A UAV remote sensing monitoring method for tea diseases based on super-resolution and detection network

Abstract

This invention relates to a drone-based remote sensing monitoring method for tea diseases using super-resolution and detection networks, comprising the following steps: Data acquisition: First, low-altitude images of tea leaves are acquired, with the drone flying at an altitude of 2m to 4m, serving as the training set for the network; Data processing: The cropped low-altitude images are labeled using LabelImg software; then, the low-altitude images are divided into training and validation sets for the network at an 8:2 ratio; Network training: First, a super-resolution network model is trained, including sequentially training a PSNR model and a GAN model; then, a detection network model is trained; Super-resolution and detection: The drone flies at an altitude of 12m, acquiring high-altitude images of tea leaves, which are then fed into the trained super-resolution network for super-resolution to restore detail information. The super-resolution images are then cropped and fed into the trained detection network for tea disease detection. This invention can save drone flight costs and time, helping tea farmers to detect tea diseases in tea gardens in real time.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a UAV remote sensing monitoring method for tea diseases based on super-resolution and detection networks. Background Technology

[0002] With the development of imaging and image processing technologies, methods for detecting crop diseases can be broadly categorized into two directions: traditional machine learning methods and convolutional neural network methods. Traditional machine learning methods often require manually designed features, but these handcrafted features lack sufficient representational power, making it difficult to fully extract discriminative feature information from crop disease images. Furthermore, traditional machine learning models for crop disease identification exhibit poor robustness and are unsuitable for detecting small-target crop diseases over large areas in complex backgrounds.

[0003] Compared to traditional machine learning methods, convolutional neural networks (CNNs) can automatically extract features for crop disease identification, possessing high feature description and model generalization capabilities, and significantly reducing the cost of manual intervention. From a data perspective, CNNs require a large amount of data for training, placing high demands on both the quality and quantity of data. For agricultural disease detection applications, data sources include camera or mobile phone images and drone footage. Data acquisition and processing require significant human resources. While cameras and mobile phones can capture high-resolution images, the process is cumbersome and prone to focus issues. Drones offer faster image acquisition, but for small targets like tea leaves, the drone's altitude cannot be too high to obtain high-quality data.

[0004] Most existing methods for detecting tea leaf lesions are based on high-resolution tea leaf images, which are not suitable for detecting low-resolution tea leaf images taken by drones at high altitudes. From a network model perspective, detection networks are becoming increasingly lightweight to achieve practical applications. Currently, the smallest model among mainstream detection networks, such as YOLOv5s in the YOLOv5 series, is around 15MB. However, this model is still not lightweight enough. How to further reduce the model's size while maintaining detection accuracy is a hot topic. Summary of the Invention

[0005] (1) Technical problems to be solved

[0006] This invention proposes a UAV remote sensing monitoring method for tea diseases based on super-resolution and detection networks. The method includes data acquisition, data processing, network training, super-resolution, and detection. This invention can save UAV flight costs and time, and help tea farmers monitor tea diseases in tea gardens in real time.

[0007] (2) Technical solution

[0008] An embodiment of the present invention proposes a UAV remote sensing monitoring method for tea diseases based on super-resolution and detection networks, comprising the following steps:

[0009] Data acquisition: First, low-altitude images of tea leaves were acquired, with the drone flying at an altitude of 2m to 4m; then, high-altitude images of tea leaves were acquired, with the drone flying at an altitude of about 12m. All tea leaf images had a resolution of 8000*6000.

[0010] Data processing: First, the 13 low-altitude images were cropped into 1300 images with a resolution of 800*600; second, the cropped low-altitude images were labeled using LabelImg software; then, the low-altitude images were divided into the training set and validation set of the network in an 8:2 ratio.

[0011] Network training: First, train the super-resolution network model, including training the PSNR model and the GAN model sequentially; then train the detection network model.

[0012] Super-resolution and detection: The drone images of tea leaves taken at a height of 12m are fed into a trained super-resolution network to perform super-resolution and restore detailed information. The super-resolution images are then cropped and fed into a trained detection network for tea disease detection.

[0013] Furthermore, the super-resolution network model is a Real-ESRGAN network model, which consists of a generator and a discriminator. The generator is composed of dense residual blocks stacked together, and the discriminator is a U-Net network with regularization terms.

[0014] Furthermore, before training the super-resolution network model, the following steps are taken: constructing the network dataset. Specifically, the prepared low-altitude tea leaf dataset is subjected to two degradation blurring processes. The first degradation process includes Gaussian blur, downsampling, noise addition, and JPEG compression. The second degradation process includes Gaussian blur, downsampling, and noise addition.

[0015] Furthermore, the specific steps for training the super-resolution network model include:

[0016] A PSPN-guided generator network is trained based on the minimum absolute error to obtain an optimal model;

[0017] The network parameters are initialized using the obtained optimal model, and the network model is further trained using a discriminator. This results in the final model of the super-resolution network in this paper.

[0018] Furthermore, the detection network model is improved based on the YOLOv5s network. ShuffleNetV2 residual blocks are cascaded as the backbone of the network. The convolutional part in the neck of the YOLOv5s network is replaced with depthwise separable convolutions, and the original CIoU loss function is replaced with SIoU loss function.

[0019] Furthermore, the ShuffleNetV2 architecture includes two modes. In the first mode, after the feature map input, there is a channel splitting operation. This operation divides the feature map with c input channels into two branches. The left branch does not perform any operation, while the right branch contains three convolution operations. The two 1*1 convolutions have been replaced by ordinary convolutions instead of the grouped convolutions in ShuffleNetV1. Finally, the data in the channels of these two branches are merged using the Concat and ChannelShuffle operations. The first mode does not have a channel splitting operation.

[0020] Furthermore, the depthwise separable convolution consists of two parts: channel-wise convolution and point-wise convolution.

[0021] Furthermore, the detection network model uses a feature pyramid structure when performing target detection. The feature pyramid structure employs upsampling and convolution operations from bottom to top, undergoing two 2x upsampling and two convolution processes.

[0022] Furthermore, the SIoU loss function includes: diagonal loss, distance loss, shape loss, and IoU loss.

[0023] Furthermore, the diagonal loss is defined as:

[0024]

[0025] Where c h Let σ be the height difference between the center points of the ground truth bounding box and the predicted bounding box, and let σ be the distance between the centers of the ground truth bounding box and the predicted bounding box. Equal to angle α;

[0026]

[0027]

[0028]

[0029] The coordinates of the center of the true bounding box. For the coordinates of the predicted bounding box center, it can be noted that when α is When the angle is 0, the angle loss is 0. During training, if... Then minimize α; otherwise minimize β.

[0030] The distance loss is defined as:

[0031]

[0032] in:

[0033]

[0034] Here (c w ,c h () represents the width and height of the smallest bounding rectangle of the ground truth bounding box and the predicted bounding box;

[0035] The shape loss is defined as:

[0036]

[0037] in:

[0038]

[0039] (w,h) and (w) gt ,h gt ) are the width and height of the predicted bounding box and the ground truth bounding box, respectively. θ controls the degree of attention to shape loss. In order to avoid focusing too much on shape loss and reducing the movement of the predicted bounding box, the range of the θ parameter is specified as [2,6].

[0040] The IoU loss is defined as:

[0041]

[0042] In summary, the final SIoU loss function is defined as follows:

[0043]

[0044] (3) Beneficial effects

[0045] This invention presents a two-stage network. The first stage uses a GAN network to super-resolution images of tea leaves taken by a drone, while the second stage utilizes an improved Yolov5s algorithm to detect tea leaf lesions. The GAN network's generator consists of cascaded dense residual modules, and the discriminator uses a U-Net network with spectral normalization regularization. The detection network's backbone is composed of the lightweight ShuffleNetV2 module, trained using the SIoU loss function. The trained super-resolution detection model can recover low-resolution tea leaf images captured by a drone at a flight altitude of at least 12m and detect small-target lesions. This model saves on drone flight costs and time, helping tea farmers monitor tea diseases in tea gardens in real time.

[0046] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a method for remote sensing monitoring of tea diseases using unmanned aerial vehicles according to an embodiment of the present invention.

[0049] Figure 2 This is a diagram of the Real-ESRGAN super-resolution network structure in one embodiment of the present invention.

[0050] Figure 3 This is a network structure diagram of a super-resolution network generator in one embodiment of the present invention.

[0051] Figure 4 This is a schematic diagram of a detection network in one embodiment of the present invention.

[0052] Figure 5 This is a network structure diagram of the detection network backbone using ShuffleNetV2 structure blocks according to an embodiment of the present invention.

[0053] Figure 6 This is a diagram of a depthwise separable convolutional network structure according to an embodiment of the present invention.

[0054] Figure 7 This is a network structure diagram of a characteristic pyramid structure according to an embodiment of the present invention.

[0055] Figure 8 These are original drone images taken at a height of 12m using the method described in this embodiment of the invention, along with a partial view thereof.

[0056] Figure 9 This is a super-resolution overall structure diagram and a partial display of the original drone image taken at a height of 12m using the method of this embodiment of the invention.

[0057] Figures 10-12 This is a comparison image of the super-resolution result and the original image in an embodiment of the present invention. Detailed Implementation

[0058] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following detailed description of the embodiments and the accompanying drawings are used to illustrate the principles of the present invention by way of example, but should not be used to limit the scope of the present invention. That is, the present invention is not limited to the described embodiments, and any modifications, substitutions and improvements to the parts, components and connection methods are covered without departing from the spirit of the present invention.

[0059] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0060] The following will refer to the appendix. Figure 1 -Appendix Figure 3 This application will be described in detail with reference to the embodiments.

[0061] An embodiment of the present invention proposes a UAV remote sensing monitoring method for tea diseases based on super-resolution and detection networks, comprising the following steps:

[0062] Data acquisition: First, low-altitude images of tea leaves were acquired, with the drone flying at an altitude of 2m to 4m; then, high-altitude images of tea leaves were acquired, with the drone flying at an altitude of about 12m. All tea leaf images had a resolution of 8000*6000.

[0063] Data processing: First, the low-altitude images were cropped into 100 images with a resolution of 800*600; second, the cropped low-altitude images were labeled using LabelImg software; then, the low-altitude images were divided into the training set and validation set of the network in an 8:2 ratio.

[0064] Network training: First, train the super-resolution network model, including training the PSNR model and the GAN model sequentially; then train the detection network model.

[0065] Super-resolution and detection: The drone images of tea leaves taken from high altitude are fed into a trained super-resolution network to perform super-resolution and restore detailed information. The super-resolution images are then cropped and fed into a trained detection network for tea disease detection.

[0066] Specifically, in one embodiment of the present invention, the super-resolution network model is a Real-ESRGAN network model, which consists of a generator and a discriminator. The generator is an ESRGAN network generator, and the discriminator is an improved U-Net network.

[0067] Specifically, in another embodiment of the present invention, before training the super-resolution network model, the method further includes: constructing a network dataset, specifically including: performing two downgrading blurring processes on the prepared low-altitude tea leaf dataset, the first downgrading process including Gaussian blur, downsampling, noise addition and JPEG compression, and the second downgrading process including Gaussian blur, downsampling and noise addition.

[0068] Specifically, in one embodiment of the present invention, the specific steps for training the super-resolution network model include: training a PSPN-guided network based on the minimum absolute error to obtain an optimal model; initializing the network parameters using the obtained optimal model; and training the network model to obtain the final model of the super-resolution network in this paper.

[0069] Specifically, in one embodiment of the present invention, the detection network model is improved based on the YOLOv5s network, using ShuffleNetV2 residual block concatenation as the backbone of the network, replacing the convolutional part in the neck part of the YOLOv5s network with depthwise separable convolution, and replacing the original CIoU loss function with the SIoU loss function.

[0070] Specifically, in one embodiment of the present invention, the ShuffleNetV2 structure includes two modes. In the first mode, after the feature map is input, there is a channel splitting operation. This operation divides the feature map with c input channels into two branches. The left branch does not perform any operation, and the right branch contains three convolution operations. The two 1*1 convolutions have been replaced by ordinary convolutions instead of the grouped convolutions in ShuffleNetV1. Finally, the data in the channels of these two branches are merged using Concat and Channel Shuffle operations. There is no channel splitting operation in the first mode.

[0071] Specifically, in one embodiment of the present invention, the depthwise separable convolution consists of two parts: channel-wise convolution and point-wise convolution.

[0072] Specifically, in one embodiment of the present invention, the detection network model uses a feature pyramid structure when performing target detection. The feature pyramid structure employs upsampling and convolution operations from bottom to top, undergoing two double upsampling and two convolution processes.

[0073] Specifically, in one embodiment of the present invention, the SIoU loss function includes: diagonal loss, distance loss, shape loss, and IoU loss.

[0074] Specifically, in one embodiment of the present invention, the diagonal loss is defined as:

[0075]

[0076] Where c h Let σ be the height difference between the center points of the ground truth bounding box and the predicted bounding box, and let σ be the distance between the centers of the ground truth bounding box and the predicted bounding box. Equal to angle α;

[0077]

[0078]

[0079]

[0080] The coordinates of the center of the true bounding box. For the coordinates of the predicted bounding box center, it can be noted that when α is When the angle is 0, the angle loss is 0. During training, if... Then minimize α; otherwise minimize β.

[0081] The distance loss is defined as:

[0082]

[0083] in:

[0084]

[0085] Here (c w ,c h () represents the width and height of the smallest bounding rectangle of the ground truth bounding box and the predicted bounding box;

[0086] The shape loss is defined as:

[0087]

[0088] in:

[0089]

[0090] (w,h) and (w) gt ,h gt ) are the width and height of the predicted bounding box and the ground truth bounding box, respectively. θ controls the degree of attention to shape loss. In order to avoid focusing too much on shape loss and reducing the movement of the predicted bounding box, the range of the θ parameter is specified as [2,6].

[0091] The IoU loss is defined as:

[0092]

[0093] In summary, the final SIoU loss function is defined as follows:

[0094]

[0095] The following example illustrates the UAV remote sensing monitoring method for tea diseases based on super-resolution and detection networks, according to another embodiment of the present invention. (See attached document.) Figure 1 As shown in the figure, the UAV remote sensing monitoring method for tea diseases based on super-resolution and detection networks in this invention includes:

[0096] 1. Data Acquisition: First, low-altitude images of tea leaves were acquired, with the drone flying at an altitude of 2m to 4m; then, high-altitude images of tea leaves were acquired, with the drone flying at an altitude of approximately 12m (all tea leaf images have a resolution of 8000*6000).

[0097] 2. Data processing: (1) Image cropping: Each 8000*6000 image is cropped to 800*600, that is, one image is cropped into 100 images. (2) Data annotation: LabelImg software is used to annotate the cropped low-altitude images. (3) Training set division: The low-altitude images are divided into the network's training set and validation set in an 8:2 ratio.

[0098] 3. Network training: (1) Training the super-resolution network model: The training of the super-resolution model is divided into two steps. First, train the PSNR model, and then train the GAN model. (2) Training the detection network model.

[0099] 4. Super-resolution and detection: First, the aerial images of tea leaves taken by the drone are fed into a trained super-resolution network to restore details. Then, the super-resolution images are cropped and fed into a trained detection network for tea disease detection.

[0100] Specifically, see the appendix. Figure 2 As shown, in this embodiment, the super-resolution network adopts the Real-ESRGAN network, which consists of two parts: a generator and a discriminator. The generator is the generator of the ESRGAN network, and the discriminator is the improved U-Net network.

[0101] The generator is mainly composed of stacked RRDB (Residual-in-Residual Dense Block) blocks, which are composed of dense blocks. This structure combines the ideas of multi-level residual networks and dense connections. The specific structure is as follows: Figure 3As shown, the generator of this network is a very deep network. It has been proven that the deeper the network, the better the receptive field and the greater the amount of information it can utilize from the surrounding environment. However, deeper networks often have more complex nonlinear mappings and more complex calculations. The role of dense blocks is to avoid information loss during forward and backward propagation. For x4 super-resolution, the network performs the generator exactly as in ESRGAN; while for x2 and x1 super-resolution, the network first performs pixel-unshuffle (the inverse operation of pixel-shuffle, which can be understood as enlarging the image size by compressing image channels) to reduce the image resolution, thereby expanding the number of image channels. Then, the processed image is fed into the network for super-resolution reconstruction.

[0102] The RRDB module employs a dense residual structure with no BN layers. Each RRDB block consists of three DenseBlocks with no BN layers. Removing BN layers has been proven to improve performance and reduce computational complexity in various PSNR-oriented tasks, including super-resolution and deblurring. BN layers normalize features using the mean and variance of a batch of data during training and use the estimated mean and variance of the entire training set during testing. When the statistical differences between the training and test sets are large, BN layers tend to introduce unpleasant artifacts and limit generalization ability. We have empirically observed that BN layers are more likely to introduce artifacts when the network is deep and trained on a GAN architecture. These artifacts sometimes appear in the middle of iterations and under different settings, violating the requirement for stable performance during training. Therefore, BN layers are removed for stable training and consistent performance. Furthermore, removing BN layers helps improve generalization ability and reduce computational complexity and memory usage.

[0103] The discriminator is a U-Net network with spectral standard regularization, meaning that spectral standard regularization is added to the last layer of the U-Net network. The U-Net discriminator can judge the authenticity of each generated pixel at the pixel level, which can ensure the overall realism of the generated image while focusing on the details.

[0104] Super-resolution network training steps: This embodiment of the invention uses a GAN network for image super-resolution. The working principle of a GAN network is as follows: First, the generator learns the features of the input image and outputs a fake image similar to the input image. Then, the fake image is input into the discriminator, which judges the similarity between the fake image and the original image and feeds back the difference between the fake image and the original image to the generator. The generator will adjust and update according to the feedback. The entire process will be continuously looped under the guidance of the network loss function until the network model is saturated.

[0105] Network dataset construction: The prepared low-altitude photographed tea leaf dataset undergoes two downgrading blurring processes. Each downgrading process includes Gaussian blur, downsampling, noise addition, and JPEG compression, but the second downgrading process does not include JPEG compression. (Corresponding to...) Figure 2 In the network input, x1 represents the original image, x2 represents the image after the first downgrade process, and x4 represents the image obtained after the second downgrade process.

[0106] Network Training: The training of the super-resolution network in this paper consists of two steps: First, a PSPN-guided network is trained using L1 Loss (Minimum Absolute Error) to obtain an optimal model. Second, the network is trained based on the model obtained in the first step. The parameters of the network are initialized using the model obtained in the first step. The loss function in the second step is a combination of L1 Loss, perceptual loss, and GAN loss. The network model trained in the second step is the final model of the super-resolution network in this paper.

[0107] Detection Network: In this embodiment of the invention, the detection network is an improvement on the YOLOv5s baseline network. ShuffleNetV2 residual blocks are concatenated as the backbone of the network. At the same time, depthwise separable convolutions replace the convolutional parts in the neck of the original YOLOv5s network, and the SIoU loss function replaces the original CIoU loss function.

[0108] (1) ShufflleNetV2 architecture

[0109] In this embodiment of the invention, the backbone of the detection network uses ShuffleNetV2 structural blocks for feature extraction. The ShuffleNetV2 series of networks can be used for classification, and its backbone is responsible for feature extraction. In this embodiment, the ShuffleNetV2 series of networks is used instead of CSPdarknet53 in YOLOv5s for feature extraction, which reduces the number of network model parameters while maintaining accuracy. ShuffleNetV2 is an upgraded version based on ShuffleNetV1, and it achieves higher accuracy than ShuffleNetV1 and MobileNetV2 at the same complexity. The basic building blocks of this model can be roughly divided into two types, the first being as shown in the attached figure. Figure 5As shown in the left part, after the feature map input, there is a channel split operation. This operation divides the feature map with c input channels into two branches. The left branch does not perform any operation, while the right branch contains three convolution operations. The two 1x1 convolutions, previously grouped convolutions in ShuffleNetV1, are replaced with regular convolutions. Finally, the data from these two branches are merged using Concat and Channel Shuffle operations. This not only ensures that the input and output channels of the basic module are the same but also avoids the Add operation, speeding up the model's inference. Finally, the ChannelShuffle operation is performed. It is worth noting that... Figure 5 The left side of the module no longer has channel separation operation, so the number of output channels of this basic module is twice the number of input channels, and the operation process of its left and right branches is basically the same.

[0110] (2) Depthwise separable convolution

[0111] Using depthwise separable convolutions instead of regular convolutions can reduce network parameters; see appendix. Figure 6 As shown, this embodiment of the invention uses a large number of depthwise separable convolutions to reduce the weight of the detection network. Depthwise separable convolution mainly consists of two processes: a combination of channel-wise convolution (DW) and pointwise convolution (PW). Compared to conventional convolution operations, it has a lower number of parameters and lower computational cost. In DW convolution, one kernel is responsible for one channel, and each channel is convolved by only one kernel. The number of channels in the feature map generated by this process is exactly the same as the number of channels in the input kernel. The number of feature maps after DW convolution is the same as the number of channels in the input layer, which cannot expand the feature maps. Moreover, this approach performs convolution operations independently on each channel of the input layer, failing to effectively utilize the feature information of different channels at the same spatial location. Therefore, PW convolution is needed to combine these feature maps to generate new feature maps. The operation of PW convolution is very similar to that of conventional convolution, with a kernel size of 1*1*M, where M is the number of channels in the previous layer. Therefore, the convolution operation here will weight and combine the feature maps from the previous step in the depth direction to generate new feature maps. The number of output feature maps is equal to the number of convolution kernels.

[0112] (3) Characteristic pyramid structure

[0113] In target detection, feature pyramids are typically used to detect targets of different scales. A feature pyramid structure includes a bottom-up path and a top-down path. In this embodiment of the invention, upsampling and convolution operations are employed in the bottom-up process of the feature pyramid, such as... Figure 7As shown, the feature fusion process involves two 2x upsampling operations (the feature map is magnified twice) and two convolution processes. To fully utilize the image information obtained from the feature extraction part, the feature fusion part of this paper performs multiple fusions on the feature maps from the feature extraction part. In the bottom-up process, after two upsampling operations, concatenation operations are performed with the shallow and middle-layer feature maps from the feature extraction part, respectively. Before upsampling, a convolution operation is performed on the feature map. After the concatenation operation, the number of channels is adjusted using DWConv convolution to ensure that the number of input and output channels is consistent in the bottom-up process. In the top-down process, fusion is performed with the middle-layer and deep features from the feature extraction part, respectively. To reduce the number of parameters, only DWConv is used in the top-down process, and Conv is no longer used. Before the concatenation operation, the DWConv convolution kernel size is 3*3 with a stride of 2, which performs 2x downsampling on the feature map. After the concatenation operation, the DWConv convolution kernel size is 3*3 with a stride of 1, which adjusts the number of channels to ensure that the number of input and output channels is consistent.

[0114] (4) SIou loss function

[0115] Regarding IoU loss functions, such as GIoU, DIoU, and CIoU, they do not take into account the orientation between the ground truth bounding box and the predicted bounding box, resulting in slow convergence. In response, SIoU introduces the vector angle between the ground truth bounding box and the predicted bounding box and redefines the relevant loss function, which specifically includes four parts: diagonal loss, distance loss, shape loss, and IoU loss.

[0116] a. Angle cost, defined as follows:

[0117]

[0118] Where c h Let σ be the height difference between the center points of the ground truth bounding box and the predicted bounding box, and let σ be the distance between the centers of the ground truth bounding box and the predicted bounding box. It equals angle α.

[0119]

[0120]

[0121]

[0122] The coordinates of the center of the true bounding box. For the coordinates of the predicted bounding box center, it can be noted that when α is When the angle is 0, the angle loss is 0. During training, if... If so, minimize α; otherwise, minimize β.

[0123] b. Distance loss

[0124]

[0125] in:

[0126]

[0127] Here (c w ,c h () represents the width and height of the smallest bounding rectangle of the ground truth bounding box and the predicted bounding box.

[0128] c. Shape loss

[0129]

[0130] in:

[0131]

[0132] (w,h) and (w) gt ,h gt ) are the width and height of the predicted bounding box and the ground truth bounding box, respectively. θ controls the degree of attention paid to shape loss. In order to avoid paying too much attention to shape loss and reducing the movement of the predicted bounding box, the range of the θ parameter is specified as [2,6].

[0133] d. IoU loss

[0134]

[0135] In summary, the final SIoU loss function is defined as follows:

[0136]

[0137] Experimental results of embodiments of the present invention

[0138] 1. Over-scoring results

[0139] like Figure 8 The image shown is an original picture of a tea garden taken by a drone at a flight altitude of 12m (resolution 8000*6000). The portion of the image cropped out with a red box has a resolution of 800*600. Due to the high shooting altitude, the tea leaf image lacks detail, which is not conducive to lesion detection. Therefore, a super-resolution network is needed to restore the detail information. The large image was cropped into 100 smaller images with a resolution of 800*600 and then fed into a trained super-resolution network for 1x and 4x super-resolution respectively. The super-resolution results are shown in Figure 9. Figure 10 , 11 This shows a comparison between the super-resolution image and the original image.

[0140] 2. Test Results

[0141] To improve detection accuracy, the x4 super-resolution result (in which the lesion is relatively large compared to the pixel size, which is beneficial for target detection) is used as the input to the detection network. The detection effect is as follows: Figure 12 Table 1 shows the comparison information between the detection network in this paper and YOLOv5s. As shown in the table, the detection network in this paper is much smaller than YOLOv5s, and the accuracy only decreases slightly.

[0142] Table 1. Performance Comparison between the Improved Network and the Original Yolov5s Network

[0143] network P R mAP SIZE Yolov5s 0.767 0.746 0.806 14.4M Network structure improvement 0.735 0.732 0.778 0.7M Structural improvement + SIoU loss function 0.743 0.765 0.794 0.7M

[0144] This invention presents a two-stage network. The first stage uses a GAN network to super-resolution images of tea leaves taken by a drone, while the second stage utilizes an improved Yolov5s algorithm to detect tea leaf lesions. The GAN network's generator consists of cascaded dense residual modules, and the discriminator uses a U-Net network with spectral normalization regularization. The detection network's backbone is composed of the lightweight ShuffleNetV2 module, trained using the SIoU loss function. The trained super-resolution detection model can recover low-resolution tea leaf images captured by a drone at a flight altitude of at least 12m and detect small-target lesions. This model saves on drone flight costs and time, helping tea farmers monitor tea diseases in tea gardens in real time.

[0145] Existing agricultural disease monitoring technologies target samples in simple background environments. This invention addresses the detection of tea lesions in natural settings. In natural environments, tea lesions are similar in color to the soil, making them difficult to distinguish. Tea leaves are easily affected by sunlight, leading to overexposure; exposed leaves appear whitish, and this color closely matches the lesion color, further complicating the differentiation. Furthermore, wind-induced swaying of the tea bushes during photography can cause ghosting in the images. The detection network in this invention can effectively distinguish between reflective leaves and lesioned leaves, while the super-resolution network can mitigate ghosting to some extent.

[0146] Using drones for photography covers a large area, resulting in a very small pixel ratio for tea lesions in the image, which is detrimental to target detection. Furthermore, when drones shoot from high altitudes, the already small tea lesions become even smaller, making them difficult for the human eye to discern. The super-resolution network in this invention can improve the resolution of tea images and restore their detailed information.

[0147] Most existing technologies are based on detection networks to detect diseases. This paper integrates super-resolution networks and detection networks to perform 4x super-resolution on tea leaf images taken from high altitudes, thereby solving the problem of difficult detection of small targets at low resolution.

[0148] The super-resolution network used in this embodiment of the invention can effectively recover low-resolution tea leaf images, and x4 super-resolution can effectively improve the accuracy of disease detection. The detection network proposed in this embodiment is more lightweight than the original Yolov5s network, and the accuracy decreases by only 1.2%, which has virtually no impact on the detection effect. Furthermore, it introduces the SIoU loss function, which performs better than the CIoU loss function used in Yolov5s. This embodiment combines the super-resolution network with the detection network to achieve disease detection in low-resolution tea leaf images, helping UAVs collect effective data at higher flight altitudes and reducing UAV flight costs. The experimental data in this embodiment are tea leaf images in natural scenes, and the data is greatly affected by the environment; the network proposed in this embodiment can adapt well to the natural environment.

[0149] For embodiments of the method, please refer to the relevant descriptions of the device embodiments. The present invention is not limited to the specific steps and structures described above and shown in the figures. Furthermore, for the sake of brevity, detailed descriptions of known method techniques are omitted herein.

[0150] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art without departing from the scope of the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks, characterized in that, Includes the following steps: Data acquisition: First, low-altitude images of tea leaves were acquired, with the drone flying at an altitude of 2m to 4m; then, high-altitude images of tea leaves were acquired, with the drone flying at an altitude of about 12m. All tea leaf images had a resolution of 8000*6000. Data processing: First, the low-altitude images were cropped into 100 images with a resolution of 800*600; second, the cropped low-altitude images were labeled using LabelImg software; then, the low-altitude images were divided into the training set and validation set of the network in an 8:2 ratio. Network training: First, train the super-resolution network model, including training the PSNR model and the GAN model sequentially; then train the detection network model. Super-resolution and detection: The drone images of tea leaves taken from high altitude are fed into a trained super-resolution network to perform super-resolution and restore detailed information. The super-resolution images are then cropped and fed into a trained detection network for tea disease detection. The super-resolution network model is a Real-ESRGAN network model, which consists of a generator and a discriminator. The generator is composed of dense residual structure blocks stacked together, and the discriminator is a U-Net network with regularization terms. The detection network model is an improvement on the YOLOv5s baseline network. ShuffleNetV2 residual blocks are concatenated as the backbone of the network. Depthwise separable convolutions replace the convolutional parts in the neck of the YOLOv5s network. The SIoU loss function replaces the original CIoU loss function. The ShuffleNetV2 architecture includes two modes. In the first mode, after the feature map input, there is a ChannelSpilt operation, which divides the input feature map with c channels into two branches. The left branch does not perform any operation, while the right branch contains three convolution operations. The two 1*1 convolutions have been replaced with ordinary convolutions instead of the grouped convolutions in ShuffleNetV1. Finally, the data in the channels of these two branches are merged using Concat and Channel Shuffle operations. The first mode does not have a channel separation operation.

2. The method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks according to claim 1, characterized in that, Before training the super-resolution network model, the following steps are required: constructing the network dataset. Specifically, the prepared low-altitude tea leaf dataset is subjected to two degradation blurring processes. The first degradation process includes Gaussian blur, downsampling, noise addition, and JPEG compression. The second degradation process includes Gaussian blur, downsampling, and noise addition.

3. A method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks according to claim 1 or 2, characterized in that, The specific steps for training a super-resolution network model include: A PSPN-guided network is trained based on the minimum absolute error to obtain an optimal generator model. The obtained generator model is used to initialize the parameters of the network, and then the super-resolution network is trained in combination with the discriminator network. The trained network model is the final model of the super-resolution network in this paper.

4. The method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks according to claim 1, characterized in that, The depthwise separable convolution consists of two parts: channel-wise convolution and point-wise convolution.

5. The method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks according to claim 1, characterized in that, The detection network model uses a feature pyramid structure when performing feature fusion. The feature pyramid structure employs upsampling and convolution operations from bottom to top, and undergoes two double upsampling and two convolution processes. Using a feature pyramid structure to fuse the feature maps of the network's feature extraction part can retain more depth information of the network, which is beneficial for the network to detect small targets.

6. The method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks according to claim 1, characterized in that, The SIoU loss function includes: diagonal loss, distance loss, shape loss, and IoU loss.

7. A method for UAV remote sensing monitoring of tea diseases based on super-resolution and detection networks according to claim 6, characterized in that, The diagonal loss is defined as: (1) in The height difference between the center points of the ground truth bounding box and the predicted bounding box. The distance between the centers of the ground truth bounding box and the predicted bounding box. equal to angle ; (2) (3) (4) The coordinates of the center of the true bounding box. For the predicted box center coordinates, it can be noted that when for When the angle is 0, the angle loss is 0. During training, if... Then minimize Otherwise minimize ; The distance loss is defined as: (5) in: (6) Here The width and height of the minimum bounding rectangle between the ground truth bounding box and the predicted bounding box; The shape loss is defined as: (7) in: (8) and These are the width and height of the predicted bounding box and the ground truth bounding box, respectively. To control the degree of attention given to shape loss, and to avoid reducing the movement of the prediction box due to excessive focus on shape loss, the following rules are specified. The parameter range is [2,6]; The IoU loss is defined as: (9) In summary, the final SIoU loss function is defined as follows: (10) 。

Images