-particle size composition label dataset into a training set, a validation set, and a test set, and train the constructed MTA-Net particle size composition prediction model. The MTA-Net particle size distribution prediction model includes a convolutional feature extraction backbone, a mid-layer texture guidance module for horizontal multi-scale local binary patterns, a high-level channel-space joint attention enhancement module, and a particle size distribution prediction module. The convolutional feature extraction backbone is used to extract multi-level convolutional features from the input sedimentation simulation experimental image, which evolve from shallow local features to high-level semantic features. The mid-layer texture guidance module of the horizontal multi-scale local binary mode is used to extract explicit multi-scale texture features and perform residual channel-wise modulation on the mid-layer convolution features output by the convolution feature extraction backbone. The high-level channel-spatial joint attention enhancement module is used to adaptively filter the high-level semantic features output by the convolutional feature extraction backbone, and enhance the feature channels and spatial regions related to particle size composition prediction. The particle size distribution prediction module is used to convert the attention-enhanced high-level semantic features into the composition ratio of multiple preset particle size intervals. S3: Input the sedimentation simulation experiment image to be tested into the trained MTA-Net particle size composition prediction model, and output the corresponding particle size composition ratio.

2. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 1, characterized in that, Step S1, which involves obtaining sedimentation simulation experiment images and grain size composition data, includes: S11: Determine the cutting location, direction, and route based on the morphology of the sedimentary body and the requirements for the layout of the experimental profile, and cut the sedimentary body accordingly; S12: Within the same target area, acquire images of sedimentation simulation experiments and extract regions of interest to form image samples; S13: Collect surface sediment samples corresponding to the space of the region of interest, and perform particle size testing to obtain the particle size test results; S14: Establish a one-to-one correspondence between the image sample and the particle size test result based on the image sampling location and sample number.

3. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 2, characterized in that, The preprocessing described in step S1 includes: unifying the size, tensorizing, and standardizing the values ​​of the image samples; summarizing the particle size test results into multiple preset particle size intervals, calculating the volume fraction of each particle size interval and normalizing it to form a particle size composition label with non-negative components and a sum of 1.

4. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 1, characterized in that, The convolutional feature extraction backbone in step S2 includes the following process: the preprocessed deposition simulation experiment image is sequentially passed through the initial convolutional layer, the first residual module, and the second residual module to obtain mid-level convolutional features containing information on particle edges, local grayscale changes, and spatial structure. The output mid-level convolutional features are first input into the mid-level texture guidance module of the lateral multi-scale local binary mode for texture modulation. The modulated mid-level convolutional features are then sequentially input into the third residual module and the fourth residual module to extract high-level semantic features.

5. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 1, characterized in that, The transverse multi-scale local binary mode mid-layer texture guidance module described in step S2 includes the following processes: The deposition simulation experiment image was converted into a grayscale image, and lateral enhancement sampling points were set within a neighborhood radius r. The set of lateral enhancement sampling points is as follows: ； Among them, the first four lateral enhancement sampling points are used to characterize the lateral arrangement of particles and the continuity of layering, while the remaining lateral enhancement sampling points are used to supplement vertical grayscale changes and local structural information. The grayscale values ​​of each lateral enhancement sampling point are compared with the center pixel to form a local binary code, and mapped to 59 classes using the uniform-u2 mode. The effective coding area is divided into 4×4 sub-regions, and 59-dimensional histograms are statistically analyzed and spliced ​​in a fixed order to obtain a 944-dimensional multi-scale texture vector. The 944-dimensional multi-scale texture vector is sequentially passed through layer normalization, a first fully connected layer, batch normalization, an activation layer, and a second fully connected layer, mapped to a modulation vector with the same number of output channels as the second residual module, and reconstructed into a C×1×1 channel modulation tensor M. The mid-layer texture modulation is represented as: ； in, The mid-level convolutional features output by the second residual module. The channel modulation tensor is obtained by explicit multi-scale texture feature mapping and reconstruction. This indicates element-wise multiplication broadcast via a channel. This is the learnable scaling factor.

6. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 1, characterized in that, The high-level channel-space joint attention enhancement module described in step S2 includes the following process: The high-level channel-space joint attention enhancement module adopts a convolutional block attention module and is set after the complete output of the fourth residual module. The convolutional block attention module first weights different feature channels through channel attention, and then weights different spatial positions through spatial attention to obtain the high-level semantic features after attention enhancement.

7. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 1, characterized in that, The particle size distribution prediction module described in step S2 includes a global average pooling layer, a fully connected prediction head, and a Softmax composition normalization layer. First, the spatial dimension of high-level semantic features is compressed through the global average pooling layer to obtain image-level global features. Subsequently, the image-level global features are input into the fully connected prediction head, which outputs unnormalized prediction values ​​equal to the number of preset particle size intervals. Finally, composition normalization is performed through a Softmax composition normalization layer to obtain a particle size composition ratio vector where each component is non-negative and the sum is 1.

8. The method for predicting grain size distribution in MTA-Net deposition simulation experimental images according to claim 7, characterized in that, The fully connected prediction head includes a fully connected layer, a batch normalization layer, an activation layer, a Dropout layer, and an output layer connected in sequence, and finally outputs five unnormalized prediction values.

9. An electronic device, characterized in that, The method includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the MTA-Net deposition simulation experimental image particle size distribution prediction method according to any one of claims 1 to 8.

Images