Soybean growth point detection and emergence rate method fusing super-resolution algorithm

Abstract

The present application provides a soybean growth point detection emergence rate method fused with a super-resolution algorithm, which solves the problems of detail loss of growth points caused by low resolution of unmanned aerial vehicle high-altitude aerial image, lack of optimization of existing super-resolution technology for agricultural scenes, and low detection accuracy caused by overlapping of densely planted seedlings. The present application is realized by the following steps: obtaining soybean seedling stage image and pretreating; inputting the standardized image dataset into a super-resolution model based on RealESRGAN architecture and embedded with a unified adaptive multi-branch high-frequency enhancement module for super-resolution reconstruction to obtain a high-resolution reconstructed image; inputting the high-resolution reconstructed image into a YOLOv8 growth point detection model to output growth point detection results; eliminating false detection boxes with confidence lower than a set threshold, counting the number of effective growth points and calculating the emergence rate. The present application balances high-altitude collection efficiency and detection accuracy, has low relative error, and is suitable for automatic detection of soybean seedling emergence rate.

Description

Technical Field

[0001] This invention relates to the field of seedling emergence detection, and more particularly to a method for detecting the emergence rate of soybean growth points by incorporating super-resolution algorithms. Background Technology

[0002] The soybean seedling stage is a crucial period that determines population structure, emergence uniformity, and later yield formation. Indicators such as seedling quantity, plant spacing, row spacing, coverage, and growth uniformity are key to field management and yield prediction. With the development of unmanned aerial vehicle (UAV) technology, UAV-based soybean seedling detection has gradually become a key research focus both domestically and internationally. However, existing technologies still have the following shortcomings.

[0003] In drone aerial photography, it is difficult to balance image resolution and acquisition efficiency. While drone technology provides large-scale seedling data for precision agriculture due to its high throughput, low cost, and non-destructive nature, low-altitude flights have limited coverage, low efficiency, and are limited by endurance. High-altitude flights, while offering wide coverage and high efficiency, suffer from reduced resolution and loss of seedling details, affecting data accuracy. Existing super-resolution technologies are mostly designed for general natural images and lack targeted optimization for small targets in agricultural scenarios (such as soybean growth points). After reconstruction, problems such as texture distortion and poor structural consistency easily occur, making it difficult to effectively restore the edge and texture details of growth points. In densely planted scenarios, overlapping seedlings lead to low detection accuracy. In planting scenarios with multiple seedlings per hole and large leaves, the leaves of multiple seedlings overlap in drone images, making it difficult to distinguish leaf features. This makes it difficult for traditional image feature-based detection methods to identify individual seedlings, resulting in high false negative and false positive rates, and making it impossible to achieve accurate seedling emergence rate statistics.

[0004] Therefore, there is an urgent need for a seedling detection method that can improve the resolution of low-resolution images, enhance the clarity of soybean seedling details, and improve detection accuracy, so as to achieve accurate detection of soybean seedlings and obtain field information. Summary of the Invention

[0005] To address the challenges of balancing image resolution and acquisition efficiency in UAV aerial photography, the insufficient targeting of existing super-resolution technologies for small agricultural targets, and low detection accuracy due to overlapping seedlings in densely planted scenarios, this invention proposes a method for detecting soybean growth points and emergence rate by incorporating super-resolution algorithms. The method includes: Step 1: Collect soybean seedling images and preprocess them to obtain a standardized image dataset; Step 2: Input the standardized image dataset into the trained super-resolution model for super-resolution reconstruction and output high-resolution reconstructed images; wherein, the super-resolution model is based on the Real-ESRGAN architecture and introduces a unified adaptive multi-branch high-frequency enhancement module; Step 3: Input the high-resolution reconstructed image into the growth point detection model and output the growth point detection results; Step 4: Remove false detection boxes with confidence levels below the set threshold from the growth point detection results, count the number of valid growth points, and calculate the germination rate based on the number of valid growth points and the theoretical number of seedlings.

[0006] Furthermore, in step one, The acquisition of soybean seedling images specifically involves using a drone equipped with an RGB camera to take aerial photos during the soybean emergence period. The aerial photography altitude is 30m and the flight speed is 3m / s. The preprocessing includes cropping, stitching, and coordinate calibration of the images acquired by the UAV to obtain a standardized image set with a resolution of 2048×2048 pixels; and dividing the standardized image dataset into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0007] Furthermore, in step two, The unified adaptive multi-branch high-frequency enhancement module is embedded at the bottleneck position between the Real-ESRGAN backbone network and the upsampling module. The unified adaptive multi-branch high-frequency enhancement module includes a high-frequency saliency map construction submodule, a multi-scale convolutional fusion submodule, a channel-based high-frequency modulation submodule, and a structure-adaptive residual fusion submodule, which are used to perform fine modeling of high-frequency information in three dimensions: space, channel, and structure.

[0008] Furthermore, the high-frequency saliency map construction submodule specifically comprises: The Laplacian operator is used to convolve each channel of the input features to extract the local high-frequency Laplacian response;

[0009] in, L It is a 3×3 Laplace operator; X These are intermediate features of the Real-ESRGAN backbone network; Low-frequency reconstruction is constructed using downsampling and upsampling operations. X low And obtain the high-frequency residual through difference. R ; The high-frequency response of the Laplace and the high-frequency residual are concatenated and then fused to obtain the comprehensive high-frequency map.

[0010] The absolute value of the comprehensive high-frequency map is normalized sample by sample to obtain a high-frequency saliency map, which serves as the structural guidance signal for subsequent multi-scale convolution and channel modulation.

[0011] In the formula, It is a very small constant. .

[0012] Furthermore, the multi-scale convolutional fusion submodule specifically comprises: Set the multi-scale convolution kernel set to 3, 5, and 7, and perform convolution extraction on the input features respectively; The convolutional outputs of each branch are stacked along the new dimension after group normalization and ReLU activation. The high-frequency saliency map is input into a 1×1 convolution and then normalized by Softmax to generate dynamic weights; The dynamic weights are then weighted and fused with the stacked multi-scale features to obtain the multi-scale fused features. Z msf .

[0013] Furthermore, the channel-type high-frequency modulation submodule specifically comprises: Global average pooling is performed on the multi-scale fused features to obtain the channel description vector; The channel description vectors are processed through two layers of multilayer perceptron and a sigmoid activation function to generate basic channel weights. ; Calculate the mean of the high-frequency saliency plot This yields a structural complexity index that reflects the overall high-frequency intensity of the image; The final modulation weight is :

[0014] The multi-scale fused features are multiplied point-by-point by the final modulation weights to output the channel-modulated features. .

[0015] Furthermore, the structural adaptive residual fusion submodule specifically comprises: The mean and standard deviation of the high-frequency saliency map are calculated to obtain a compact description of the image structural complexity. This compact description is then input into a lightweight multilayer perceptron, and fusion coefficients are generated using a sigmoid activation function. The intermediate features of the Real-ESRGAN backbone network are then weighted and fused with the channel-modulated features according to the fusion coefficients, and the residual fused features are output.

[0016] In the formula, The fusion coefficient is... ; The high-frequency significance plot shows the mean and standard deviation.

[0017] Furthermore, the loss function used during the training process of the super-resolution model is:

[0018] in, , , , , as well as These are the weighting coefficients for each loss term; For pixel loss; To perceive loss; To combat the losses; For high-frequency loss; For structural loss;

[0019] in, H This is a high-frequency extraction function; To reconstruct the high-frequency information of the image; This refers to high-frequency information in real images;

[0020] in, SSIM It is a structural similarity index; During the training of the super-resolution model, the Adam optimizer was used with an initial learning rate of 1×10⁻⁶. -4 With β1=0.9, β2=0.99, weight decay of 0, total training iterations of 400,000, batch size of 12, and the MultiStepLR learning rate scheduling strategy, the learning rate is decayed by 50% at the 400,000th iteration, and an exponential moving average is introduced to the generator parameters with a decay factor of 0.999. The generator and discriminator are optimized alternately at a 1:1 update frequency.

[0021] Furthermore, the growth point detection model uses YOLOv8n as the base model, employs the SGD optimizer for parameter updates, with an initial learning rate of 0.0001, momentum of 0.95, batch size of 8, and training epochs of 250. The growth point detection model filters out low-confidence predictions by setting a confidence threshold of 0.25 during the inference phase, and then uses non-maximum suppression with an intersection-union ratio threshold of 0.45 to remove duplicate responses that are too close in spatial location, thus obtaining the growth point detection results.

[0022] Furthermore, in step four, the set threshold is 0.5; the seedling emergence rate is calculated as follows: seedling emergence rate = (number of effective growth points detected / theoretical number of seedlings) × 100%.

[0023] The beneficial effects of this invention are: To address the challenges of balancing image resolution and acquisition efficiency in UAV aerial photography, the insufficient targeting of existing super-resolution technologies for small agricultural targets, and low detection accuracy due to overlapping seedlings in densely planted scenarios, this invention proposes a soybean growth point detection method that integrates super-resolution algorithms, offering the following improvements: 1. This invention proposes a super-resolution model adapted to agricultural scenarios and a unified adaptive multi-branch high-frequency enhancement module, which can perform targeted super-resolution reconstruction of low-resolution images acquired from high altitudes. At the same time, the reconstructed high-definition images can clearly restore the details of soybean growth points, solving the problem of balancing image resolution and acquisition efficiency.

[0024] 2. This invention significantly improves the detail recovery quality of small targets such as growth points. Through a unified adaptive multi-branch high-frequency enhancement module and its four sub-modules (high-frequency saliency map construction sub-module, multi-scale convolutional fusion sub-module, channel-based high-frequency modulation sub-module, and structure-adaptive residual fusion sub-module), this invention achieves fine modeling of high-frequency information in three dimensions: space, channel, and structure. This module can adaptively adjust the enhancement intensity according to the image structural complexity, using large convolutional kernels to enhance texture details in high-frequency regions and maintaining small convolutional kernels in flat regions to avoid over-sharpening, effectively restoring the texture and edge features of growth points.

[0025] 3. This invention solves the problem of false detection and missed detection caused by overlapping seedlings in dense scenes. It directly uses the soybean growth point as the detection target, which can accurately locate the growth point and effectively avoid repeated detection and missed detection caused by overlapping leaves, thus improving the positioning accuracy of the detection frame.

[0026] 4. This invention achieves fully automated processing from image acquisition, super-resolution reconstruction, growth point detection to emergence rate calculation; it can acquire images on a large scale and process them quickly through drones, significantly improving detection efficiency compared to manual methods, and effectively improving the accuracy of emergence rate calculation, providing technical support for precise management of soybean seedlings.

[0027] This solution addresses the small target characteristics of soybean growth points by designing a super-resolution reconstruction framework adapted to agricultural scenarios. Through high-frequency enhancement, it effectively restores the texture and edge features of growth points. It outperforms other comparative methods in terms of detail recovery, frequency domain consistency, and texture naturalness, and can more accurately reconstruct the structural and texture features in real-world scenes, providing reliable technical support for high-quality super-resolution reconstruction. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of the unified adaptive multi-branch high-frequency enhancement module network structure of the present invention; Figure 2 This is a schematic diagram illustrating the super-resolution reconstruction effect of images acquired by the UAV at a height of 30 meters according to the present invention; Figure 3 This is a schematic diagram illustrating the super-resolution reconstruction effect of images acquired by the UAV at a height of 5 meters according to the present invention; Figure 4 This is a schematic diagram of the soybean growth point detection results of the present invention. Detailed Implementation

[0030] 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.

[0031] Combination Figures 1-4 This invention proposes a method for detecting soybean seedling emergence rate by integrating super-resolution algorithms, the method comprising: Step 1: Collect soybean seedling images and preprocess them to obtain a standardized image dataset; Furthermore, in step one, The acquisition of soybean seedling images specifically involved using a drone equipped with an RGB camera to take aerial photos during the soybean emergence period. The aerial photography altitude was 100m, the flight speed was 5m / s, the forward overlap rate was 80%, and the lateral overlap rate was 70%. The preprocessing includes cropping, stitching, and coordinate calibration of the images acquired by the UAV to obtain a standardized image set with a resolution of 2048×2048 pixels; and dividing the standardized image dataset into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0032] Using a DJI M300 RTK drone equipped with a Zenmuse P1 RGB camera, aerial photography was conducted on a clear, windless morning at an altitude of 30m and a flight speed of 3m / s, capturing 826 raw images. After software cropping, stitching, and coordinate calibration, a standardized image dataset of 2048×2048 pixels was obtained, which was then divided into a training set of 661 images, a validation set of 83 images, and a test set of 82 images in an 8:1:1 ratio. The images were then converted to PNG format for later use.

[0033] Step 2: Input the standardized image dataset into the trained super-resolution model for super-resolution reconstruction and output high-resolution reconstructed images; wherein, the super-resolution model is based on the Real-ESRGAN architecture and introduces a unified adaptive multi-branch high-frequency enhancement module; Furthermore, in step two, The unified adaptive multi-branch high-frequency enhancement module is embedded at the bottleneck position between the Real-ESRGAN backbone network and the upsampling module. The unified adaptive multi-branch high-frequency enhancement module includes a high-frequency saliency map construction submodule, a multi-scale convolutional fusion submodule, a channel-based high-frequency modulation submodule, and a structure-adaptive residual fusion submodule, which are used to perform fine modeling of high-frequency information in three dimensions: space, channel, and structure.

[0034] The input to the Unified Adaptive Multi-Branch High-Frequency Enhancement Module (UAMHF) is intermediate features from the RRDB backbone.

[0035] This feature contains rich semantic information, but it still falls short in representing high-frequency structures (such as edges, textures, and fine-grained patterns). The core idea of ​​UAMHF is to first explicitly separate high-frequency components, then perform multi-scale convolution and channel modulation based on high-frequency saliency, and finally controllably synthesize the enhanced high-frequency features and the original features through structure-adaptive residual fusion, thereby obtaining reconstructed features with better structural consistency and texture details.

[0036] Real The ESRGAN generator structure can be divided into three stages: initial feature extraction, RRDB backbone network, and upsampling reconstruction. Let the input image be...

[0037] Shallow features are obtained after the initial convolutional layer.

[0038] Subsequently, shallow features are input into the RRDB backbone network, and deep semantic features are extracted through a multi-layer residual dense structure:

[0039] The main output is added to the shallow features to form the bottleneck feature:

[0040] Traditional Real ESRGAN directly inputs this feature into the upsampling module, but because the upsampling operation dilutes the high-frequency structure, the high-frequency information in the bottleneck feature often cannot be fully utilized. Therefore, this method inserts the UAMHF module at the bottleneck position, enabling it to explicitly enhance the high-frequency structure before upsampling.

[0041] This placement offers several advantages. Firstly, RRDB has already completed global semantic modeling of the input image, and the bottleneck features are relatively stable at the semantic level. Introducing UAMHF at this point allows for a focus on compensating and enhancing high-frequency components without disrupting the overall semantic structure. Secondly, placing high-frequency enhancement before upsampling avoids further smoothing of the high-frequency structure by upsampling, enabling the enhanced high-frequency information to directly participate in the spatial reconstruction process, thus preserving clearer edges and more complete local structures in the final output. Furthermore, UAMHF acts as a residual... This will not change the overall residual topology of the generator, which is beneficial for maintaining the stability of the training process and optimizing convergence characteristics.

[0042] Furthermore, the high-frequency saliency map construction submodule specifically comprises: The Laplacian operator is used to convolve each channel of the input features to extract the local high-frequency Laplacian response;

[0043] in, L It is a 3×3 Laplace operator; X For Real Intermediate features of the ESRGAN backbone network; Low-frequency reconstruction is constructed using downsampling and upsampling operations. X low And obtain the high-frequency residual through difference. R ; The high-frequency response of the Laplace and the high-frequency residual are concatenated and then fused to obtain the comprehensive high-frequency map.

[0044] The absolute value of the comprehensive high-frequency map is normalized sample by sample to obtain a high-frequency saliency map, which serves as the structural guidance signal for subsequent multi-scale convolution and channel modulation.

[0045] In the formula, It is a very small constant. .

[0046] The high-frequency saliency map construction branch aims to explicitly extract the high-frequency structure in the input features and transform it into a saliency map that can be used for subsequent modulation.

[0047] First, using a fixed 3×3 Laplace operator...

[0048] Convolution is performed on each channel of the input features to extract local high-frequency responses:

[0049] This operator exhibits a strong response to edge and textured regions, while maintaining near-zero output for flat regions, thus achieving explicit separation of high-frequency structures. Subsequently, to obtain a low-frequency approximation, the input features are reconstructed using downsampling and upsampling operations to build a low-frequency reconstruction:

[0050] And the high-frequency residuals are obtained through difference:

[0051] After concatenating the Laplace high-frequency response and the differential high-frequency residual, a comprehensive high-frequency map is obtained by convolutional fusion:

[0052] To ensure that the high-frequency saliency plot has a stable statistical range across different samples, its absolute value is normalized sample by sample:

[0053] in Used for numerical stability. The final high-frequency saliency map.

[0054] This serves as a structural guide signal for subsequent multi-scale convolution and channel modulation.

[0055] Furthermore, the multi-scale convolutional fusion submodule specifically comprises: Set the multi-scale convolution kernel set to 3, 5, and 7, and perform convolution extraction on the input features respectively; The convolutional outputs of each branch are stacked along the new dimension after group normalization and ReLU activation. The high-frequency saliency map is input into a 1×1 convolution and then normalized by Softmax to generate dynamic weights; The dynamic weights are then weighted and fused with the stacked multi-scale features to obtain the multi-scale fused features.Z msf .

[0056] The multi-scale convolutional fusion branch aims to dynamically select the convolutional kernel scale based on the high-frequency saliency map, enabling the model to favor large convolutional kernels in high-frequency regions to enhance texture details, while maintaining small convolutional kernels in low-frequency regions to avoid over-sharpening. This branch achieves spatially adaptive convolution through parallel convolutions and saliency-guided dynamic weighting. Specifically, multi-scale convolutional fusion involves: Let the set of multi-scale convolution kernels be...

[0057] The corresponding convolution output is

[0058] in Indicated by The convolutional kernel performs feature extraction. To enhance the stability and discriminability of features at different scales, the convolutional output of each branch undergoes GroupNorm and ReLU activation processing.

[0059] Stacking it along the new dimension yields

[0060] To generate dynamic weights based on the high-frequency saliency map, the saliency map is input into a 1×1 convolution and then normalized using Softmax:

[0061] in

[0062] The final multi-scale fusion feature is

[0063] This mechanism enables the selection of convolution kernels to be adaptively related to the image structure, thereby achieving differentiated processing in different regions.

[0064] The high-frequency modulation channel attention is specifically: Multi-scale fusion features Perform global average pooling to obtain the channel description vector.

[0065] The basic channel weights are then generated through a two-layer MLP (with ReLU activation):

[0066] in Sigmoid is used to constrain weights to... .

[0067] To introduce structural sensitivity, the mean of the high-frequency saliency plot is calculated:

[0068] This mean reflects the overall high-frequency intensity of the image and can be considered a global indicator of structural complexity. Unlike traditional SE which relies solely on gaps, this structure-driven modulation factor enables the attention mechanism to enhance channel responses in high-frequency rich regions and suppress unnecessary enhancements in flat regions.

[0069] The final modulation weights are:

[0070] in As an amplification factor, it enables the high-frequency saliency map to play a modulation role in the channel dimension.

[0071] The final channel modulation output is

[0072] in This indicates point-by-point multiplication. This design enables channel attention to not only rely on global semantic statistics but also to adaptively adjust the enhancement intensity based on the image structural complexity, thereby more effectively highlighting high-frequency sensitive channels.

[0073] Furthermore, the channel-type high-frequency modulation submodule specifically comprises: Global average pooling is performed on the multi-scale fused features to obtain the channel description vector; The channel description vectors are processed through two layers of multilayer perceptron and a sigmoid activation function to generate basic channel weights. ; Calculate the mean of the high-frequency saliency plot This yields a structural complexity index that reflects the overall high-frequency intensity of the image; The final modulation weight is :

[0074] The multi-scale fused features are multiplied point-by-point by the final modulation weights to output the channel-modulated features. .

[0075] Furthermore, the structural adaptive residual fusion submodule specifically comprises: The mean and standard deviation of the high-frequency saliency map are calculated to obtain a compact description of the image structural complexity. This compact description is then input into a lightweight multilayer perceptron, and fusion coefficients are generated using a sigmoid activation function. The intermediate features of the Real-ESRGAN backbone network are then weighted and fused with the channel-modulated features according to the fusion coefficients, and the residual fused features are output.

[0076] In the formula, The fusion coefficient is... ; The high-frequency significance plot shows the mean and standard deviation.

[0077] The structural adaptive residual fusion module is specifically as follows: For saliency map Calculate the mean and standard deviation:

[0078] The mean reflects the overall high-frequency intensity, while the standard deviation reflects the degree of texture variation. Together, they constitute a compact description of the image's structural complexity.

[0079] Then Input a lightweight MLP to generate fusion coefficients

[0080] in The Sigmoid constraint ensures that the enhancement process is stable and controllable.

[0081] The final output is

[0082] in For input features, This represents the enhanced features after channel modulation. The adaptive residual fusion mechanism effectively avoids over-enhancement or under-enhancement caused by a fixed ratio, ensuring that the enhancement intensity remains consistent with the image structural complexity, thereby improving the overall reconstruction quality and stability.

[0083] To enable the optimization objective to directly affect the high-frequency region, this method introduces high-frequency loss and structural consistency loss into the original loss system. This allows the training objective to form a gradient direction consistent with the high-frequency enhancement path of UAMHF during backpropagation. The high-frequency loss constrains the difference between the reconstructed image and the real image in the high-frequency domain through the Laplacian operator, and is defined as follows:

[0084] This loss term directly applies to the second-order gradient of the image, enabling the model to obtain explicit high-frequency gradient signals during backpropagation. Since the high-frequency saliency map construction branch of UAMHF is also based on the Laplacian operator and low-frequency difference, the high-frequency loss is consistent with the saliency map construction mechanism at the gradient level, allowing the saliency map to gradually converge into a more accurate high-frequency structure indication during training, thereby improving the effectiveness of multi-scale convolution selection and channel modulation.

[0085] To further maintain the continuity and geometric stability of the local structure, this study introduces structural consistency loss, which is defined as follows:

[0086] This loss, measured jointly by three components—luminance, contrast, and structure—enables the model to enhance texture details while maintaining edge continuity and local structural consistency. Since UAMHF's structure-adaptive residual fusion module predicts the fusion coefficients using the mean and variance of the saliency map, the structural consistency loss directly influences the gradient of the fusion coefficients during backpropagation, causing the enhancement intensity to gradually converge to a balance between restoring details and suppressing noise during training.

[0087] Combining the above loss terms, the total loss function proposed in this study is:

[0088] Compared with the original loss system, this loss function not only explicitly introduces constraints of the high-frequency domain and the structural domain into the optimization objective, but also forms a close collaboration with the four sub-modules of UAMHF in gradient propagation: the high-frequency saliency map obtains a more accurate structural response under the guidance of the high-frequency loss, the multi-scale convolution avoids unstable enhancement under the constraint of the structural consistency loss, the channel modulation maintains structural coherence under the action of the SSIM gradient, and the residual fusion achieves adaptive convergence of enhancement intensity under the combined influence of the high-frequency and structural gradients.

[0089] During the training of the super-resolution model, the Adam optimizer was used with an initial learning rate of 1×10⁻⁶. -4 With β1=0.9, β2=0.99, weight decay of 0, total training iterations of 400,000, batch size of 12, and the MultiStepLR learning rate scheduling strategy, the learning rate is decayed by 50% at the 400,000th iteration, and an exponential moving average is introduced to the generator parameters with a decay factor of 0.999. The generator and discriminator are optimized alternately at a 1:1 update frequency.

[0090] Step 3: Input the high-resolution reconstructed image into the growth point detection model and output the growth point detection results; Furthermore, the growth point detection model uses YOLOv8n as the base model, employs the SGD optimizer for parameter updates, with an initial learning rate of 0.0001, momentum of 0.95, batch size of 8, and training epochs of 250. The growth point detection model filters out low-confidence predictions by setting a confidence threshold of 0.25 during the inference phase, and then uses non-maximum suppression with an intersection-union ratio threshold of 0.45 to remove duplicate responses that are too close in spatial location, thus obtaining the growth point detection results.

[0091] The Labelme software was used to annotate soybean growth points on the high-resolution images after super-resolution reconstruction, resulting in 12,500 valid points and generating YOLO format annotation files. YOLOv8n was used as the base model, employing the SGD optimizer with an initial learning rate of 0.0001, momentum of 0.95, batch size of 8, and training epochs of 250, without any additional data augmentation. During the inference phase, the image size was adapted to 640×640, and low-confidence predictions and duplicate responses were removed using a 0.25 confidence threshold and a 0.45 IoU threshold for non-maximum suppression, yielding the growth point detection results.

[0092] Step 4: Remove false detection boxes with confidence levels below the set threshold from the growth point detection results, count the number of valid growth points, and calculate the germination rate based on the number of valid growth points and the theoretical number of seedlings.

[0093] Furthermore, in step four, the set threshold is 0.5; the seedling emergence rate is calculated as follows: seedling emergence rate = (number of effective growth points detected / theoretical number of seedlings) × 100%.

[0094] The super-resolution reconstructed images of the experimental field were input into the optimized YOLOv8 model for batch inference. False detection boxes with a confidence level below 0.5 were removed, resulting in 158,235 valid growth points. Using the formula: emergence rate = (number of valid growth points / theoretical number of seedlings) × 100%, the soybean emergence rate in this experimental field was calculated to be: (158,235 / 166,500) × 100% = 95.03%. The seedling emergence rate of the experimental field was 94.87% when manually sampled. The relative error between the results obtained by this method and the manual results was only 0.17%. Furthermore, the aerial photography of 100 mu of field took only 25 minutes, which is 6 times more efficient than traditional low-altitude aerial photography. This effectively solves the problems of low resolution and poor detection accuracy of high-altitude images, while avoiding missed detections and false detections caused by dense overlapping seedlings.

[0095] The above provides a detailed description of the soybean growth point detection seedling rate method based on the fusion of super-resolution algorithms proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A soybean growth point detection emergence rate method using a fusion super-resolution algorithm, characterized by, Includes the following steps: Step 1: Collect soybean seedling images and preprocess them to obtain a standardized image dataset; Step two: input the standardized image dataset into the trained super-resolution model for super-resolution reconstruction, and output high-resolution reconstructed images; wherein the super-resolution model is based on Real ESRGAN architecture, and a unified adaptive multi-branch high-frequency enhancement module is introduced. Step 3: Input the high-resolution reconstructed image into the growth point detection model and output the growth point detection results; Step 4: Remove false detection boxes with confidence levels below the set threshold from the growth point detection results, count the number of valid growth points, and calculate the germination rate based on the number of valid growth points and the theoretical number of seedlings.

2. The method of claim 1, wherein, In step one, The acquisition of soybean seedling images specifically involves using a drone equipped with an RGB camera to take aerial photos during the soybean emergence period. The aerial photography altitude is 30m and the flight speed is 3m / s. The preprocessing includes cropping, stitching, and coordinate calibration of the images acquired by the UAV to obtain a standardized image set with a resolution of 2048×2048 pixels; and dividing the standardized image dataset into a training set, a validation set, and a test set in a ratio of 8:1:

1.

3. The method of claim 1, wherein, In step two, Unified adaptive multi-branch high-frequency enhancement module embedded in real Bottleneck position between ESRGAN backbone network and up-sampling module The unified adaptive multi-branch high-frequency enhancement module includes a high-frequency saliency map construction submodule, a multi-scale convolutional fusion submodule, a channel-based high-frequency modulation submodule, and a structure-adaptive residual fusion submodule, which are used to perform fine modeling of high-frequency information in three dimensions: space, channel, and structure.

4. The method of claim 3, wherein, The high-frequency saliency map construction submodule is specifically as follows: The Laplacian operator is used to convolve each channel of the input features to extract the local high-frequency Laplacian response; in, L It is a 3×3 Laplace operator; X For Real Intermediate features of the ESRGAN backbone network; Low-frequency reconstruction is constructed using downsampling and upsampling operations. X low And obtain the high-frequency residual through difference. R ; The high-frequency response of the Laplace and the high-frequency residual are concatenated and then fused to obtain the comprehensive high-frequency map. The absolute value of the comprehensive high-frequency map is normalized sample by sample to obtain a high-frequency saliency map, which serves as the structural guidance signal for subsequent multi-scale convolution and channel modulation. wherein is a very small constant, .

5. The method of claim 3, wherein, The multi-scale convolutional fusion submodule is specifically as follows: Set the multi-scale convolution kernel set to 3, 5, and 7, and perform convolution extraction on the input features respectively; The convolutional outputs of each branch are stacked along the new dimension after group normalization and ReLU activation. The high-frequency saliency map is input into a 1×1 convolution and then normalized by Softmax to generate dynamic weights; The dynamic weight is fused with the stacked multi-scale features to obtain multi-scale fusion features Z msf .

6. The method of claim 3, wherein, The channel-type high-frequency modulation submodule is specifically: Global average pooling is performed on the multi-scale fused features to obtain the channel description vector; The channel description vectors are processed through two layers of multilayer perceptron and a sigmoid activation function to generate basic channel weights. ; averaging the high frequency saliency map to obtain a structure complexity index reflecting the overall high frequency intensity of the image; The final modulation weight is : The multi-scale fusion feature is point-wise multiplied with the final modulation weight, and a channel-modulated feature is output .

7. The method of claim 3, wherein, The structural adaptive residual fusion submodule is specifically as follows: The mean and standard deviation of the high-frequency saliency map are calculated to obtain a compact description of the image structural complexity; this compact description is input into a lightweight multilayer perceptron, and fusion coefficients are generated using a sigmoid activation function; Real The intermediate features of the ESRGAN backbone network are weighted and fused with the channel-modulated features according to the fusion coefficients, and the residual fused features are output. In the formula, The fusion coefficient is... ; The high-frequency significance plot shows the mean and standard deviation.

8. The method according to claim 1, characterized in that, The loss function used during the training of the super-resolution model is: wherein, , , , , and are weight coefficients for each loss term; is a pixel loss; perceptual loss; adversarial loss; high frequency loss; structural loss; wherein, H is a high frequency extraction function; is high frequency information of the reconstructed image; is high frequency information of the real image; wherein SSIM is the structural similarity index; During the training of the super-resolution model, the Adam optimizer was used with an initial learning rate of 1×10⁻⁶. -4 With β1=0.9, β2=0.99, weight decay of 0, total training iterations of 400,000, batch size of 12, and the MultiStepLR learning rate scheduling strategy, the learning rate is decayed by 50% at the 400,000th iteration, and an exponential moving average is introduced to the generator parameters with a decay factor of 0.

999. The generator and discriminator are optimized alternately at a 1:1 update frequency.

9. The method according to claim 1, characterized in that, The growth point detection model uses YOLOv8n as the base model and SGD optimizer for parameter updates. The initial learning rate is 0.0001, the momentum is 0.95, the batch size is 8, and the number of training rounds is 250. The growth point detection model filters out low-confidence predictions by setting a confidence threshold of 0.25 during the inference phase, and then uses non-maximum suppression with an intersection-union ratio threshold of 0.45 to remove duplicate responses that are too close in spatial location, thus obtaining the growth point detection results.

10. The method of claim 1, wherein, In step four, The set threshold is 0.5; the seedling emergence rate is calculated as follows: seedling emergence rate = (number of effective growth points detected / theoretical number of seedlings) × 100%.

Images