Aggregate water content detection method and device, storage medium and equipment

Abstract

An aggregate moisture content detection method and device, a storage medium and equipment. The method comprises: acquiring image feature data and reflected spectrum intensity data of an aggregate to be detected; performing feature mapping on the reflected spectrum intensity data to obtain reflected spectrum feature data aligned with the dimension of the image feature data; performing feature fusion on the image feature data and the feature-mapped reflected spectrum feature data to obtain fused aggregate feature data; and obtaining the moisture content of the aggregate to be detected based on the fused aggregate feature data. The above scheme can improve the detection accuracy of the aggregate moisture content.

Description

Technical Field

[0001] This invention relates to the field of aggregate testing technology, specifically to a method, apparatus, storage medium, and equipment for detecting aggregate moisture content. Background Technology

[0002] Concrete aggregates refer to the granular, loose materials that act as a skeleton or filler in concrete. They are specifically divided into coarse aggregates and fine aggregates. Coarse aggregates include pebbles and crushed stone, while fine aggregates include natural sand and manufactured sand. The moisture content of the aggregates is one of the most critical factors affecting concrete quality. Excessive moisture content will result in an overly thin final concrete. Therefore, it is necessary to monitor the moisture content of the aggregates in real time to facilitate concrete mixing.

[0003] However, existing methods for detecting aggregate moisture content have poor accuracy. Summary of the Invention

[0004] The problem this invention aims to solve is: how to improve the accuracy of aggregate moisture content detection.

[0005] To address the above problems, embodiments of the present invention provide a method for detecting aggregate moisture content, the method comprising:

[0006] Acquire image feature data and reflectance spectral intensity data of the aggregate to be tested;

[0007] The reflectance spectral intensity data is feature-mapped to obtain reflectance spectral feature data aligned with the dimensions of the image feature data;

[0008] The image feature data and the feature-mapped reflectance spectral feature data are fused to obtain the fused aggregate feature data.

[0009] Based on the fused aggregate characteristic data, the moisture content of the aggregate to be tested is obtained.

[0010] In one possible embodiment, acquiring image feature data of the aggregate to be tested includes:

[0011] Acquire images of the aggregate to be tested;

[0012] Preprocess the acquired images;

[0013] The preprocessed image is input into a preset image feature extraction network to obtain the image feature data of the aggregate to be tested.

[0014] In one possible embodiment, the preset image feature extraction network includes: an EfficientNetV2 backbone network and at least one DICE module; wherein the EfficientNetV2 backbone network is used to extract global semantic features of the preprocessed image, and the DICE module performs fine-grained feature extraction.

[0015] In one possible embodiment, the EfficientNetV2 backbone network includes: a 3*3 convolutional module, a first MBConv module, and a second MBConv module, wherein the first MBConv module is a 1x expanded inverted bottleneck k*k convolutional module, and the second MBConv module is a 6x expanded inverted bottleneck k*k convolutional module, where k*k is the depth of the convolutional kernel.

[0016] In one possible embodiment, the DICE module includes:

[0017] The local fusion submodule is adapted to perform depth-separable convolution operations on the output of the EfficientNetV2 backbone network along the channel direction, width direction and height direction respectively, and to concatenate the results of the depth-separable convolution operations in the three directions.

[0018] The global fusion submodule is suitable for performing channel shuffling on the concatenation results of depth-separable convolution operations in three directions, performing dimensionality reduction on the results after channel shuffling, and then using a channel attention mechanism to adaptively adjust the channel feature weights.

[0019] In one possible embodiment, the step of fusing the image feature data and reflectance spectral feature data to obtain fused aggregate feature data includes:

[0020] The image feature data and reflectance spectrum feature data are stitched together to obtain the stitched feature data;

[0021] The spliced ​​feature data is input into a fused fully connected network to obtain fused aggregate feature data.

[0022] In one possible embodiment, the aggregate is natural sand.

[0023] This invention also provides an aggregate moisture content detection device, the device comprising:

[0024] The acquisition unit is suitable for acquiring image feature data and reflectance spectral intensity data of the aggregate to be tested;

[0025] The feature mapping unit is adapted to perform feature mapping on the reflectance spectral intensity data to obtain reflectance spectral feature data aligned with the dimensions of the image feature data;

[0026] The fusion unit is adapted to fuse the image feature data with the feature-mapped reflectance spectral feature data to obtain fused aggregate feature data.

[0027] The prediction unit is adapted to obtain the moisture content of the aggregate to be tested based on the fused aggregate characteristic data.

[0028] This invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of any of the methods described above.

[0029] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the steps of any of the methods described above when running the computer program.

[0030] Compared with the prior art, the technical solution of the embodiments of the present invention has the following advantages:

[0031] By applying the solution of this invention, image feature data and reflectance spectral intensity data of the aggregate to be tested are acquired, and then the moisture content of the aggregate to be tested is predicted based on these two feature data. This is equivalent to predicting the moisture content of the aggregate to be tested based solely on the reflectance spectral intensity data. The image feature data of the aggregate to be tested can capture the color and texture information of the aggregate to be tested. Therefore, when predicting the moisture content of the aggregate to be tested based on the feature data after the fusion of the image feature data and the reflectance spectral intensity data, the influence of the color of the aggregate to be tested and the surface water film on the moisture content of the aggregate to be tested can be corrected, thereby improving the accuracy of the moisture content prediction. Attached Figure Description

[0032] Figure 1 This is a flowchart of a method for detecting aggregate moisture content in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of an aggregate moisture content detection process in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the structure of an EffiDiceNet network according to an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the calculation process of a first MBConv module in an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the calculation process of a second MBConv module in an embodiment of the present invention;

[0037] Figure 6 This is a schematic diagram of the structure of a DICE module in an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram illustrating the principle of a channel attention mechanism in an embodiment of the present invention;

[0039] Figure 8 This is a schematic diagram of the structure of an aggregate moisture content detection device according to an embodiment of the present invention. Detailed Implementation

[0040] Traditional aggregate moisture content testing relies on manual experience and offline laboratory methods (such as the drying method, which requires 2-3 hours to measure moisture content), which has drawbacks such as time lag, large subjective bias, and inability to provide real-time guidance for production.

[0041] To improve timeliness and accuracy, most current methods employ intelligent detection solutions for aggregate moisture content testing. Specifically, near-infrared sensors are used to illuminate the aggregate surface to obtain the intensity of the aggregate's reflectance spectrum. Then, the moisture content of the aggregate can be estimated using the pre-obtained mapping relationship between the reflectance spectrum intensity and the actual moisture content of the aggregate.

[0042] However, during the aforementioned intelligent detection process, the intensity of the reflected spectrum obtained by the near-infrared sensor is subject to certain errors due to interference from environmental factors such as aggregate color and surface water film, resulting in a sharp increase in errors in the high moisture content area of ​​the aggregate.

[0043] Specifically, different colored aggregates have different reflectivities for near-infrared light. Darker aggregates absorb more light, resulting in a lower reflectance spectral intensity, which can lead to the aggregate being misjudged as having high moisture content, while lighter-colored aggregates are the opposite. Furthermore, near-infrared spectroscopy cannot distinguish between adsorbed water (internal moisture) and surface water films. Surface water films form a highly reflective layer on the aggregate surface, which can cause near-infrared sensors to mistakenly interpret as sufficient internal moisture, leading to a lower reflectance spectral intensity and also resulting in the aggregate being misjudged as having high moisture content. For example, when the actual moisture content of natural sand is >10%, the error can reach 2.5%, resulting in a 5%–8% waste of cement and an increased risk of uncontrolled water-cement ratio.

[0044] To address this problem, the present invention provides a method for detecting aggregate moisture content. This method combines image feature data and reflectance spectral intensity data of the aggregate to be tested to predict its moisture content, which is equivalent to predicting the moisture content of the aggregate based solely on reflectance spectral intensity data. The image feature data of the aggregate to be tested can capture its color and texture information, thereby correcting the influence of the aggregate's color and surface water film on its moisture content and improving the accuracy of moisture content prediction.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Reference Figure 1 This invention provides a method for detecting the moisture content of aggregates, which may include the following steps:

[0047] Step 11: Obtain image feature data and reflectance spectral intensity data of the aggregate to be tested.

[0048] In practice, the surface of the aggregate to be tested is irradiated non-contactly using an infrared spectral sensor, and the collected reflectance spectral intensity data is directly used as the obtained reflectance spectral intensity data.

[0049] In some embodiments, reflectance spectral intensity data of different regions on the surface of the aggregate to be tested can also be collected, and the average value of the reflectance spectral intensity data of each collected region can be calculated, and the average value result can be used as the obtained reflectance spectral intensity data.

[0050] In specific implementation, refer to Figure 2 High-definition industrial cameras can be used to capture real-time images of the aggregate under test, obtaining raw image data. Regions of interest within the raw image data can then be cropped to obtain RGB image sequences (≥10 frames per second) for subsequent image feature extraction. For example, a high-definition industrial camera can be installed above a conveyor belt carrying natural sand to capture real-time images of the natural sand on the conveyor belt.

[0051] In practice, after obtaining the RGB image sequence captured by the camera, the acquired images can be preprocessed first, and then the preprocessed images can be input into a preset image feature extraction network to obtain the image feature data of the aggregate to be tested.

[0052] The process involves preprocessing the acquired images, including but not limited to denoising, white balance, and exposure correction, to eliminate interference caused by complex environmental factors such as dust, lighting changes, and vibration, and to ensure that the image quality meets the input requirements of the preset image feature extraction network.

[0053] In practice, the preset image feature extraction network can be implemented using various network structures. There are no restrictions here, as long as the preset image feature extraction network can distinguish between color changes and actual moisture content.

[0054] Step 12: Perform feature mapping on the reflectance spectral intensity data to obtain reflectance spectral feature data aligned with the dimensions of the image feature data.

[0055] In specific implementation, refer to Figure 2The reflectance spectral intensity data has a dimension of 256, while the image feature data typically has a dimension of 64. By performing feature mapping on the reflectance spectral intensity data, reflectance spectral feature data with the same dimension as the image feature data can be obtained. For example, when the image feature data has a dimension of 64, the reflectedance spectral feature data after feature mapping will also have a dimension of 64.

[0056] In practical implementation, feature mapping of reflectance spectral intensity data can be performed using linear normalization combined with a fully connected network. Specifically, linear normalization is used to compress the reflectance spectral intensity data corresponding to different wavelengths into the [0, 1] interval. The result of linear normalization is then input into a fully connected network, which projects the linearly normalized vector onto a latent space of the same dimension as the image feature data, thus obtaining the feature-mapped reflectance spectral feature data.

[0057] Step 13: Perform feature fusion between the image feature data and the feature-mapped reflectance spectral feature data to obtain fused aggregate feature data.

[0058] In practice, after the image feature data and the feature-mapped reflectance spectral feature data are concatenated, the fused aggregate feature data can be obtained. The dimension of the fused aggregate feature data is the sum of the dimensions of the two types of data. For example, when both the image feature data and the feature-mapped reflectance spectral feature data are 64-dimensional, the fused aggregate feature data will be 128-dimensional.

[0059] In specific implementation, refer to Figure 2 Image feature data and feature-mapped reflectance spectral feature data can be input into a fusion fully connected network to achieve cross-modal information complementarity and fusion of the two.

[0060] Step 14: Based on the fused aggregate characteristic data, obtain the moisture content of the aggregate to be tested.

[0061] In practice, after obtaining the aggregate characteristic data after fusion, the moisture content of the aggregate can be determined by finding the mapping relationship between the pre-obtained reflection spectrum intensity and the actual moisture content of the aggregate.

[0062] Since the aggregate characteristic data after fusion can better reflect the true characteristics of the aggregate to be tested, the determination of the moisture content of the aggregate to be tested using the aggregate characteristic data after fusion can make the moisture content test results of the aggregate to be tested more accurate.

[0063] To better distinguish between background and moisture characteristics of the aggregate under test, in one embodiment of the present invention, the preset image feature extraction network can be EffiDiceNet. EffiDiceNet can distinguish between color changes and true moisture characteristics, thereby reducing the influence of the aggregate color and surface water film on the moisture content of the aggregate under test and improving the accuracy of moisture content prediction.

[0064] Specifically, EffiDiceNet may include: an EfficientNetV2 backbone network and at least one DICE (Dimension-wise Interactive Convolutional Enhancement) module; wherein, the EfficientNetV2 backbone network is used to extract global semantic features (such as humidity distribution and color distribution) of the preprocessed image, and the DICE module performs fine-grained feature extraction (such as edges and reflective areas).

[0065] In its implementation, the EffiDiceNet network consists of depth, width, and resolution. Depth refers to the number of layers in the deep learning network representing the aggregate image, while width refers to the number of channels in each layer, and resolution refers to the size of the aggregate image. By balancing these three components, an efficient and accurate EffiDiceNet network can be obtained.

[0066] In embodiments of the present invention, the EfficientNetV2 backbone network can use lightweight convolutional blocks such as Moving Inverted Bottleneck Convolutional Modules (MBConv blocks) as the basic network structure. By stacking multiple MBConv blocks, a highly efficient EfficientNetV2 backbone network can be constructed. Compared with the EfficientNetV2 network structure, under the same computational cost, the EfficientNetV2 backbone network in the embodiments of the present invention has higher feature extraction capabilities, especially in complex backgrounds, and can better distinguish the background and moisture characteristics of the aggregate to be tested.

[0067] In practice, the number of MBConv blocks in the EfficientNetV2 backbone network can be set according to the network depth and task requirements. For example, when the input image is a high-resolution image, the number of MBConv blocks can be increased to extract more image details. When a lightweight model is required, the number of MBConv blocks can be reduced.

[0068] In one embodiment, the overall structure of the EfficientNetV2 backbone network is as follows: Figure 3 As shown. (Refer to...) Figure 3The EfficientNetV2 backbone network may include: a 3*3 convolutional module Conv3*3, a first MBConv block, and a second MBConv block. The first MBConv module can be a 1x expanded inverted bottleneck k*k convolutional module MBConv1 k*k, and the second MBConv module can be a 6x expanded inverted bottleneck k*k convolutional module MBConv6 k*k, where k*k is the depth of the convolutional kernel.

[0069] Specifically, the 3*3 convolutional module Conv3*3 is used for initial image feature extraction. The input image size of the 3*3 convolutional module Conv3*3 is 3*H*W, and the output image size is 24*(H / 2)*(W / 2), where H is the height of the image, W is the width of the image, and 3 and 24 refer to the number of channels in the image.

[0070] Reference Figure 4 The first MBConv module, MBConv1 k*k, can sequentially perform a k*k depth convolution operation, a scalable depthwise separable convolution operation SE, and a 1*1 convolution operation Con 1*1 on the input. (See reference...) Figure 5 The second MBConv module, MBConv6 k*k, can sequentially perform a 1*1 convolution operation Con 1*1, a depthwise k*k convolution operation, a scalable depthwise separable convolution operation SE, and a 1*1 convolution operation Con 1* on the input. Specifically, when the second MBConv module MBConv6 k*k performs a 1*1 convolution operation Con 1*1 on the input, it increases the number of channels in the input image by a factor of 6. Convolution operations are primarily used to adjust the number of channels, while depthwise separable convolution operations are used to reduce computational cost and the number of parameters.

[0071] The EfficientNetV2 backbone network in this embodiment of the invention can not only reduce computational overhead by balancing depth, width and resolution, but also maintain high accuracy in different scenarios, thereby improving the detection accuracy of the moisture content of the aggregate to be tested.

[0072] While the EfficientNetV2 backbone network excels in computational efficiency and overall feature extraction, its depthwise separable convolutional structure effectively reduces model parameters and computational complexity by separating the extraction process of channels and space. However, this structure may suffer from insufficient interaction between channel and spatial dimensional information when dealing with targets with complex shapes or irregular structures. A lack of sensitive modeling of details such as edge contours may affect the recognition performance of different water content features. Furthermore, its structure may weaken the ability to integrate spatial contextual information, thus limiting its performance in recognizing targets with occlusion or background interference.

[0073] Therefore, in the embodiments of this invention, the EffiDiceNet network further includes at least one DICE module. The core idea of ​​the DICE module is borrowed from the DiceNet network, which uses dimension-wise convolutions to construct efficient feature extraction units, enhancing cross-dimensional information interaction capabilities while maintaining lightweight design. Specifically, addressing the unique challenges of feature extraction from aggregates in complex industrial environments, this invention makes key improvements and optimizations to the specific structure of the feature extraction units in the DiceNet network, making it more suitable for moisture content detection tasks, ultimately forming the DICE module in the embodiments of this invention. This DICE module improves the structure of traditional convolutions, enabling more independent and efficient feature extraction across channel and spatial dimensions.

[0074] In specific implementation, refer to Figure 6 The DICE module may include: a local fusion submodule and a global fusion submodule. Wherein:

[0075] The local fusion submodule is adapted to perform depth-separable convolution operations on the output of the EfficientNetV2 backbone network along the channel direction, width direction and height direction respectively, and to concatenate the results of the depth-separable convolution operations in the three directions. This allows for fine-grained feature extraction of the input image in local regions, thereby enhancing the network's ability to perceive information at different levels.

[0076] The global fusion submodule is suitable for performing channel shuffling on the stitched results of depth-separable convolution operations in three directions. After channel shuffling, the results are dimensionality-reduced, and then a channel attention mechanism is used to adaptively adjust the feature weights. This allows for further integration of image features in the channel, width, and height directions after local fusion, thereby improving the model's overall perception capability, enhancing the model's understanding of the overall morphology of online natural sand, strengthening the integrity description of aggregates with different moisture contents, and reducing the impact of local noise on the detection results.

[0077] Specifically, suppose the image features Y of the input image R of the DICE module are... in Mapped to Y in ∈R C*W*H Where C is the channel of the input image R, W is the channel width of the input image R, and H is the height of the input image R. After the input image R is input to the local fusion submodule, the local fusion submodule will perform independent feature extraction operations on the input image R in the channel direction, width direction and height direction. All of them can use depthwise separable convolution to reduce computational complexity and improve parameter utilization efficiency.

[0078] Specifically, for processing along the channel direction, depthwise separable convolution is performed along the channel dimension, resulting in an image feature map Y. C ∈R C*W*H For processing the width and height directions, the input features are first dimensionality-permutated, then depthwise separable convolutions are performed, and finally the original dimensions are restored to obtain the feature maps Y. W Y H ∈R C*W*H The feature maps obtained in these three directions have the same size, respectively enhancing channel information, horizontal structure information, and vertical structure information. Concatenating these three maps yields the feature map Y. cat ∈R 3C*W*H This provides rich spatial structure information support for the subsequent global fusion stage, which will greatly improve the detection accuracy of the aggregates to be tested.

[0079] Compared to the feature extraction unit in existing Disseminated networks, the DICE module in this embodiment replaces all standard convolutions in each dimension (C, H, W) with depthwise separable convolutions during the local fusion stage, thereby significantly reducing computational complexity and the number of parameters. Depthwise separable convolution is itself an efficient convolution decomposition method; combining it with the dimensionality decomposition idea of ​​DICE forms a "dual decomposition," making the model very lightweight and easy to deploy in embedded devices on the production line. Furthermore, the pointwise convolution in the depthwise separable convolution is specifically responsible for combining channel information, enabling the model to better fuse different feature modalities from infrared sensors and images, thereby enhancing the model's nonlinear expressive power in the channel dimension.

[0080] After completing the local fusion, due to the feature mapping Y obtained in the local fusion stage... cat It is composed of three sets of features from different directions, and the channel order does not fully interact, which may lead to information redundancy. Therefore, the DICE module introduces a channel shuffling operation, which breaks the original order relationship between channels, enabling features from different sources to be effectively integrated, thereby improving information flow.

[0081] Specifically, the spliced ​​Y cat The channels are divided into three groups according to their orientation. Then, the channels are transposed. Finally, the transposed features are reassembled back into their original shape, resulting in a Y-shaped feature. s ∈R 3C*W*H By shuffling the channels, the information exchange between local features can be enhanced, allowing the features of channels, height, and width to be fully mixed, thereby improving the overall understanding of the aggregate under test.

[0082] After the channel is shuffled, feature Y s The features are still in a high-dimensional state, and directly using them for subsequent processing would significantly increase computational costs. Therefore, the DICE module uses a 1×1 pointwise convolution to reduce the dimensionality of the features, decreasing the number of channels from 3C to C, thus obtaining the dimensionality-reduced feature map Y. down ∈R C*W*H Despite the dimensionality reduction process, the importance of different channels still varies. Therefore, the DICE module further employs a channel attention mechanism to adaptively adjust feature weights to enhance the representation capability of key regions.

[0083] Specifically, Figure 7 This is a schematic diagram illustrating the principle of a channel attention mechanism in an embodiment of the present invention. (Refer to...) Figure 7 For the dimensionality-reduced feature map Y down The feature weights can be adjusted by sequentially performing three steps: global pooling, a two-layer fully connected network, and channel weighting. Specifically, the feature mapping Y... down Perform global average pooling to calculate the global response Y along the channel dimension. gap ∈R C*1*1 To establish dependencies between channels, the DICE module can use a two-layer fully connected network to connect Y channels. gap Mapping is performed, where the first fully connected layer network maps Y... gap The dimension of the vector is reduced from C×1×1 to C / r×1×1 (where r is the reduction ratio), and then activated using the ReLU function to obtain the weight vector Y. w1 The second fully connected layer upscales the dimensionality-reduced features back to C×1×1, activates them using the Sigmoid function, and outputs a weight vector Y. w2 Finally, the obtained weight vector Y w2 As a channel attention weight, it acts on Y down Up, and soon Y w2 With Y down Multiplying them together adjusts the channel weights, resulting in the output feature Y. out Relative to Y down Y out The dimensions remain the same, but the features are enhanced.

[0084] Due to the output feature Y out The weights of important feature channels were enhanced, features in irrelevant directions were suppressed, and the feature representation capability of the excavation site area was improved. Furthermore, by adopting a method of first reducing and then increasing dimensionality, the information dimension is compressed and then restored, which can extract more discriminative feature information while reducing computational complexity. This achieves the focus on key features of the excavation site area and the effective suppression of redundant information.

[0085] Compared to the feature extraction unit in existing Disseminator networks, the DICE module in this embodiment adds a lightweight channel weighting operation after dimensionality reduction during the global fusion stage. Specifically, the global fusion submodule first obtains global statistical information for each channel through global average pooling (GAP), then generates a weight vector through two fully connected layers (with dimensionality reduction and ReLU activation in the middle, and Sigmoid activation in the last layer). Each value in this vector represents the importance of the corresponding channel. Finally, this weight vector is multiplied by the original feature map, allowing the model to learn to "focus" on feature channels that are more critical to determining water content (e.g., channels corresponding to certain color or texture changes), while suppressing unimportant or noisy channels. This greatly improves the model's sensitivity to subtle moisture features and its resistance to interference.

[0086] In practical implementation, the EffiDiceNet network can use only one DICE module, or two or more DICE modules, depending on the required detection accuracy. When using two or more DICE modules, they are connected sequentially to form a feature extraction chain, thereby gradually reducing spatial size, increasing the number of channels, and improving feature representation capabilities.

[0087] In one embodiment of the present invention, reference is made to... Figure 3 The EffiDiceNet network may include three DICE modules: a first DICE module (DICE1), a second DICE module (DICE2), and a third DICE module (DICE3). The first DICE module (DICE1) is connected to the output of the EfficientNetV2 backbone network, the second DICE module (DICE2) is connected to the output of the first DICE module (DICE1), and the third DICE module (DICE3) is connected to the output of the second DICE module (DICE2). The structure of each DICE module can be as follows... Figure 6 As shown, the output of the EfficientNetV2 backbone network passes through three DICE modules in sequence, which can gradually reduce the spatial size, increase the number of channels, and improve the feature representation capability.

[0088] by Figure 3 Taking the EffiDiceNet network structure shown in the figure as an example, assuming the input image size of the EffiDiceNet network is 3×H×W, the stride of the 3*3 convolutional module Conv3*3, the first MBConv block and the second MBConv block, the first DICE module DICE1, the second DICE module DICE2 and the third DICE module DICE3 are 2, 2, 2, 2, 1 and 2 respectively. Then the changes in image size after passing through each module are shown in Table 1:

[0089] Table 1

[0090]

[0091] This shows that after passing through the EffiDiceNet network, 256-dimensional image feature data can be obtained. This 256-dimensional image feature data can then be fused with reflectance spectral feature data to achieve cross-modal information complementarity.

[0092] The aggregate moisture content detection method in this embodiment combines the DICE module with the EfficientNetV2 backbone network. The first few layers of the existing EfficientNetV2 network's MBConv modules are retained, while the later layers' MBConv modules are replaced by the DICE module. Since aggregate detection tasks typically involve high-resolution input images requiring processing of large areas, the retained MBConv module, with its lower computational complexity and higher efficiency, ensures rapid and efficient extraction of low-level features of online natural sand in the early stages. This minimizes computational overhead and improves the model's processing speed. Furthermore, it effectively extracts these basic features, enhancing the performance of the subsequent DICE module when processing complex features. The DICE module does not directly use the feature extraction units in the Disseminator network; instead, the improved DICE module replaces the later stages' MBConv modules in the EfficientNetV2 backbone network as "plug-and-play" components, thus balancing efficiency and performance.

[0093] The EfficientNet network front end retains the MBConv module, enabling rapid and efficient extraction of low-level general features (such as edges and contours). The EfficientNet network back end embeds the DICE module, which performs efficient detail capture and channel-spatial information interaction at a more abstract feature level, specifically handling subtle representations caused by changes in water content, forming a hybrid architecture. This combined network architecture integrates the advantages of two efficient networks (EfficientNetV2 and DiceNet), constructing a specialized network (EffiDiceNet) more suitable for this task, effectively improving the accuracy of water content detection.

[0094] Experimental verification shows that the solution adopted in this invention achieves an order-of-magnitude improvement in moisture content detection accuracy, reducing the online detection error of natural sand moisture content from ±2.5% to within 0.5%, fundamentally overcoming the measurement distortion caused by aggregate color differences and surface water film interference in near-infrared technology. Simultaneously, this invention, through multi-sensor fusion and the EffiDiceNet model, significantly improves robustness to multi-source interference (such as dust and light), eliminating the need for frequent manual calibration for sand sources from different origins, and shortening the calibration time for a single batch from several hours to less than ten minutes, achieving a technological leap from intermittent auxiliary measurement to real-time intelligent closed-loop control.

[0095] To enable those skilled in the art to better understand and implement the present invention, the apparatus, testing system, electronic device, and computer-readable storage medium corresponding to the above method are described in detail below.

[0096] Reference Figure 8 This invention also provides an aggregate moisture content detection device 80, which may include: an acquisition unit 81, a feature mapping unit 82, a fusion unit 83, and a prediction unit 84. Wherein:

[0097] The acquisition unit 81 is adapted to acquire image feature data and reflectance spectral intensity data of the aggregate to be tested;

[0098] The feature mapping unit 82 is adapted to perform feature mapping on the reflectance spectral intensity data to obtain reflectance spectral feature data aligned with the dimensions of the image feature data;

[0099] The fusion unit 83 is adapted to perform feature fusion between the image feature data and the feature-mapped reflectance spectrum feature data to obtain fused aggregate feature data.

[0100] The prediction unit 84 is adapted to obtain the moisture content of the aggregate to be tested based on the fused aggregate characteristic data.

[0101] The acquisition unit 81, feature mapping unit 82, fusion unit 83, and prediction unit 84 can be implemented by referring to the description of the corresponding steps in the above-mentioned method for detecting aggregate moisture content, and will not be repeated here.

[0102] This invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of any of the above methods.

[0103] In specific implementations, the computer-readable storage medium may include ROM, RAM, disk, or optical disk, etc.

[0104] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program that can run on the processor, and the processor executes the steps of any of the methods described above when running the computer program.

[0105] Regarding the modules / units included in the various devices and products described in the above embodiments, they can be software modules / units, hardware modules / units, or a combination of both. For example, for various devices and products applied to or integrated into a chip, all of their modules / units can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs that run on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits; for various devices and products applied to or integrated into a chip module, all of their modules / units can be implemented using hardware methods such as circuits, and different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using hardware methods such as circuits. The components can be implemented using software programs that run on the processor integrated within the chip module. The remaining (if any) modules / units can be implemented using hardware methods such as circuits. For various devices and products applied to or integrated into the terminal, each of its components / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or in different components within the terminal. Alternatively, at least some modules / units can be implemented using software programs that run on the processor integrated within the terminal, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits.

[0106] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for detecting aggregate moisture content, characterized in that, include: Acquire image feature data and reflectance spectral intensity data of the aggregate to be tested; The reflectance spectral intensity data is feature-mapped to obtain reflectance spectral feature data aligned with the dimensions of the image feature data; The image feature data and the feature-mapped reflectance spectral feature data are fused to obtain the fused aggregate feature data. Based on the fused aggregate characteristic data, the moisture content of the aggregate to be tested is obtained; Acquiring image feature data of the aggregate to be tested includes: acquiring an image of the aggregate to be tested; preprocessing the acquired image; and inputting the preprocessed image into a preset image feature extraction network to obtain the image feature data of the aggregate to be tested. The preset image feature extraction network includes: an EfficientNetV2 backbone network and at least one DICE module; wherein, the EfficientNetV2 backbone network is used to extract global semantic features of the preprocessed image, and the DICE module performs fine-grained feature extraction; The DICE module includes: The local fusion submodule is adapted to perform depth-separable convolution operations on the output of the EfficientNetV2 backbone network along the channel direction, width direction and height direction respectively, and to concatenate the results of the depth-separable convolution operations in the three directions. The global fusion submodule is suitable for performing channel shuffling on the concatenation results of depth-separable convolution operations in three directions, performing dimensionality reduction on the results after channel shuffling, and then using a channel attention mechanism to adaptively adjust the channel feature weights.

2. The method for detecting aggregate moisture content as described in claim 1, characterized in that, The EfficientNetV2 backbone network includes: a 3*3 convolutional module, a first MBConv module, and a second MBConv module. The first MBConv module is a 1x expanded inverted bottleneck k*k convolutional module, and the second MBConv module is a 6x expanded inverted bottleneck k*k convolutional module, where k*k is the depth of the convolutional kernel.

3. The method for detecting aggregate moisture content as described in claim 1, characterized in that, The step of fusing the image feature data and reflectance spectral feature data to obtain fused aggregate feature data includes: The image feature data and reflectance spectrum feature data are stitched together to obtain the stitched feature data; The spliced ​​feature data is input into a fused fully connected network to obtain fused aggregate feature data.

4. The method for detecting aggregate moisture content as described in claim 1, characterized in that, The aggregate is natural sand.

5. A device for detecting aggregate moisture content, characterized in that, include: The acquisition unit is suitable for acquiring image feature data and reflectance spectral intensity data of the aggregate to be tested. The feature mapping unit is adapted to perform feature mapping on the reflectance spectral intensity data to obtain reflectance spectral feature data aligned with the dimensions of the image feature data; The fusion unit is adapted to fuse the image feature data with the feature-mapped reflectance spectral feature data to obtain fused aggregate feature data. The prediction unit is adapted to obtain the moisture content of the aggregate to be tested based on the fused aggregate feature data; The acquisition unit is adapted to acquire images of the aggregate to be tested and to preprocess the acquired images. The preprocessed image is input into a preset image feature extraction network to obtain the image feature data of the aggregate to be tested; The preset image feature extraction network includes: an EfficientNetV2 backbone network and at least one DICE module; wherein, the EfficientNetV2 backbone network is used to extract global semantic features of the preprocessed image, and the DICE module performs fine-grained feature extraction; The DICE module includes: The local fusion submodule is adapted to perform depth-separable convolution operations on the output of the EfficientNetV2 backbone network along the channel direction, width direction and height direction respectively, and to concatenate the results of the depth-separable convolution operations in the three directions. The global fusion submodule is suitable for performing channel shuffling on the concatenation results of depth-separable convolution operations in three directions, performing dimensionality reduction on the results after channel shuffling, and then using a channel attention mechanism to adaptively adjust the channel feature weights.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement the steps of the method according to any one of claims 1 to 4.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, characterized in that, When the processor runs the computer program, it performs the steps of the method according to any one of claims 1 to 4.

Images