Method and device for constructing leaf detection model based on hyperspectral imaging technology

Abstract

The embodiment of the invention provides a method for constructing a leaf detection model based on a hyperspectral imaging technology. The method comprises the following steps: acquiring a hyperspectral image of a sample leaf; preprocessing the hyperspectral image of the sample leaf to obtain a target hyperspectral image; a competitive self-adaptive reweighted sampling method is adopted to screen band images of the target hyperspectral image, and a plurality of target band images are obtained; and detecting the plurality of target band images by using a pre-constructed deep learning model. In the scheme, a pre-constructed deep learning model is adopted to cope with the complexity of hyperspectral image data, the model overcomes the tedious process of artificial feature extraction in a traditional method through end-to-end automatic feature extraction, space and spectral information is effectively fused, the detection precision of the water content and chlorophyll content of the leaves is remarkably improved, and the detection accuracy of the leaf moisture content and chlorophyll content is improved. And meanwhile, the method has relatively high generalization ability and convergence speed, and is suitable for detecting large-scale seedling leaves.

Description

Technical Field

[0001] This invention relates to the field of leaf inspection, and in particular to a method and apparatus for constructing a leaf inspection model based on hyperspectral imaging technology. Background Technology

[0002] Seedlings are young plants cultivated from seeds or other propagation materials during tree reproduction, possessing a certain growth potential and vitality. The quality of seedlings directly affects the later growth rate, disease and pest resistance, physiological health, stress resistance, and final yield of the plant.

[0003] Currently, seedling testing typically employs chemical analysis methods. For example, the moisture content of leaves in *Betula sclerophylla* seedlings is determined using the drying and weighing method, which involves drying the sample to a constant weight and calculating the water loss to determine the moisture content. Chlorophyll is measured using spectrophotometry, which measures the absorbance of chlorophyll extract at a fixed wavelength and then calculates the chlorophyll content using a formula.

[0004] Traditional chemical methods require a high level of operator skill, and improper operation can lead to biased results. Furthermore, the testing costs are high, especially when multiple repeated tests are needed, which significantly increases the cost and limits their application in large-scale seedling monitoring. Therefore, there is a need to develop an efficient, non-destructive method suitable for large-scale seedling leaf testing. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for constructing a leaf detection model based on hyperspectral imaging technology, thereby addressing the problem that existing detection methods are costly and unsuitable for large-scale seedling monitoring. The specific technical solution is as follows:

[0006] In a first aspect of this invention, a method for constructing a leaf detection model based on hyperspectral imaging technology is provided, comprising:

[0007] Acquire hyperspectral images of the sample blade;

[0008] The hyperspectral image of the sample blade is preprocessed to obtain the target hyperspectral image; wherein the preprocessing includes at least one of masking, redundancy clipping, and noise reduction.

[0009] A competitive adaptive reweighted sampling method is used to filter the band images of the target hyperspectral image, resulting in multiple target band images;

[0010] Multiple target band images are detected using a pre-built deep learning model, which is used to extract spectral features and fuse spatial features for each target band image.

[0011] Optionally, the detection of multiple target band images using a pre-built deep learning model includes:

[0012] The spectral data of the multiple target band images are extracted using a 3D residual module and the spatial data are fused. 3D max pooling is used to reduce the dimensionality of the extracted data while retaining key features, including leaf water content and chlorophyll content.

[0013] A channel attention mechanism is used to highlight the key features through adaptive weights, thereby obtaining the first target key features;

[0014] The spatial features of the multiple target band images are extracted using a 2D residual module, and the dimensionality of the extracted spatial data is further reduced by 2D max pooling to obtain the second target key features, which include blade surface details and texture information.

[0015] By integrating the key features of the first target and the key features of the second target, the state characterization of the sample blade is obtained.

[0016] Optionally, the channel attention mechanism includes: a squeezing phase and an excitation phase;

[0017] The extrusion process includes:

[0018]

[0019] Among them, Z C This represents the global descriptor for the c-th channel, x. i,j,k This represents the value of the c-th channel at position (i,j) of the input feature map, where H and W represent the height and width of the feature map, respectively.

[0020] The incentive phase includes:

[0021]

[0022] Where σ represents the sigmoid activation function, and g represents a fully connected layer containing non-linear activation, used to transform the global descriptor and generate channel weight coefficients Sc;

[0023] The channel attention mechanism, which highlights the key features through adaptive weights, includes:

[0024]

[0025] in, This represents the value of the c-th channel at position (i,j) in the weighted feature map.

[0026] Optionally, each residual module includes a skip connection and a main path; the skip connection is used to adjust the number of channels of the input feature map to match the number of channels of the main path output.

[0027] The main path consists of three convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer uses a 1x1 convolution to adjust the number of channels in the input feature map. The second layer uses a 3x3 convolution kernel to capture local features of the input data. The third layer uses a 1x1 convolution to adjust the number of channels and integrate the features extracted by the previous layer.

[0028] Optionally, the preprocessing of the hyperspectral image of the sample blade includes:

[0029] The sample blade is masked to obtain a mask image;

[0030] Extract the minimum bounding rectangle from the mask image, and crop the area outside the minimum bounding rectangle to obtain the mask border image;

[0031] The mask border image is scaled to a preset pixel size using interpolation.

[0032] A smoothing algorithm is used to denoise the scaled mask border image.

[0033] In a second aspect of the invention, a blade detection device is also provided, comprising:

[0034] The image acquisition module is used to acquire hyperspectral images of the sample blades.

[0035] The preprocessing module is used to preprocess the hyperspectral image of the sample blade to obtain the target hyperspectral image; wherein the preprocessing includes at least one of masking, redundancy clipping, and noise reduction.

[0036] The image filtering module is used to filter the band images of the target hyperspectral image using a competitive adaptive reweighted sampling method to obtain multiple target band images;

[0037] The image detection module is used to detect multiple target band images using a pre-built deep learning model, wherein the deep learning model is used to extract spectral features and fuse spatial features for each target band image.

[0038] Optionally, the image detection module is specifically used for:

[0039] The spectral data of the multiple target band images are extracted using a 3D residual module and the spatial data are fused. 3D max pooling is used to reduce the dimensionality of the extracted data while retaining key features, including leaf water content and chlorophyll content.

[0040] A channel attention mechanism is used to highlight the key features through adaptive weights, thereby obtaining the first target key features;

[0041] The spatial features of the multiple target band images are extracted using a 2D residual module, and the dimensionality of the extracted spatial data is further reduced by 2D max pooling to obtain the second target key features, which include blade surface details and texture information.

[0042] By integrating the key features of the first target and the key features of the second target, the state characterization of the sample blade is obtained.

[0043] Optionally, the channel attention mechanism includes: a squeezing phase and an excitation phase;

[0044] The extrusion process includes:

[0045]

[0046] Among them, Z C This represents the global descriptor for the c-th channel, x. i,j,k This represents the value of the c-th channel at position (i,j) of the input feature map, where H and W represent the height and width of the feature map, respectively.

[0047] The incentive phase includes:

[0048]

[0049] Where σ represents the sigmoid activation function, and g represents a fully connected layer containing non-linear activation, used to transform the global descriptor and generate channel weight coefficients Sc;

[0050] The channel attention mechanism, which highlights the key features through adaptive weights, includes:

[0051]

[0052] in, This represents the value of the c-th channel at position (i,j) in the weighted feature map.

[0053] Optionally, each residual module includes a skip connection and a main path; the skip connection is used to adjust the number of channels of the input feature map to match the number of channels of the main path output.

[0054] The main path consists of three convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer uses a 1x1 convolution to adjust the number of channels in the input feature map. The second layer uses a 3x3 convolution kernel to capture local features of the input data. The third layer uses a 1x1 convolution to adjust the number of channels and integrate the features extracted by the previous layer.

[0055] In a third aspect of the present invention, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0056] Memory, used to store computer programs;

[0057] A processor, when executing a program stored in memory, implements any of the steps described above.

[0058] The beneficial effects of this invention are:

[0059] This invention provides a method for constructing a leaf detection model based on hyperspectral imaging technology. The method involves acquiring hyperspectral images of sample leaves; preprocessing the hyperspectral images of the sample leaves to obtain target hyperspectral images; using a competitive adaptive reweighted sampling method to filter the band images of the target hyperspectral images to obtain multiple target band images; and using a pre-built deep learning model to detect the multiple target band images. In this scheme, to address the complexity of hyperspectral image data, a pre-built deep learning model is employed. This model overcomes the cumbersome process of manual feature extraction in traditional methods through end-to-end automatic feature extraction and effectively integrates spatial and spectral information, significantly improving the detection accuracy of leaf water content and chlorophyll content. It also possesses strong generalization ability and convergence speed, making it suitable for large-scale seedling leaf detection.

[0060] This invention proposes a multi-dimensional optimized technical solution based on traditional hyperspectral leaf detection, which has the following main advantages:

[0061] 1. Automatic selection mechanism for key bands based on CARS

[0062] This invention introduces a competitive adaptive reweighted sampling (CARS) method to select feature bands in hyperspectral images, effectively removing redundant bands, highlighting target features, significantly reducing the model input dimension and improving detection efficiency. It is superior to conventional algorithms such as SPA and has stronger feature extraction targeting and modeling performance.

[0063] 2. A multidimensional feature extraction model integrating 3D-2D residual network structures

[0064] We construct a deep neural network structure (3D-2D-ResNet-SE) that integrates 3D and 2D residual modules, taking into account both the spectral and spatial dimensions of hyperspectral images. It extracts spectral features at the band level and texture information at the image level, respectively, thus avoiding the problem that traditional methods cannot simultaneously take into account both spatial and spectral dimensions.

[0065] 3. Introduce SE channel attention mechanism to enhance the model's discriminative ability.

[0066] By utilizing the Squeeze-and-Excitation (SE) attention mechanism, adaptive weighting is applied to spectral and spatial feature channels, guiding the model to focus on channels that are more sensitive to water content and chlorophyll content, effectively improving the model's detection accuracy and generalization ability for physiological indicators.

[0067] The method of this invention not only significantly improves the detection accuracy of leaf water content and chlorophyll content, but also has the advantages of strong ability to process high-dimensional data, fast model convergence speed, and high deployment flexibility. It is suitable for non-destructive detection of the health status of large-scale seedlings and has high practical value and industrial promotion prospects. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0069] Figure 1 This is a flowchart of the method for constructing a leaf detection model based on hyperspectral imaging technology according to the present invention;

[0070] Figure 2 This is the leaf spectral image extraction process of the present invention;

[0071] Figure 3 This is a structural diagram of the model of the present invention;

[0072] Figure 4 This is a schematic diagram of the SE channel attention mechanism of the present invention;

[0073] Figure 5 This is a residual structure diagram of the present invention;

[0074] Figure 6 This describes how the predictive performance of the model in this invention varies with the input size.

[0075] Figure 7 The training loss varies for different models in this invention;

[0076] Figure 8 This is a visualization of the water content characteristics under different drought treatments according to the present invention;

[0077] Figure 9This is a visualization of chlorophyll content characteristics under different drought treatments according to the present invention.

[0078] Figure 10 A schematic diagram of the gantry mechanism structure in an embodiment of the present invention;

[0079] Figure 11 for Figure 10 A magnified view of a section at point A in the middle;

[0080] Figure 12 for Figure 10 A magnified view of a section at point B in the middle;

[0081] Figure 13 This is a schematic diagram of the structure of the two first crossbars in this invention;

[0082] Figure 14 for Figure 13 A magnified view of a section at point C;

[0083] Figure 15 This is a schematic diagram of the structure of the second crossbar in this invention;

[0084] Figure 16 for Figure 15 A magnified view of a section at point D;

[0085] Figure 17 for Figure 15 A structural diagram from another angle;

[0086] Figure 18 for Figure 17 A magnified view of a section at point E in the middle;

[0087] Figure 19 This is a schematic diagram of the structure of the vertical rod and the second movable plate in this invention;

[0088] Figure 20 This is a schematic diagram of the structure of the sleeve plate and the vertical rod in this invention;

[0089] Figure 21 for Figure 20 A structural diagram from another angle;

[0090] Figure 22 This is a schematic diagram of the structure of the second moving plate and the second slider in this invention;

[0091] Figure 23 for Figure 22 A magnified view of a section at point F. Detailed Implementation

[0092] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] To address the problems in the existing technology, this invention provides a method for constructing a leaf detection model based on hyperspectral imaging technology, which solves the problem that the existing detection methods have high detection costs and are not suitable for large-scale seedling monitoring.

[0094] It should be noted that the method for constructing a leaf detection model based on hyperspectral imaging technology provided in this embodiment of the invention can be applied to electronic devices. In practical applications, the electronic device can be a smartphone, tablet computer, laptop computer, etc., which is all reasonable.

[0095] The following section first introduces the method for constructing a leaf detection model based on hyperspectral imaging technology provided in the embodiments of the present invention.

[0096] like Figure 1 As shown in the embodiments of the present invention, the method for constructing a leaf detection model based on hyperspectral imaging technology may include the following steps:

[0097] S101, acquire hyperspectral images of the sample blade.

[0098] For example, the sample leaves in this embodiment can be obtained by sampling. During the sampling process, considering that there may be physiological differences between different plant layers, in order to fully reflect the physiological response of the seedling leaves, three leaves can be randomly selected from the upper, middle and lower layers of each selected plant, for a total of nine leaves, and the average spectral data and hyperspectral image data of a total of 1199 leaves can be obtained as the analysis sample of a single plant.

[0099] The procedure for determining leaf physiological parameters is as follows: First, a portable chlorophyll fluorometer is used to measure the chlorophyll content on live leaves to ensure the authenticity and accuracy of the data. Next, leaves are quickly harvested, and their spectral characteristics are immediately measured on-site using a miniature near-infrared spectrometer to minimize the impact of post-harvest water evaporation on the results. Given the complexity of hyperspectral image acquisition equipment and the time-consuming data acquisition process, samples are collected in groups of three leaves. This improves data acquisition efficiency and enhances data stability and representativeness. Finally, each leaf sample is numbered and accurately measured for its water content.

[0100] The hyperspectral image of the sample blade in this embodiment can be obtained by any spectroscopic instrument in the prior art, such as a near-infrared hyperspectral imaging camera or a portable near-infrared spectrometer.

[0101] S102, preprocess the hyperspectral image of the sample blade to obtain the target hyperspectral image; wherein the preprocessing includes at least one of masking, redundancy clipping, and noise reduction.

[0102] For example, in one implementation, preprocessing the hyperspectral image of the sample blade may include:

[0103] The sample blade is masked to obtain a mask image;

[0104] Extract the minimum bounding rectangle from the mask image, and crop the area outside the minimum bounding rectangle to obtain the mask border image;

[0105] The mask border image is scaled to a preset pixel size using interpolation.

[0106] A smoothing algorithm is used to denoise the scaled mask border image.

[0107] Since the hyperspectral images of the sample blades were acquired in groups of three blades, preprocessing can be performed to obtain individual blade hyperspectral images and facilitate input for subsequent deep learning models. For example: Figure 2 As shown, by using a masking operation, only the spectral image of a single leaf is retained, while the spectral values ​​of the background area and the other two leaves are set to zero. This eliminates interference signals from non-target areas such as soil and impurities, resulting in a clear hyperspectral image of a single leaf, i.e., the mask image (mask picture). Next, the minimum bounding rectangle of the leaf (mask border image) is extracted based on the mask image to ensure that the cropped image completely covers the leaf area, avoiding redundant background. Considering the differences in leaf size and the input requirements of the deep learning model, interpolation is used to uniformly scale the cropped hyperspectral image to 256×256 pixels, retaining 166 spectral bands (removing the first and last 5 noise bands from the original 176 bands). Subsequently, Savitzky-Golay (SG) smoothing is applied to the image to further remove noise and optimize image quality.

[0108] S103, a competitive adaptive reweighted sampling method is used to filter the band images of the target hyperspectral image to obtain multiple target band images.

[0109] In this embodiment, by screening the target hyperspectral image, the bands most relevant to leaf water content and chlorophyll content can be selected, redundant information can be removed, and data dimensionality can be reduced, thereby reducing computational complexity and improving prediction accuracy and generalization ability. For example, in predicting the water content of *Betula sclerophylla* leaves, the Competitive Adaptive Reweighted Sampling (CARS) method screened 39 characteristic wavelengths, particularly in the near-infrared region of 700–1000 nm. These wavelengths are closely related to the pigment absorption characteristics and cell structure of the leaves, and may reflect the high absorption characteristics of water in the near-infrared band. In predicting the chlorophyll content of *Betula sclerophylla* leaves, the CARS method also selected 28 characteristic wavelengths, covering the range from 400 nm to nearly 1000 nm. These wavelengths may be related to the biochemical composition and structural information of the leaves.

[0110] S104, using a pre-built deep learning model to detect multiple target band images, wherein the deep learning model is used to extract spectral features and fuse spatial features for each target band image.

[0111] For example, in one implementation, detecting multiple target band images using a pre-built deep learning model may include:

[0112] The spectral data of the multiple target band images are extracted using a 3D residual module and the spatial data are fused. 3D max pooling is used to reduce the dimensionality of the extracted data while retaining key features, including leaf water content and chlorophyll content.

[0113] A channel attention mechanism is used to highlight the key features through adaptive weights, thereby obtaining the first target key features;

[0114] The spatial features of the multiple target band images are extracted using a 2D residual module, and the dimensionality of the extracted spatial data is further reduced by 2D max pooling to obtain the second target key features, which include blade surface details and texture information.

[0115] By integrating the key features of the first target and the key features of the second target, the state characterization of the sample blade is obtained.

[0116] For example, such as Figure 3 As shown, the deep learning model in this embodiment is a 3D-2DResNet-SE model, which is a 3D residual module-2D residual module-residual connection structure-channel attention mechanism (Squeeze-and-Excitation, SE) model.

[0117] This model extracts deep features of different spectral bands of leaves through three 3D residual modules. The 3D convolutional kernel can simultaneously capture spectral information and fuse spatial information. Then, a (2×2×2) 3D max pooling is used to reduce the data dimensionality and retain key features. Next, an SE channel attention module is introduced to highlight key features related to leaf water content and chlorophyll content through adaptive weights. Then, two 2D residual modules are used to further optimize the extraction of spatial features and deeply explore the details and texture information of the leaf surface. After each module, a (2×2) 2D max pooling is applied to further reduce the data dimensionality. Finally, through the multi-dimensional fusion of spatial and spectral features, a comprehensive and accurate representation of the leaf state is constructed.

[0118] To effectively enhance key spectral features relevant to the predicted target and suppress irrelevant information, an SE attention mechanism can be introduced. For example... Figure 4 As shown, the SE attention mechanism adaptively adjusts the weights of each channel, enabling the model to pay more attention to key features related to leaf water content and chlorophyll content detection, thereby improving detection accuracy.

[0119] The SE attention mechanism employs a three-step process of "Squeeze-Excitation-Scale," adaptively adjusting the importance of channel features to allow the model to intelligently focus on relevant information. It is commonly used to improve accuracy in tasks such as classification and detection, and is a typical implementation of channel attention mechanisms. First, in the squeezing phase... sq (·), The SE attention mechanism uses global average pooling to compress the spatial dimension of each channel of the input feature map, generating a global descriptor for each channel. For example, global average pooling is performed on the feature map U (shape C×H×W, where C is the number of channels and H / W is the spatial dimension) output by convolution, compressing the spatial information of each channel into a single numerical value, and outputting a 1×1×C vector, which can be represented numerically as:

[0120]

[0121] Among them, Z C This represents the global descriptor for the c-th channel, x. i,j,c This represents the value of the c-th channel at position (i,j) of the input feature map, where H and W represent the height and width of the feature map, respectively.

[0122] Next, in the incentive phase F ex (·, W), the SE attention mechanism performs a non-linear transformation on these global descriptors through two fully connected layers to generate weight coefficients for each channel, including:

[0123]

[0124] Where σ represents the sigmoid activation function, and g represents a fully connected layer with non-linear activation, used to transform the global descriptor and generate channel weight coefficients Sc.

[0125] Finally, in the scaling phase F Scale (·,·), the SE attention mechanism applies these weight coefficients to the original feature map, weighting the features of each channel to enhance key features and suppress irrelevant information, including:

[0126]

[0127] in, This represents the value of the c-th channel at position (i,j) in the weighted feature map.

[0128] Residual Connection Structure: Hyperspectral images are high-dimensional. As the network deepens, in addition to increased computational resource consumption and model overfitting, gradient vanishing and gradient exploding problems also occur. In this embodiment, each residual module includes a skip connection and a main path. Skip connections directly pass the input x1 to the output. In this way, the network can learn a "direct mapping" of the input features, reducing the gradient vanishing problem during training.

[0129] For example, the residual module adopts the following structure: Figure 5 As shown, in one implementation, each residual module includes a skip connection and a main path; the skip connection is used to adjust the number of channels of the input feature map to match the number of channels of the main path output.

[0130] The main path consists of three 2D / 3D convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer uses a 1x1 convolution to adjust the number of channels in the input feature map. The second layer uses a 3x3 convolution kernel to capture local features of the input data. The third layer uses a 1x1 convolution to adjust the number of channels and integrate the features extracted by the previous layer.

[0131] By stacking multiple such residual modules, the network can progressively learn more complex and abstract feature representations, and avoid the problems of gradient vanishing and exploding, thus improving the training stability and efficiency of the model.

[0132] In this invention, hyperspectral images of sample leaves are acquired; these images are preprocessed to obtain target hyperspectral images; a competitive adaptive reweighted sampling method is used to filter the band images of the target hyperspectral images, resulting in multiple target band images; and a pre-built deep learning model is used to detect these multiple target band images. To address the complexity of hyperspectral image data, this solution employs a pre-built deep learning model. This model overcomes the cumbersome manual feature extraction process of traditional methods through end-to-end automatic feature extraction and effectively integrates spatial and spectral information, significantly improving the detection accuracy of leaf water content and chlorophyll content. It also exhibits strong generalization ability and convergence speed, making it suitable for large-scale detection of seedling leaves.

[0133] For example, to further illustrate the scheme of this application, this application evaluated the impact of different band selection strategies on the modeling performance of hyperspectral images. This application extracts corresponding bands from the original hyperspectral image to construct sub-images for key bands selected by two feature wavelength selection algorithms, CARS and SPA (Successive Projection Algorithm), and uses these sub-images as input to train the 3D-2D-ResNet-SE model. The experimental results in Table 1 show that the band images selected by the CARS method exhibit the best modeling performance in the prediction models of leaf water content and chlorophyll content. In the prediction of leaf water content, the model constructed from the band images selected by the CARS method achieved the highest coefficient of determination (R²). 2 =0.8639), an improvement of 0.0261 compared to the full-band model, and superior to the band images selected by the SPA method, an improvement of 0.0104. A similar trend was also observed in the prediction of chlorophyll content, with the band images selected by CARS also showing the best performance, R 2 The value was 0.8496, an improvement of 0.014 compared to the full-band model, and an improvement of 0.0067 compared to the SPA selection method. The advantage of the CARS method lies in its ability to accurately select bands strongly correlated with the target variable (water content or chlorophyll content), thereby reducing the impact of redundant information on the model. Through a competitive weighting strategy, CARS can automatically identify and select the bands that contribute most to the prediction task, effectively improving prediction accuracy. Compared to SPA, the CARS method shows a more significant improvement in model accuracy and can reduce the number of bands by approximately 75%, thus reducing the computational burden of the model and improving processing speed. This optimization has important practical significance for applications requiring real-time or large-scale data processing. In summary, the CARS method not only improves the model's predictive performance but also increases computational efficiency, possessing greater operability and application potential, especially excelling in fields such as big data analysis and real-time monitoring.

[0134] Table 1 Performance of 3D-2D-ResNet-SE Model

[0135]

[0136] For example, this application further investigated the impact of hyperspectral image input size on model prediction performance and training efficiency, with specific results shown in Table 2 and... Figure 6 As shown, four different image resolutions (64×64, 128×128, 192×192, and 256×256 pixels) were used as input to evaluate the impact of image size variation on model performance. The results show that as image resolution increases, the model's performance in predicting leaf water content (LWC) and chlorophyll content (LCC) generally improves. In LWC prediction, the test set R... 2 Improved from 0.7419 at 64×64 resolution to 0.8639 at 256×256; R in LCC prediction 2 The value increased from 0.7386 to 0.8496. This indicates that higher resolution images provide richer spectral and spatial information, helping the model extract key features related to leaf physiological state, thereby improving prediction accuracy. At the same time, the increase in image size also significantly increases the computational cost of model training. Taking LWC prediction as an example, the training time per round increased from 4.8 seconds with a 64×64 input to 24.7 seconds with a 256×256 input; in the LCC prediction task, this time increased from 4.5 seconds to 23.8 seconds. Therefore, while increasing the input size helps optimize model performance, it also incurs a corresponding computational cost.

[0137] Table 2. Model prediction performance under different input sizes

[0138]

[0139] also, Figure 6 It also shows that improvements in model accuracy exhibit diminishing marginal returns. For example, LWC's R... 2 The increase is 0.0271 for the 128×128 to 192×192 resolution upgrade, while the upgrade from 192×192 to 256×256 is only 0.0174; LCC's R 2 The increases were 0.0109 and 0.0276 respectively. This trend indicates that the performance gain from further increasing image resolution has gradually diminished. In practical applications, the choice of input image resolution should be a trade-off between prediction accuracy and computational resources. While high-resolution images can enhance model performance, in large-scale data processing or resource-constrained environments, moderately reducing the resolution can help improve overall processing efficiency without significantly compromising accuracy, thus achieving a more cost-effective modeling effect.

[0140] For example, to comprehensively evaluate the predictive performance of the model in this application, four feature extraction methods—2DResnet, 3DResnet, Goognet, and Densenet—were selected for comparative analysis. The loss variation of different deep learning models in predicting the leaf water content and chlorophyll content of *Betula platyphylla* seedlings is shown. All models used band data selected by the CARS method, with 39 bands used to predict chlorophyll content and 28 bands used to predict water content. To ensure consistency in evaluation conditions, the input image size for all models was uniformly set to 256×256 pixels.

[0141] Figure 7 The training loss changes of various comparative models are shown. For water content prediction, the loss of all models decreased rapidly in the early stages of training, demonstrating the models' rapid learning of data features. As training progressed, the loss gradually stabilized, indicating that the models gradually converged. Specifically, 2DResnet and 3DResnet showed faster loss reduction in the early stages, but fluctuated significantly throughout the training process, especially in the mid-to-late stages. Densenet and Googlenet showed more stable loss reduction and demonstrated better stability in the mid-training stage. In contrast, the 3D-2DResNet-SE model showed the fastest convergence speed, with the loss decreasing rapidly and stabilizing within the first 20 epochs, demonstrating its superior performance in handling hyperspectral data. For chlorophyll content prediction, the models performed similarly, with the loss also decreasing rapidly in the early stages of training. 2DResnet and 3DResnet showed faster loss reduction in the early stages of training, but exhibited some fluctuation in the mid-to-late stages. Densenet and Googlenet showed more stable loss reduction, demonstrating better stability. The 3D-2DResNet-SE model also exhibited the fastest convergence speed, with the loss decreasing rapidly and stabilizing within the first 20 rounds, further validating the effectiveness of the model in processing complex spectral data.

[0142] Table 3 Comparison of prediction performance of different models

[0143]

[0144] Table 3 shows a performance comparison of different deep learning models in predicting the water content and chlorophyll content of *Betula sclerophylla* leaves. The results show that the 3D-2DResNet-SE model performs best in both prediction tasks. For predicting leaf water content, this model achieves a high R-value on the test set. 2 The R² score reached 0.8639, and the MAE was 4.1980, both outperforming other models and demonstrating higher prediction accuracy and stronger stability. In contrast, while the 2DResNet model performed well on the training set, it showed significant overfitting on the test set, with its R² score being [missing data].2 The R² value is 0.8332, and the MAE is 4.6288, indicating weak generalization ability. Although the 3DResNet model introduces three-dimensional convolution to enhance the processing of spectral information, its dependence on spatial features may not fully capture the intrinsic correlations between different bands, resulting in poor prediction performance on the test set. 2 The R² value was 0.8034, and the MAE was 5.1198. The DenseNet model improved the efficiency of feature information transmission and reuse through a dense connection mechanism, but it failed to effectively suppress noise when processing hyperspectral images, resulting in a large prediction error and an MAE of 4.8097, especially performing worst in water content prediction. The GoogLeNet model uses a multi-scale convolutional structure to capture image features at different scales; however, its ability to fuse spectral band-related information is limited, leading to slightly inferior performance in chlorophyll content prediction. 2 The accuracy is 0.8225, and the MAE is 2.4881. In summary, the 3D-2DResNet-SE model combines the advantages of 3D and 2D convolutions and adaptively adjusts feature weights through the SE module, effectively extracting and fusing spatial and spectral information. This overcomes the limitations of other models and demonstrates the best predictive performance. Overall, the 3D-2DResNet-SE model, with its optimized structure and feature extraction capabilities, exhibits significant advantages in both prediction accuracy and generalization ability.

[0145] To visually demonstrate the recognition capabilities of the constructed 3D-2D-ResNet-SE model, this application employs gradient-weighted class activation mapping (GRM) for model visualization. Specifically, the Grad-CAM++ method is used. This method optimizes weight calculation to generate high-resolution heatmaps, more clearly revealing the feature regions the model focuses on during prediction. Compared to traditional Grad-CAM, Grad-CAM++ provides more accurate visualization when processing images with complex structures or multiple target instances.

[0146] Figure 8This study provides visualizations of the moisture content characteristics of *Betula sclerophylla* leaves after drought treatments lasting different durations. The left side of the image shows the leaves in their natural colors, while the right side shows the corresponding moisture content heatmaps, with colors transitioning from blue (representing low concern or low moisture content) to red (representing high concern or high moisture content). The image sequence from day 1 to day 11 shows significant changes in leaf moisture status as the drought treatment duration increases. In the early stages (days 1-3), the heatmaps show higher concern in the central leaf region, likely due to abundant veins and active water transport in this area. As time progresses (days 5-7), signs of leaf drying begin to appear, with the red areas in the heatmaps expanding, indicating an intensifying impact of drought on the central leaf region. By days 9 and 11, moisture loss is even more severe, and the uneven distribution of red patches in the heatmaps reveals that the drought's effects are spreading across the entire leaf, particularly the central and peripheral areas. The expansion and deepening of these red areas indicate a significant decrease in moisture content in these parts.

[0147] Figure 9 A series of characteristic images of chlorophyll content in *Betula sclerophylla* leaves under different drought treatment durations are presented. The image sequence from day one to day eleven shows significant changes in chlorophyll content and distribution as the drought treatment duration increases. In the early stages of drought (days one and three), the leaves show a relatively uniform red area, indicating a relatively even distribution of chlorophyll and that photosynthesis was not significantly affected. However, as the drought continued (days five and seven), more blue areas began to appear at the leaf edges, suggesting a decrease in chlorophyll content and a possible decline in photosynthetic activity in these areas. By days nine and eleven, although drought stress continued, the red area still covered a large area of ​​the leaves, suggesting that the leaves may possess a certain degree of drought resistance, able to maintain chlorophyll content and photosynthetic activity to some extent. This may indicate that certain physiological mechanisms in the leaves help mitigate the effects of drought and protect the photosynthetic system.

[0148] Compared to the embodiments of the above methods, embodiments of the present invention provide a blade detection device, which may include:

[0149] The image acquisition module is used to acquire hyperspectral images of the sample blades.

[0150] The preprocessing module is used to preprocess the hyperspectral image of the sample blade to obtain the target hyperspectral image; wherein the preprocessing includes at least one of masking, redundancy clipping, and noise reduction.

[0151] The image filtering module is used to filter the band images of the target hyperspectral image using a competitive adaptive reweighted sampling method to obtain multiple target band images;

[0152] The image detection module is used to detect multiple target band images using a pre-built deep learning model, wherein the deep learning model is used to extract spectral features and fuse spatial features for each target band image.

[0153] This invention acquires hyperspectral images of sample leaves; preprocesses the hyperspectral images of the sample leaves to obtain target hyperspectral images; uses a competitive adaptive reweighted sampling method to filter the band images of the target hyperspectral images to obtain multiple target band images; and uses a pre-built deep learning model to detect the multiple target band images. In this scheme, to address the complexity of hyperspectral image data, a pre-built deep learning model is used. This model overcomes the cumbersome process of manual feature extraction in traditional methods through end-to-end automatic feature extraction, and effectively integrates spatial and spectral information, significantly improving the detection accuracy of leaf water content and chlorophyll content. It also has strong generalization ability and convergence speed, making it suitable for large-scale detection of seedling leaves.

[0154] Optionally, the image detection module is specifically used for:

[0155] The spectral data of the multiple target band images are extracted using a 3D residual module and the spatial data are fused. 3D max pooling is used to reduce the dimensionality of the extracted data while retaining key features, including leaf water content and chlorophyll content.

[0156] A channel attention mechanism is used to highlight the key features through adaptive weights, thereby obtaining the first target key features;

[0157] The spatial features of the multiple target band images are extracted using a 2D residual module, and the dimensionality of the extracted spatial data is further reduced by 2D max pooling to obtain the second target key features, which include blade surface details and texture information.

[0158] By integrating the key features of the first target and the key features of the second target, the state characterization of the sample blade is obtained.

[0159] Optionally, the channel attention mechanism includes: a squeezing phase and an excitation phase;

[0160] The extrusion process includes:

[0161]

[0162] Among them, Z C This represents the global descriptor for the c-th channel, x.i,j,k This represents the value of the c-th channel at position (i,j) of the input feature map, where H and W represent the height and width of the feature map, respectively.

[0163] The incentive phase includes:

[0164]

[0165] Where σ represents the sigmoid activation function, and g represents a fully connected layer containing non-linear activation, used to transform the global descriptor and generate channel weight coefficients Sc;

[0166] The channel attention mechanism, which highlights the key features through adaptive weights, includes:

[0167]

[0168] in, This represents the value of the c-th channel at position (i,j) in the weighted feature map.

[0169] Optionally, each residual module includes a skip connection and a main path; the skip connection is used to adjust the number of channels of the input feature map to match the number of channels of the main path output.

[0170] The main path consists of three convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer uses a 1x1 convolution to adjust the number of channels in the input feature map. The second layer uses a 3x3 convolution kernel to capture local features of the input data. The third layer uses a 1x1 convolution to adjust the number of channels and integrate the features extracted by the previous layer.

[0171] The above step S101 involves acquiring a hyperspectral image of the sample blade, which is specifically achieved using a gantry mechanism. The gantry mechanism is as follows: Figure 10-23 As shown.

[0172] The gantry mechanism includes two sets of uprights 1 and two first crossbars 2. The two sets of uprights 1 are respectively installed at the bottom ends of the two first crossbars 2. A second crossbar 3 is provided at the top of the two first crossbars 2. The second crossbar 3 moves back and forth at the top of the two first crossbars 2 through a first drive mechanism provided at both ends. A second moving plate 8 is provided on one side of the second crossbar 3. The second moving plate 8 moves back and forth on one side of the second crossbar 3 through a second drive mechanism. A sleeve plate 21 is installed on the outer wall of the side of the second moving plate 8 away from the second crossbar 3. A vertical rod 4 is slidably installed inside the sleeve plate 21. A mounting plate 22 is installed at the bottom end of the vertical rod 4. A camera 23 is installed on one outer wall of the mounting plate 22. A third drive mechanism is provided inside the sleeve plate 21 for driving the vertical rod 4 to move the camera 23 up and down together. In this embodiment, the camera 23 can be an RGB camera, a multispectral camera, a hyperspectral camera, or a thermal imaging camera. It can be flexibly replaced according to actual usage requirements. An integrated industrial control computer is provided on the side of the mounting plate 22 away from the camera 23, so that the camera 23 has real-time data processing and storage capabilities.

[0173] The first driving mechanism includes first movable plates 14 mounted at the bottom of both ends of the second crossbar 3. Multiple first sliders 16 are mounted at the bottom of each of the two first movable plates 14, and first slide rails 11, which slide in cooperation with the multiple first sliders 16, are mounted at the top of each of the two first crossbars 2. The multiple first sliders 16 at the bottom of the first movable plates 14 slide in cooperation with the first slide rails 11 at the top of the first crossbars 2, thereby limiting and guiding the back-and-forth movement of the first movable plates 14 at the top of the first crossbars 2.

[0174] Both first crossbars 2 have a first rack 12 installed inside, and both first moving plates 14 have a first drive motor 7 installed on their tops. The output ends of both first drive motors 7 pass through the first moving plates 14 and are fitted with first gears 15 that mesh with the first racks 12. After starting, the first drive motors 7 can drive the first gears 12 to rotate. Since the first racks 12 are fixedly installed inside the first crossbars 2, when the first gears 15 are driven by the first drive motors 7 to rotate forward and backward, they will move forward or backward along the first racks 12. This will cause the first moving plates 14 to move synchronously on the top of the first slide rails 11, thereby causing the second crossbar 3 to move back and forth on the tops of the two first crossbars 2. This allows for the X-axis movement adjustment of multiple components, including the second crossbar 3, the vertical bar 4, and the camera 23.

[0175] One of the first crossbars 2 has a first rectangular plate 5 installed on one side of its outer wall, and a first drag chain 6 is installed on one end of the second crossbar 3 near the first rectangular plate 5. The first drag chain 6 is placed on top of the first rectangular plate 5.

[0176] The second driving mechanism includes multiple second sliders 25 mounted on the outer wall of the second moving plate 8 near the second crossbar 3. Two second slide rails 17 are mounted on the side of the second crossbar 3 near the second moving plate 8, and the multiple second sliders 25 are slidably connected to their respective second slide rails 17. The multiple second sliders 25 on one side of the outer wall of the second moving plate 8 slidingly engage with the second slide rails 17 mounted on the outer wall of the second crossbar 3, thereby limiting and guiding the back-and-forth movement of the second moving plate 8 on the outer wall of the second crossbar 3.

[0177] A second drive motor 20 is mounted on the outer wall of the second movable plate 8 near the sleeve plate 21. A second gear 24 is mounted on the output end of the second drive motor 20 through the second movable plate 8. A second rack 18 is mounted on the outer wall of the second crossbar 3 near the second movable plate 8. The second gear 24 meshes with the second rack 18. After the second drive motor 20 is started, it can drive the second gear 24 to rotate. Since the second rack 18 is fixedly set inside the second crossbar 3, when the second gear 24 is driven by the second drive motor 20 to rotate forward and backward, it will drive itself to move left or right along the second rack 18. This will cause the second movable plate 8 to move synchronously on the second slide rail 17, and then drive the second movable plate 8 to move back and forth on one side of the second crossbar 3, so as to realize the Y-axis movement adjustment of multiple components such as the second movable plate 8, the vertical rod 4, and the camera 23.

[0178] A second rectangular plate 10 is installed on the outer wall of the second crossbar 3 away from the second movable plate 8. A second drag chain 9 is installed on the outer wall of the second movable plate 8 near the second crossbar 3. The second drag chain 9 is placed on top of the second rectangular plate 10. A third drag chain 13 is also installed on the outer wall of the second movable plate 8 away from the second crossbar 3.

[0179] A third slide rail 28 is installed on one outer wall of the vertical rod 4, and the inner wall of one side of the sleeve 21 is slidably engaged with the third slide rail 28. The slidable connection between the inner wall of one side of the sleeve 21 and the third slide rail 28 can guide and limit the up and down movement of the vertical rod 4 inside the sleeve 21.

[0180] The third drive mechanism includes a third drive motor 26 mounted on the outer wall of one side of the sleeve plate 21. A third gear is mounted on the output end of the third drive motor 26 through the sleeve plate 21. A third rack 27 is mounted on the outer wall of one side of the vertical rod 4. The third gear meshes with the third rack 27. The third gear is located inside the sleeve plate 21, and its driving method is the same as that of the first gear 15 and the second gear 24 mentioned above. Therefore, the position of the third gear is not marked in the figure. The specific driving method of the third gear is to drive it to rotate forward and backward inside the sleeve plate 21 by the drive of the third drive motor 26. Since the third rack 27 is fixedly set on one side of the vertical rod 4, when the third gear is driven by the third drive motor 26 to rotate forward and backward, it will drive itself to move up or down along the third rack 27, thereby driving the vertical rod 4 to move up and down synchronously inside the sleeve plate 21, realizing the Z-axis movement adjustment of the vertical rod 4 and multiple components such as the camera 23.

[0181] In this embodiment, during the operation of the electrically driven equipment such as the first drive motor 7, the second drive motor 20, and the third drive motor 26, the related cable harnesses need to move and bend frequently, which may lead to wear and damage to the cable harnesses. However, by setting up the first cable chain 6, the second cable chain 9, and the third cable chain 13, and installing the required various lines inside the first cable chain 6, the bending and pulling of the cable harnesses and wear can be reduced, thereby extending the service life of the cable harnesses. Furthermore, the first cable chain 6, the second cable chain 9, and the third cable chain 13 can withstand mechanical movement and vibration, preventing the cables and harnesses from loosening or falling off during equipment operation, thereby ensuring the stable operation of the equipment. In addition, the first cable chain 6, the second cable chain 9, and the third cable chain 13 typically have low noise and vibration reduction characteristics, which helps to reduce the noise and vibration generated during equipment operation and improve the user experience of the equipment.

[0182] When it is necessary to use camera 23 to detect plants at different positions and heights, the first drive motor 7 is activated to rotate the first gear 15, which in turn moves the first moving plate 14 synchronously on the top of the first slide rail 11. This, in turn, causes the second crossbar 3 to move back and forth on the top of the two first crossbars 2, enabling the X-axis movement adjustment of multiple components, including the second crossbar 3, the vertical bar 4, and the camera 23. The plant being photographed is planted in the area below the two first crossbars 2.

[0183] After the second drive motor 20 is started, it can drive the second gear 24 to rotate, which will drive it to move left or right along the second rack 18. This will cause the second moving plate 8 to move synchronously on the second slide rail 17, and then drive the second moving plate 8 to move back and forth on one side of the second crossbar 3, so as to realize the Y-axis movement adjustment of multiple components such as the second moving plate 8, the vertical bar 4 and the camera 23.

[0184] Furthermore, the third drive motor 26 drives the third gear to rotate in both directions inside the sleeve 21, causing it to move up or down along the third rack 27, which in turn drives the vertical rod 4 to move up and down synchronously inside the sleeve 21, thus enabling the vertical rod 4 and multiple components such as the camera 23 to move and adjust in the Z-axis direction.

[0185] During use, the camera 23 can be moved along the X-axis of the second horizontal bar 3, the Y-axis of the second moving plate 8, and the Z-axis of the vertical bar 4, enabling the pod to move precisely in the X, Y, and Z axes. This ensures that the camera 23 can accurately detect plants at different positions and heights. Since the first drive motor 7, the second drive motor 20, and the third drive motor 26 are all servo motors, the stability and repeatability of the relevant components during movement can be ensured by using servo motors in conjunction with guide rails.

[0186] The multi-source data acquisition pod, composed of components such as the vertical rod 4 and camera 23, supports both stationary and dynamic scanning modes to meet diverse detection needs. Camera 23 can be flexibly replaced with an RGB camera, multispectral camera, hyperspectral camera, or thermal imaging camera as required. An industrial control computer is integrated within the pod, providing real-time data processing and storage capabilities. Furthermore, the pod's structural design supports rapid module replacement for easy functional expansion. Equipped with a touch interface and software system, users can easily set the pod's movement trajectory and acquisition parameters. The pod also supports remote data transmission and equipment status monitoring. Through these methods, it can meet the detection needs of different plant species and growth stages.

[0187] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for constructing a leaf detection model based on hyperspectral imaging technology, characterized in that, include: Acquire hyperspectral images of the sample blade; The hyperspectral image of the sample blade is preprocessed to obtain the target hyperspectral image; The preprocessing includes at least one of masking, redundancy trimming, and noise reduction. A competitive adaptive reweighted sampling method is used to filter the band images of the target hyperspectral image, resulting in multiple target band images; Multiple target band images are detected using a pre-built deep learning model, which is used to extract spectral features and fuse spatial features for each target band image.

2. The method according to claim 1, characterized in that, The method of detecting multiple target band images using a pre-built deep learning model includes: The spectral data of the multiple target band images are extracted using a 3D residual module and the spatial data are fused. 3D max pooling is used to reduce the dimensionality of the extracted data while retaining key features, including leaf water content and chlorophyll content. A channel attention mechanism is used to highlight the key features through adaptive weights, thereby obtaining the first target key features; The spatial features of the multiple target band images are extracted using a 2D residual module, and the dimensionality of the extracted spatial data is further reduced by 2D max pooling to obtain the second target key features, which include blade surface details and texture information. By integrating the key features of the first target and the key features of the second target, the state characterization of the sample blade is obtained.

3. The method according to claim 2, characterized in that, The channel attention mechanism includes a squeezing phase and an incentive phase; The extrusion process includes: Among them, Z C This represents the global descriptor for the c-th channel, x. i,j,k This represents the value of the c-th channel at position (i,j) of the input feature map, where H and W represent the height and width of the feature map, respectively. The incentive phase includes: Where σ represents the sigmoid activation function, and g represents a fully connected layer containing non-linear activation, used to transform the global descriptor and generate channel weight coefficients Sc; The channel attention mechanism, which highlights the key features through adaptive weights, includes: in, This represents the value of the c-th channel at position (i,j) in the weighted feature map.

4. The method according to claim 2 or 3, characterized in that, Each residual module includes a skip connection and a main path; the skip connection is used to adjust the number of channels in the input feature map to match the number of channels in the output of the main path. The main path consists of three convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer uses 1x1 convolutions to adjust the number of channels in the input feature map. The second layer uses 3x3 convolutional kernels to capture local features of the input data. The driving layer uses 1x1 convolutions to adjust the number of channels and integrate the features extracted by the previous layer.

5. The method according to claim 2 or 3, characterized in that, The preprocessing of the hyperspectral image of the sample blade includes: The sample blade is masked to obtain a mask image; Extract the minimum bounding rectangle from the mask image, and crop the area outside the minimum bounding rectangle to obtain the mask border image; The mask border image is scaled to a preset pixel size using interpolation. A smoothing algorithm is used to denoise the scaled mask border image.

6. A blade detection device, characterized in that, include: The image acquisition module is used to acquire hyperspectral images of the sample blades. The preprocessing module is used to preprocess the hyperspectral image of the sample blade to obtain the target hyperspectral image; The preprocessing includes at least one of masking, redundancy trimming, and noise reduction. The image filtering module is used to filter the band images of the target hyperspectral image using a competitive adaptive reweighted sampling method to obtain multiple target band images; The image detection module is used to detect multiple target band images using a pre-built deep learning model, wherein the deep learning model is used to extract spectral features and fuse spatial features for each target band image.

7. The apparatus according to claim 6, characterized in that, The image detection module is specifically used for: The spectral data of the multiple target band images are extracted using a 3D residual module and the spatial data are fused. 3D max pooling is used to reduce the dimensionality of the extracted data while retaining key features, including leaf water content and chlorophyll content. A channel attention mechanism is used to highlight the key features through adaptive weights, thereby obtaining the first target key features; The spatial features of the multiple target band images are extracted using a 2D residual module, and the dimensionality of the extracted spatial data is further reduced by 2D max pooling to obtain the second target key features, which include blade surface details and texture information. By integrating the key features of the first target and the key features of the second target, the state characterization of the sample blade is obtained.

8. The apparatus according to claim 7, characterized in that, The channel attention mechanism includes a squeezing phase and an incentive phase; The extrusion process includes: Among them, Z C This represents the global descriptor for the c-th channel, x. i,j,k This represents the value of the c-th channel at position (i,j) of the input feature map, where H and W represent the height and width of the feature map, respectively. The incentive phase includes: Where σ represents the sigmoid activation function, and g represents a fully connected layer containing non-linear activation, used to transform the global descriptor and generate channel weight coefficients Sc; The channel attention mechanism, which highlights the key features through adaptive weights, includes: in, This represents the value of the c-th channel at position (i,j) in the weighted feature map.

9. The apparatus according to claim 6 or 7, characterized in that, Each residual module includes a skip connection and a main path; the skip connection is used to adjust the number of channels in the input feature map to match the number of channels in the output of the main path. The main path consists of three convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer uses a 1x1 convolution to adjust the number of channels in the input feature map. The second layer uses a 3x3 convolution kernel to capture local features of the input data. The third layer uses a 1x1 convolution to adjust the number of channels and integrate the features extracted by the previous layer.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.

Images