Method for identifying plant withering state by using polarization imaging
By using polarization imaging technology and an improved MobileNetV1 model, the problems of unstable identification accuracy and high cost in plant wilting monitoring have been solved, achieving high-precision and low-cost wilting state identification and dynamic monitoring.
Patent Information
- Application Number
- CN202510877574.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-27
- Publication Date
- 2025-10-31
AI Technical Summary
Existing plant wilting monitoring technologies suffer from inconsistent accuracy and high costs.
Four-channel polarized light intensity images of plant leaves were acquired using polarization imaging technology. Stokes parameters and linear polarization degree were calculated, HIS pseudo-periodic color coding was constructed, polarized leaf vein features were obtained, and leaf vein skeleton diagrams were processed using an improved MobileNetV1 model to finally identify the wilting state of the plant.
It achieves high-precision identification of plant wilting state, reduces equipment costs, and improves the stability and robustness of monitoring results, enabling clear observation of changes in leaf vein information.
Smart Images
Figure CN120877092A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of plant state recognition methods, specifically relating to a method for recognizing plant wilting state using polarization imaging. Background Technology
[0002] Wilting is a vital characteristic exhibited by plants under various stress conditions, and it is also one of the most direct and earliest phenotypic indicators of water stress in plants. Rapidly and accurately identifying wilting signals in leaves or canopy can capture information about plant water deficits before a sudden drop in soil moisture content, providing a scientific basis for precision irrigation and water-saving management.
[0003] Commonly used plant wilting monitoring technologies mainly include visible light digital imaging and machine vision, thermal infrared imaging, multispectral / hyperspectral imaging, and traditional visual observation. Visible light digital imaging and machine vision utilize cameras to capture the appearance features of plant leaves and identify wilting through image processing methods based on color, morphology, or texture. However, this method is easily affected by light intensity, shooting angle, and background complexity, leading to unstable detection accuracy. Thermal infrared imaging technology measures leaf temperature to determine water stress and temperature increases caused by stomatal closure, enabling the detection of early physiological changes in leaves; however, thermal imaging instruments are expensive. Multispectral / hyperspectral imaging technology can acquire specific spectral information associated with leaf moisture or chemical composition, providing rich data dimensions and theoretically enabling high-precision inversion of moisture and physiological indicators. However, its equipment is costly and sensitive to changes in ambient light, resulting in unstable detection accuracy. Summary of the Invention The purpose of this invention is to provide a method for identifying the wilting state of plants using polarization imaging, thereby solving the problems of unstable accuracy and high cost in existing identification methods.
[0004] The technical solution adopted in this invention is a method for identifying the wilting state of plants using polarization imaging, and the specific steps are as follows: Step 1: Acquire four-channel polarized light intensity images of plant leaves and calculate Stokes parameters and degree of linear polarization; Step 2: Construct HIS pseudo-periodic color coding, and obtain polarized leaf vein feature P by fusing HIS pseudo-periodic color coding. Step 3: Obtain the leaf vein skeleton map of the plant leaf through the polarized leaf vein feature P; Step 4: Construct the improved MobileNetV1 model and train the improved MobileNetV1 model; Step 5: Preprocess the leaf vein skeleton map and input the preprocessed leaf vein skeleton map into the trained improved MobileNetV1 model to predict the wilting state of the plant.
[0005] The invention is further characterized by: In step 1, the expression for the Stokes parameter is: (1) In equation (1), S represents the Stokes vector; This represents the total light intensity, which is the superposition of polarized light intensities in the 0° and 90° directions; This represents the difference in polarization intensity between the 0° and 90° directions; This represents the intensity difference of polarized light in the +45° and -45° directions; This indicates the intensity difference between right-handed and left-handed circularly polarized light; This is called the Stokes parameter; Indicates the intensity of polarized light in the 0° direction; Indicates the intensity of polarized light in the 90° direction; This indicates the intensity of polarized light at a 45° angle. This indicates the intensity of polarized light at a 135° angle. Indicates the intensity of right-handed circularly polarized light; This indicates the intensity difference of left-handed circularly polarized light.
[0006] In step 1, the expression for the degree of linear polarization is: (2) In equation (2), DoLP Indicates the degree of linear polarization.
[0007] In step 1, the four-channel polarized light intensity image is acquired by a polarization imaging system. The polarization imaging system includes a sample stage (3). A polarizer and a focal plane polarization camera (4) are respectively set on one side of the sample stage (3). The polarizer and the focal plane polarization camera (4) are set on the same side. A micro-polarization array is set inside the focal plane polarization camera (4). The polarizer (2) includes an LED light source (1), and a linear polarizing film (2) is attached to the LED light source (1). The angle between the outgoing light from the LED light source (1) and the incident light from the focal plane polarizing camera (4) is 35°.
[0008] In step 2, the expression for HIS pseudo-periodic color encoding is: (3) In equation (3), H Indicates hue; S Indicates saturation; I Indicates brightness; i for DoLP The pixel value.
[0009] The specific process of step 3 is as follows: Step 3.1: Separate the polarized leaf vein feature P obtained in Step 2 into RGB channels to obtain images with three channels: R, G, and B. Step 3.2: Stretch the B channel image obtained in step 3.1 to obtain the stretched channel image; The expression is: (5) In equation (5), ; This represents the stretched channel image; A B-channel image representing the polarized leaf vein feature P; Step 3.3, for Normalization is performed to obtain the normalized image; The expression is: (8) Normalized image Convert to binary image Transform binary image A diagram of the vein framework of a plant leaf; The expression is: (9).
[0010] The improved MobileNetV1 model consists of a backbone network and a classification head; The backbone network consists of a channel attention module, an initial convolutional layer, and 13 depthwise separable convolutional blocks; The classification head consists of global average pooling, Dropout layer, fully connected layer, activation function, BN layer, fully connected layer, activation function, BN layer, and Sigmoid layer; The activation functions in the classification head consist of ReLU activation function and L2 regularization.
[0011] In step 4, a phased training strategy is adopted during training. First, the backbone network is frozen and only the classification head is trained. Then, the backbone network is unfrozen for fine-tuning. Finally, cosine annealing with learning rate decay across the entire number of steps is applied for further parameter tuning to obtain the trained and improved MobileNetV1 model.
[0012] In step 4, the loss function used during training is:
[0013] in γ = 2.0 represents the model's predicted probability for the current sample, α = 0.75 emphasizes samples that are difficult to classify, and assigns higher weights to the foreground class.
[0014] In step 5, the specific preprocessing procedure is as follows: Step 5.1: The leaf vein skeleton map of the plant leaf obtained in Step 3 is separated into LAB channels to obtain L channel image, A channel image and B channel image; the L channel image is enhanced by limiting contrast adaptive histogram equalization, the enhanced L channel image is recombined with A channel and B channel, and the recombined image is normalized to [0,1]. Step 5.2: Use the perceptual hash algorithm to remove duplicates from the normalized image obtained in Step 5.1; Step 5.3: The deduplicated image obtained in step 5.2 is subjected to random rotation, translation, scaling and shearing to simulate changes in camera posture and distance, and is further enhanced by horizontal flipping; at the same time, brightness jitter is applied to simulate the actual fluctuation of light intensity for data augmentation.
[0015] The beneficial effects of this invention are: (1) The present invention utilizes polarization imaging to identify the wilting state of plants. The leaf vein skeleton map is used as input. An improved MobileNetV1 model is constructed to identify the wilting state of plants with high precision. It can dynamically monitor the water changes in the early stage of plant wilting, and has low sensitivity to environmental changes and stronger robustness, thereby making the monitoring results more stable. (2) The present invention utilizes polarization imaging to identify the wilting state of plants. It acquires four-channel images through a polarization imaging system, and the equipment cost is low. (3) The present invention utilizes polarization imaging to identify the wilting state of plants. By using linear polarization images and polarization image processing technology, the changes in leaf vein information of the tested plant can be clearly displayed. The changes in leaf vein information during the process of the plant changing from healthy to wilting can be clearly observed, thereby improving the accuracy of wilting state identification. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the polarization imaging system in the method for identifying plant wilting state using polarization imaging of the present invention. Figure 2 This is a 0-degree polarization intensity map in the method for identifying plant wilting state using polarization imaging in this invention. Figure 3 This invention relates to a method for identifying plant wilting using polarization imaging; Figure 4 The leaf vein feature image P is used in the method for identifying plant wilting state using polarization imaging in this invention. Figure 5 This is a diagram of the plant vein skeleton in the method for identifying plant wilting state using polarization imaging, as described in this invention.
[0017] In the figure, 1. Light source, 2. Linear polarizing film, 3. Sample stage, 4. Focusing plane polarizing camera. Detailed Implementation
[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0019] Example 1 This invention provides a method for identifying plant wilting using polarization imaging, specifically implemented according to the following steps: Step 1: Acquire images of plant leaves using a polarization imaging system. , , , Four-channel polarized light intensity images were used to calculate Stokes parameters and the degree of linear polarization. Step 2: Construct HIS pseudo-periodic color coding according to the mapping rules, and obtain the polarized leaf vein feature P by fusing the HIS pseudo-periodic color coding. Step 3, obtain the following through polarized leaf vein feature P: Figure 4 The diagram shows the leaf vein skeleton of the plant shown. Step 4: Construct the improved MobileNetV1 model and train the improved MobileNetV1 model; Step 5: Preprocess the leaf vein skeleton map and input the preprocessed leaf vein skeleton map into the trained improved MobileNetV1 model to predict the wilting state of the plant.
[0020] Example 2 Based on Example 1, in step 1, the expression for the Stokes parameter is: (1) In equation (1), S represents the Stokes vector; This represents the total light intensity, which is the superposition of polarized light intensities in the horizontal (0°) and vertical (90°) directions; It represents the difference in polarization intensity between the horizontal (0°) and vertical (90°) directions, and usually reflects the distribution of linearly polarized light in the horizontal-vertical coordinate system; This represents the intensity difference of polarized light in the +45° and -45° directions, reflecting the contrast of linearly polarized light in the diagonal direction; It represents the intensity difference between right-handed and left-handed circularly polarized light, and is used to characterize the circular polarization component; This is called the Stokes parameter; Indicates the intensity of polarized light in the 0° direction; Indicates the intensity of polarized light in the 90° direction; This indicates the intensity of polarized light at a 45° angle. This indicates the intensity of polarized light at a 135° angle. Indicates the intensity of right-handed circularly polarized light; This indicates the intensity difference of left-handed circularly polarized light; In practical applications, the detection of the polarization state of plant leaves has revealed circular polarization information. The amount of information contained is relatively small, and the experimental system is relatively complex. Introducing too many optical components can easily cause errors. Therefore, the proportion of linearly polarized states in the optical field can be obtained without measuring the circularly polarized components.
[0021] like Figure 2 As shown, when linearly polarized light is irradiated at 90 degrees onto a plant sample, Channel polarization intensity image; The expression for the degree of linear polarization is: (2) In equation (2), DoLP Indicates the degree of linear polarization; The degree of linear polarization reflects the proportion of linearly polarized light in the total optical field polarization states; like Figure 3 As shown, the linear polarization degree image is obtained by formula (2).
[0022] Example 3 Based on Example 2, the structure of the polarization imaging system is as follows: it includes a sample stage 3, a polarizer and a focal plane polarization camera 4 are respectively arranged on one side of the sample stage 3, the polarizer and the focal plane polarization camera 4 are arranged on the same side, and a micro polarization array is arranged inside the focal plane polarization camera 4. The polarizer 2 includes an LED light source 1 (which is a near-parallel light source), and a linear polarizing film 2 is attached to the LED light source 1; the included angle between the LED light source 1, the sample stage 3, and the split-focus plane polarizing camera 4 is 35°. LED light source 1 is obliquely incident on the target object being measured, and the focal plane polarization camera 4 is receiving the reflected and scattered light after interacting with the target object; In the optical path: the near-parallel LED light source 1 emits stable unpolarized light that is incident on the linear polarization film 2. After the polarized light irradiates the surface of the target object on the sample stage 3, the reflected and scattered light after interacting with the target object is received by the split-focus plane polarization camera 4.
[0023] Example 4 Based on Example 3, in step 2, the expression for constructing the HIS pseudo-periodic color encoding is as follows: (3) In equation (3), H Indicates hue; S Indicates saturation; I Indicates brightness; i for DoLP Pixel values; The linear distribution of DoLP becomes a circular hue ( The HIS pseudo-periodic color coding constructed by this invention can reflect changes in polarization information in the hue dimension. Brightness I is due to The intensity images are obtained by normalizing the channel polarization light intensity images. Since the incident light source is 90° linearly polarized light, the intensity images received at four different angles are as follows: Channel polarized light intensity images contain reflected and scattered light of varying intensity, resulting in uneven brightness in the image. Channel polarization intensity images avoid the influence of strong and weak illumination. Therefore, using a uniform brightness intensity image as I can better remove the non-uniformity problem in polarization images.
[0024] Example 5 Based on Example 4, the specific process of step 3 is as follows: Step 3.1: Separate the polarized leaf vein feature P obtained in Step 2 into RGB channels to obtain three channel images: R, G, and B. Use the brightness I in Step 2 as the B channel image. Step 3.2: Stretch the B channel image obtained in step 3.1 to obtain the stretched channel image; The expression is: (5) In equation (5), ; This represents the stretched channel image; A B-channel image representing the polarized leaf vein feature P; Step 3.3, for Normalization is performed to obtain the normalized image; The expression is: (8) Normalized image Convert to binary image Transform binary image As a diagram of the vein framework of a plant leaf, such as Figure 5 As shown; The expression is: (9); Although the leaf vein enhancement image highlights the leaf vein information, the contrast between the leaf mesophyll information and the background is low, which causes great interference for identifying the wilting state of plants by observing only changes in leaf vein features. Therefore, this invention segments the leaf veins based on the leaf vein enhancement image to obtain an image containing only leaf veins and no background, which is the leaf vein skeleton image. The algorithm uses a double threshold for threshold segmentation. It allows for more flexible extraction of plant leaf vein images or background interference, preserving the leaf vein texture features while removing interference from uneven external lighting.
[0025] Example 6 Based on Example 5, in step 4, the improved MobileNetV1 model consists of a backbone network and a classification head; The backbone network consists of a channel attention module, an initial convolutional layer, and 13 depthwise separable convolutional blocks. The classification head consists of global average pooling, a Dropout layer (p=0.2), a fully connected layer, an activation function, a Batch Normalization (BN) layer, a fully connected layer, an activation function, a BN layer, and a Sigmoid layer. The activation functions in the classification heads all consist of ReLU activation function and L2 regularization (regularization strength λ=0.01); In this model, alpha = 0.75; The channel attention module enhances the discrimination of high-frequency texture information of plant leaf veins and suppresses irrelevant background noise; The initial convolutional layer has a 3×3 kernel, a stride of 2, and 24 channels. Each depthwise separable convolutional block consists of depthwise convolutions, batch normalization (BN) layers, ReLU activation functions, pointwise convolutions, BN layers, and ReLU activation functions. In the first depthwise separable convolutional block, the depthwise convolution kernels are 3×3 with a stride of 1, and the pointwise convolution kernels are 1×1 with 48 channels. In the second depthwise separable convolutional block, the depthwise convolution kernels are 3×3 with a stride of 2, and the pointwise convolution kernels are 1×1 with 96 channels. In the third depthwise separable convolutional block, the depthwise convolution kernels are 3×3 with a stride of 1, and the pointwise convolution kernels are 1×1 with 96 channels. In the fourth depthwise separable convolutional block, the depthwise convolution kernels are 3×3 with a stride of 2, and the pointwise convolution kernels are 1×1 with 96 channels. The first depthwise separable convolutional block has a 1×1 kernel and 192 channels. The fifth depthwise separable convolutional block has a 3×3 kernel for depthwise convolutions with a stride of 1, and a 1×1 kernel for pointwise convolutions with 192 channels. The sixth depthwise separable convolutional block has a 3×3 kernel for depthwise convolutions with a stride of 1, and a 1×1 kernel for pointwise convolutions with 192 channels. The seventh depthwise separable convolutional block has a 3×3 kernel for depthwise convolutions with a stride of 2, and a 1×1 kernel for pointwise convolutions with 384 channels. The eighth to thirteenth depthwise separable convolutional blocks all have a 3×3 kernel for depthwise convolutions with a stride of 2, and all have a 1×1 kernel for pointwise convolutions with 384 channels. Global average pooling layers are used to remove spatial redundancy and reduce parameters; Dropout layers are used to alleviate overfitting on small samples; BN layers are used to stabilize internal covariate offsets; The processing flow of the improved MobileNetV1 model is as follows: Step S1: Input the leaf vein skeleton map into the channel attention module for processing, and output the weighted multi-channel feature map; The specific process is as follows: Global average pooling and global max pooling are used to capture channel-level statistical features in the leaf vein skeleton map. The expression is as follows: (10) in, This represents the value of the feature map at spatial position (i, j) on the c-th channel; T and W represent the height and width of the feature map, respectively. This represents the scalar obtained after performing global average pooling on the c-th channel; This represents the scalar obtained after performing global max pooling on the c-th channel; By learning the nonlinear relationships between channels using a multilayer perceptron with shared parameters, a weight vector is obtained: (11) in, The Sigmoid activation function has an output weight range of [0,1]; MLP is a multilayer perceptron. The weight vector is multiplied by the leaf vein skeleton map according to the channel to obtain the weighted multi-channel feature map, expressed as: (12) Where F is the output image of the thirteenth depthwise separable convolutional block, with shape (H,W,C); The weight vector has a length of C; This represents the weighted output feature map, with the same shape as F, i.e. (H, W, C); Step S2: Input the weighted multi-channel feature map into the initial convolutional layer, input the output of the initial convolutional layer into 13 depthwise separable convolutional blocks, and input the output of the 13th depthwise separable convolutional block into the classification head to predict the wilting state of the plant.
[0026] The improved MobileNetV1 model is trained using a phased training strategy. First, the backbone network is frozen and only the classification head is trained. The learning rate for this phase is lr = 1e-3. Then, the backbone network is unfrozen and fine-tuned for several rounds. The learning rate for this phase is lr = 1e-4. Finally, cosine annealing with learning rate decay across the entire training process is applied for fine-tuning. The initial learning rate for this phase is lr = 1e-5. The final trained improved MobileNetV1 model is obtained. Finally, the network's ability to distinguish between healthy and wilted features is evaluated by using dimensional metrics such as accuracy, loss function, and F1 score. The loss function during training is:
[0027] in γ = 2.0 is used to significantly amplify the gradient signal of samples that are difficult to predict correctly and suppress the gradient of easily classified samples. α is used to balance the loss contribution of positive and negative samples (or foreground and background). α = 0.75 gives higher weight to the foreground class (usually the "interesting" class) so that its loss contribution accounts for a larger proportion and helps the model better focus on the minority class in imbalanced data.
[0028] Example 7 Based on Example 6, the specific preprocessing process in step 5 is as follows: Step 5.1: The leaf vein skeleton map of the plant leaf obtained in Step 3 is separated into LAB channels to obtain L channel image, A channel image and B channel image; the L channel image is enhanced by limiting contrast adaptive histogram equalization (CLAHE), the enhanced L channel image is recombined with A channel and B channel, and the recombined image is normalized to [0,1]. Applying CLAHE to the L channel enhances local details while controlling noise. This step significantly improves the discernibility of leaf vein texture, laying a reliable foundation for subsequent feature extraction.
[0029] Step 5.2: Use the perceptual hash algorithm to remove duplicates from the normalized image obtained in Step 5.1; Step 5.3: The deduplicated image obtained in Step 5.2 is subjected to random rotation (±30°), translation (±20% for width and height), scaling (±30%) and shearing (±30%) to simulate changes in camera posture and distance, and is further enhanced by horizontal flipping; at the same time, brightness jitter (range 0.8~1.2) is applied to simulate the actual fluctuation of light intensity for data augmentation, and geometric and photometric transformations are combined to expand the diversity of training samples and suppress overfitting.
[0030] In summary, the method of this invention integrates polarization imaging technology with leaf vein segmentation algorithms, achieving accurate extraction of plant leaf vein texture in complex backgrounds. Based on the extracted leaf vein skeleton map, an improved MobileNetV1 model is trained to identify plant wilting. This method can detect subtle structural changes in leaf veins with high sensitivity during early stages of plant water content changes, improving monitoring accuracy and demonstrating good environmental robustness and cost-effectiveness. It provides a feasible technical route for the widespread application of polarization imaging technology in modern agriculture and forestry.
Claims
1. A method for identifying plant wilting state using polarization imaging, characterized in that, The specific steps are as follows: Step 1: Acquire four-channel polarized light intensity images of plant leaves and calculate Stokes parameters and degree of linear polarization; Step 2: Construct HIS pseudo-periodic color coding, and obtain polarized leaf vein feature P by fusing HIS pseudo-periodic color coding. Step 3: Obtain the leaf vein skeleton map of the plant leaf through the polarized leaf vein feature P; Step 4: Construct the improved MobileNetV1 model and train the improved MobileNetV1 model; Step 5: Preprocess the leaf vein skeleton map and input the preprocessed leaf vein skeleton map into the trained improved MobileNetV1 model to predict the wilting state of the plant.
2. The method for identifying plant wilting state using polarization imaging according to claim 1, characterized in that, In step 1, the expression for the Stokes parameter is: （1） In equation (1), S represents the Stokes vector; This represents the total light intensity, which is the superposition of polarized light intensities in the 0° and 90° directions; This represents the difference in polarization intensity between the 0° and 90° directions; This represents the intensity difference of polarized light in the +45° and -45° directions; This indicates the intensity difference between right-handed and left-handed circularly polarized light; This is called the Stokes parameter; Indicates the intensity of polarized light in the 0° direction; Indicates the intensity of polarized light in the 90° direction; This indicates the intensity of polarized light at a 45° angle. This indicates the intensity of polarized light at a 135° angle. Indicates the intensity of right-handed circularly polarized light; This indicates the intensity difference of left-handed circularly polarized light.
3. The method for identifying plant wilting state using polarization imaging according to claim 2, characterized in that, In step 1, the expression for the degree of linear polarization is: （2） In equation (2), DoLP Indicates the degree of linear polarization.
4. The method for identifying plant wilting state using polarization imaging according to claim 1, characterized in that, In step 1, the four-channel polarized light intensity image is acquired by a polarization imaging system. The polarization imaging system includes a sample stage (3). A polarizer and a focal plane polarization camera (4) are respectively set on one side of the sample stage (3). The polarizer and the focal plane polarization camera (4) are set on the same side. A micro-polarization array is set inside the focal plane polarization camera (4). The polarizer (2) includes an LED light source (1), and a linear polarizing film (2) is attached to the LED light source (1). The angle between the outgoing light from the LED light source (1) and the incident light from the focal plane polarizing camera (4) is 35°.
5. The method for identifying plant wilting state using polarization imaging according to claim 1, characterized in that, In step 2, the expression for HIS pseudo-periodic color encoding is: （3） In equation (3), H Indicates hue; S Indicates saturation; I Indicates brightness; i for DoLP The pixel value.
6. The method for identifying plant wilting state using polarization imaging according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1: Separate the polarized leaf vein feature P obtained in Step 2 into RGB channels to obtain images with three channels: R, G, and B. Step 3.2: Stretch the B channel image obtained in step 3.1 to obtain the stretched channel image; The expression is: （5） In equation (5), ; This represents the stretched channel image; A B-channel image representing the polarized leaf vein feature P; Step 3.3, for Normalization is performed to obtain the normalized image; The expression is: （8） Normalized image Convert to binary image Transform binary image A diagram of the vein framework of a plant leaf; The expression is: （9）。 7. The method for identifying plant wilting state using polarization imaging according to claim 1, characterized in that, In step 4, the improved MobileNetV1 model consists of a backbone network and a classification head; The backbone network consists of a channel attention module, an initial convolutional layer, and 13 depthwise separable convolutional blocks; The classification head consists of global average pooling, Dropout layer, fully connected layer, activation function, BN layer, fully connected layer, activation function, BN layer, and Sigmoid layer; The activation functions in the classification head consist of ReLU activation function and L2 regularization.
8. The method for identifying plant wilting state using polarization imaging according to claim 7, characterized in that, In step 4, a phased training strategy is adopted during training. First, the backbone network is frozen and only the classification head is trained. Then, the backbone network is unfrozen for fine-tuning. Finally, cosine annealing with learning rate decay across the entire number of steps is applied for further parameter tuning to obtain the trained and improved MobileNetV1 model.
9. The method for identifying plant wilting state using polarization imaging according to claim 7, characterized in that, In step 4, the loss function used during training is: in γ = 2.0 represents the model's predicted probability for the current sample, α = 0.75 emphasizes samples that are difficult to classify, and assigns higher weights to the foreground class.
10. The method for identifying plant wilting state using polarization imaging according to claim 1, characterized in that, In step 5, the specific preprocessing procedure is as follows: Step 5.1: Perform LAB channel separation on the leaf vein skeleton image of the plant leaf obtained in Step 3 to obtain L channel image, A channel image, and B channel image; The L channel image is enhanced by limiting contrast adaptive histogram equalization. The enhanced L channel image is then recombined with the A and B channels, and the recombined image is normalized to [0,1]. Step 5.2: Use the perceptual hash algorithm to remove duplicates from the normalized image obtained in Step 5.1; Step 5.3: The deduplicated image obtained in step 5.2 is subjected to random rotation, translation, scaling and shearing to simulate changes in camera posture and distance, and is further enhanced by horizontal flipping; at the same time, brightness jitter is applied to simulate the actual fluctuation of light intensity for data augmentation.