A super-resolution reconstruction method for UAV images of kiwifruit sunburn

Abstract

This invention relates to a super-resolution reconstruction method for UAV images of kiwifruit sunburn. Compared with existing technologies, it solves the problems of poor model generalization due to the lack of accurate modeling of the actual physical degradation of UAVs, and the difficulty of feature extraction networks in simultaneously considering global semantics and local high-frequency lesion details, resulting in blurred reconstruction. The invention includes the following steps: acquisition and preprocessing of UAV images of kiwifruit sunburn; generation of multi-scale paired datasets; construction of a super-resolution reconstruction model for kiwifruit sunburn images; training of the super-resolution reconstruction model for kiwifruit sunburn images; acquisition and preprocessing of UAV images; and super-resolution reconstruction of UAV images. This invention significantly eliminates boundary artifacts in image processing by constructing a multi-scale attention and spatial-frequency fusion mechanism, accurately restoring the high-frequency edge texture details of small lesions at different scales, and reducing data acquisition time.

Description

Technical Field

[0001] This invention relates to the field of image super-resolution reconstruction technology, specifically a super-resolution reconstruction method for UAV images of kiwifruit sunburn. Background Technology

[0002] Kiwifruit is highly susceptible to sunburn disease under high temperatures and strong sunlight in summer, leading to peel necrosis and flesh softening, severely impacting fruit quality and yield. Currently, using unmanned aerial vehicles (UAVs) for low-altitude remote sensing to acquire high-resolution images, combined with deep learning object detection models, for orchard phenotypic monitoring and early disease identification has become an important development direction in the field of precision agriculture.

[0003] However, in actual drone monitoring of kiwifruit sunburn disease, there is an inherent contradiction between "cruise monitoring efficiency" and "image spatial resolution." To accurately identify the details of tiny sunburn spots, drones typically need to fly at extremely low altitudes. This not only significantly increases the safety risks of flight operations in complex orchard terrain but also severely limits the coverage area and operational efficiency of a single drone flight. When drones increase their operating altitude to pursue large-area, rapid cruising, the acquired drone images inevitably suffer severe visual degradation due to the complex physical environment of the real air-to-ground imaging link. This degradation is not a simple bicubic down-interpolation sampling under laboratory conditions but rather a complex combination of multiple levels of physical degradation, including anisotropic motion blur caused by high-frequency vibrations of the flight platform, nonlinear phase distortion caused by atmospheric turbulence, sensor thermal noise, and lossy JPEG compression during downlink transmission. These combined degradation factors result in severe loss of high-frequency edges and texture details of the spots, making it difficult for downstream target detection systems to achieve high-precision disease identification and damage assessment.

[0004] In recent years, deep learning-based image super-resolution reconstruction technology has been widely studied and attempted to be applied in the field of remote sensing. However, existing super-resolution reconstruction methods generally suffer from the following significant drawbacks when directly applied to agricultural UAV imagery: First, there is the problem of constructing paired LR-HR datasets from UAV remote sensing images. Existing studies often use ideal bicubic downsampling to synthesize training data, ignoring the optical blur, sensor thermal noise, and compression artifacts that are prevalent in real UAV imaging, resulting in the model's inability to recover true details when processing real orchard images. Second, there is the problem of difficulty in recovering key features in complex scenes, leading to poor reconstruction results. Kiwi orchards are typical unstructured and complex environments, with backgrounds filled with interfering information such as foliage occlusion, shadow variations, and dead leaves similar in color to the fruit. When processing such complex scenes, models are prone to misjudging high-frequency noise in the background as texture details for enhancement, or excessively smoothing out subtle sunscald lesion features. Third, there is a lack of systematic adaptation research for multi-scale scenes. Current research on UAV remote sensing image super-resolution mainly focuses on image restoration at a single specific altitude, lacking a systematic comparison and evaluation of algorithm performance at different flight altitudes. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies, such as the lack of accurate modeling for the real physical degradation of UAVs leading to poor model generalization, and the difficulty of feature extraction networks in taking into account both global semantics and local high-frequency lesion details, resulting in blurred reconstruction. This invention provides a super-resolution reconstruction method for UAV images of sunburnt kiwifruit to solve the above problems.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A super-resolution reconstruction method for UAV imagery of sunburnt kiwifruit includes the following steps:

[0008] Acquisition and preprocessing of UAV images of sunburn on kiwifruit;

[0009] Generation of multi-scale paired datasets: A high-fidelity LR-HR paired dataset is constructed based on the physical perception multi-order degradation model of the real imaging link, and multi-height cross-scale affine transformation and degradation simulation is performed based on continuous random scale factors.

[0010] Constructing a super-resolution reconstruction model for sunburn images of kiwifruit;

[0011] Training of a super-resolution reconstruction model for sunburn images of kiwifruit;

[0012] Acquisition and preprocessing of UAV imagery;

[0013] Super-resolution reconstruction of UAV imagery: The preprocessed low-resolution image of the UAV to be detected is input into the trained super-resolution reconstruction model of kiwifruit sunburn imagery, and a forward inference process is performed. The model uses the learned network weights to sequentially perform shallow feature extraction, deep spatial-frequency joint feature extraction and lesion texture enhancement on the input image, as well as shallow and deep feature fusion and sub-pixel spatial resolution magnification. Finally, the reconstructed high-resolution image of kiwifruit sunburn is output. The high-resolution image restores the high-frequency edges and detailed textures of the small lesions that are blurred due to high-altitude physical degradation.

[0014] The acquisition and preprocessing of the kiwifruit sunburn drone images includes the following steps:

[0015] Visible light images of trellis-type kiwifruit orchards were collected using a multi-rotor drone at four typical operating heights of 3m, 5m, 7m, and 9m. The flight parameters were set as follows: flight speed 3m / s, forward overlap 80%, lateral overlap 70%, and the collection time was limited to 10:00 to 12:00 on a sunny day.

[0016] The acquired images were subjected to standardized preprocessing, including radiometric correction, geometric correction, dehazing, and uniform cropping, to obtain a standardized image set with a size of 512×512 pixels. The images acquired at a height of 3m were used as the high-resolution ground truth (HR) set, and the images acquired at heights of 5-9m were used as the low-resolution validation set (LR).

[0017] The generation of the multi-scale pairing dataset includes the following steps:

[0018] A physical perception-based UAV degradation model (PA-UDM) is constructed that strictly corresponds to the actual physical process of UAV imaging. This model includes the following three degradation operations:

[0019] Realistic Blur Kernel Convolution Degradation: Estimating the scene-adaptive blur kernel k(x) from real 9m-height UAV images using KernelGAN unsupervised method, and performing discrete convolution. ,in, This represents the image after blurring and degradation. Indicates the scene-adaptive fuzzy kernel. This represents a two-dimensional discrete convolution operation. This represents the input image after multi-scale affine transformation; the blur kernel integrates three types of degradation components: anisotropic motion blur caused by flight platform vibration, lens defocus circle of confusion, and atmospheric turbulence nonlinear phase distortion.

[0020] Poisson-Gaussian heterogeneous mixed noise injection: to Injection satisfies Heterogeneous mixed noise, where This indicates heterogeneous mixed noise. Represents the Poisson distribution function. Let be the expected value of the Poisson distribution. This indicates that the mean is 0 and the standard deviation is 0. The Gaussian distribution function; where the Poisson component simulates photon shot noise, and the Gaussian component N(0,σ2) simulates sensor readout thermal noise, σ∈[5,15];

[0021] Adaptive JPEG lossy compression: The compression quality factor q∼U(60,85) is uniformly and randomly sampled to obtain the final result. ;in, This represents the final low-resolution image output. This represents the lossy compression function in JPEG. Let U(60,85) represent the compression quality factor, and U(60,85) represent a uniform random distribution in the interval between 60 and 85. This is the image data after noise processing;

[0022] PA-UDM degradation was applied to HR images at a height of 3m to generate simulated LR images with degradation characteristics consistent with real high-altitude images, and a high-fidelity paired LR-HR dataset was constructed.

[0023] A continuous random scaling factor δ∼U(0.2,0.8) is introduced to perform a multi-scale affine transformation on the HR image. Simulates cross-scale image degradation of UAVs at arbitrary flight altitudes of 3-15m; among them, Let U(0.2, 0.8) represent a continuous random scaling factor, where U(0.2, 0.8) represents a uniform random distribution in the interval between 0.2 and 0.8. This represents the scaled image output after multi-scale affine transformation. Represents a high-resolution ground truth image. Indicates scaling operation;

[0024] The scale-transformed images are input into the physical perception UAV degradation model to perform third-order physical degradation, generating a multi-scale paired dataset covering all operational altitudes.

[0025] The construction of the super-resolution reconstruction model for sunburned kiwi fruit images includes the following steps:

[0026] The super-resolution reconstruction model for sunburn images of kiwifruit is designed to include a shallow feature extraction module, a deep feature extraction module, and a high-quality image reconstruction module.

[0027] Configure a shallow feature extraction module;

[0028] A feature mapping layer containing 3×3 convolutional kernels is used to map the input low-resolution UAV image from the RGB color space to a high-dimensional feature space. A non-linear mapping relationship is established to extract shallow local edges and color features of the image, and a shallow feature tensor is output. ;

[0029] The shallow feature tensor It is divided into two paths: one path is input to the deep feature extraction module, and the other path is directly transmitted to the input end of the high-quality image reconstruction module through a long-span residual connection, so as to realize the reuse of low-frequency information and efficient backpropagation of gradient;

[0030] Configure a deep feature extraction module;

[0031] The deep feature extraction module is used to capture the global semantics and local high-frequency texture of the image. Internally, it cascades a multi-scale residual mixing attention group and a space-frequency joint block; the specific structure is as follows:

[0032] The multi-scale residual hybrid attention group consists of K serially stacked hybrid attention modules and overlapping cross attention modules cascaded with them;

[0033] Each hybrid attention module employs a parallel dual-branch structure. The input features are first processed by layer normalization and then divided into two parallel inputs: the global branch uses a shift-window multi-head self-attention module to capture global spatial dependencies; the local branch sequentially performs convolution, activation, and pooling operations to generate a high-dimensional channel weight matrix. This weight matrix is ​​then multiplied element-wise with the original input features before layer normalization to suppress background noise and enhance key lesion features. The results of the two branches are added together and supplemented with a local residual structure before being output to the overlapping cross-attention module. When performing self-attention window division, the overlapping cross-attention module introduces boundary overlap and constrains its sliding window stride to be strictly smaller than the window size. This allows pixels at the boundary to participate in the attention calculation of multiple adjacent windows simultaneously, eliminating feature fragmentation and boundary artifacts caused by conventional independent window division and achieving continuous interaction of spatial feature information.

[0034] The spatial-frequency joint block is located at the end of the deep feature extraction layer and adopts a dual-branch structure for cross-domain fusion: one branch extracts high-frequency edge signals in the frequency domain through convolution and real-valued 2D Fast Fourier Transform, while the other branch extracts spatial local texture features through continuous convolution. The outputs of the two branches are concatenated in the channel dimension and then fused through 1×1 convolution for dimensionality reduction, finally outputting a deep feature tensor. ;

[0035] Configure a high-quality image reconstruction module.

[0036] The high-quality image reconstruction module includes a feature fusion layer and a subpixel upsampling layer;

[0037] First, the shallow feature tensor passed from the shallow feature extraction module The deep feature tensor output by the deep feature extraction module The input is fed into the feature fusion layer and added element by element: By constructing a fusion feature tensor Achieving the aggregation of shallow low-frequency contours and deep high-frequency details; subsequently, the aggregated fused feature tensor... The input is fed into a subpixel upsampling layer, where the channel-dimensional features are rearranged to the spatial dimension through a periodic shuffling operation, thereby amplifying the spatial resolution. Finally, a 3×3 reconstruction convolutional layer maps the feature map back to the RGB channels, outputting a reconstructed high-resolution image of sunburnt kiwifruit.

[0038] The training of the super-resolution reconstruction model for sunburned kiwi fruit images includes the following steps:

[0039] Low-resolution images from multi-scale paired datasets The samples were input in batches as training samples into the initialized super-resolution reconstruction model of sunburned kiwi fruit images;

[0040] The shallow feature extraction module extracts shallow local edge and color features of the image, outputs a shallow feature tensor, and transmits it in two paths. One of the shallow feature tensors is input to the deep feature extraction module, which is processed by a multi-scale residual hybrid attention group with an embedded overlapping cross attention module to capture the global semantics of the image and realize the continuous interaction of spatial features. Then it is sent to the space-frequency joint block to extract local high-frequency textures and output a deep feature tensor.

[0041] In the high-quality image reconstruction module, the shallow feature tensor passed through another long-span residual connection is added and fused with the deep feature tensor element by element. Then, the spatial resolution is amplified by the sub-pixel upsampling layer, and the reconstructed high-resolution image prediction value is mapped and output by the reconstruction convolutional layer.

[0042] Construct a joint loss function to calculate the predicted values ​​of the reconstructed high-resolution image and the corresponding high-resolution ground truth images in the multi-scale paired dataset. Reconstruction error between;

[0043] The error gradient of the joint loss function is calculated using the backpropagation algorithm, and the network weight parameters in the shallow feature extraction module, deep feature extraction module, and high-quality image reconstruction module are updated using the optimizer. The above forward and backpropagation processes are iterated repeatedly until the model loss tends to converge or the preset maximum number of iterations is reached, thus completing the model training and saving the optimal parameter weights.

[0044] A computer-readable storage medium storing a computer program that, when executed by a processor, enables a super-resolution reconstruction method for UAV images of kiwifruit sunburn.

[0045] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, which, when executed by the processor, enables a super-resolution reconstruction method for UAV images of kiwifruit sunburn.

[0046] Beneficial effects

[0047] This invention presents a super-resolution reconstruction method for UAV imagery of kiwifruit sunscald. Compared with existing technologies, it overcomes the problem of poor generalization of traditional super-resolution models in realistic high-altitude complex physical degradation scenarios. By constructing a multi-order degradation model of a physically perceptive UAV and combining it with continuous stochastic multi-scale transformation, it not only realistically simulates complex physical degradation but also achieves adaptation to target scale changes at different flight altitudes, effectively improving the reconstruction capability for cross-scale targets. Simultaneously, based on the traditional Transformer architecture, by constructing a multi-scale attention and space-frequency fusion mechanism, it significantly eliminates boundary artifacts in image processing, accurately restores the high-frequency edge texture details of tiny lesions at different scales, and reduces data acquisition time. Attached Figure Description

[0048] Figure 1 This is a sequence diagram of the method of the present invention;

[0049] Figure 2 This is a schematic diagram of the super-resolution reconstruction model for sunburn images of kiwifruit involved in this invention.

[0050] Figure 3 This is a diagram of the multi-scale change strategy and degradation model proposed in this invention;

[0051] Figure 4 This is a comparison chart of the reconstruction effects of the present invention and other existing methods. Detailed Implementation

[0052] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0053] like Figure 1 As shown, the super-resolution reconstruction method for UAV images of kiwifruit sunburn according to the present invention includes the following steps:

[0054] The first step is the acquisition and preprocessing of drone images of sunburn on kiwifruit.

[0055] (1) Use a multi-rotor UAV to collect visible light images of trellis kiwi orchards at typical operating heights of 3m, 5m, 7m and 9m. The flight parameters are set as follows: flight speed 3m / s, forward overlap rate 80%, lateral overlap rate 70%, and the collection time is limited to 10:00 to 12:00 am on a sunny day.

[0056] (2) Perform radiometric correction, geometric correction, dehazing and uniform cropping on the acquired images in sequence to obtain a standardized image set with a size of 512×512 pixels; among them, the images acquired at a height of 3m are used as the high-resolution ground truth (HR) and the images acquired at a height of 5-9m are used as the low-resolution validation set (LR).

[0057] The second step is the generation of a multi-scale paired dataset: a multi-level degradation model of physical perception based on the real imaging link is constructed and a high-fidelity LR-HR paired dataset is generated, and multi-height cross-scale affine transformation and degradation simulation are performed based on continuous random scale factors.

[0058] In UAV images acquired at different flight altitudes, the same target often exhibits distortion due to scale variations, making pixel-level registration of true high and low altitude images difficult. This hinders the manual selection of images with identical areas. Secondly, physical and environmental factors, such as motion blur caused by airflow and sensor thermal noise, can lead to significant differences between UAV images acquired at different times, posing a challenge to acquiring paired data during actual UAV operations in the field.

[0059] The overall topology flow for multi-scale paired dataset generation and degradation model in this step is as follows: Figure 3 As shown in the figure, this illustrates the overall topology flow of multi-scale paired dataset generation and the physical perception drone degradation model. From Figure 3 As can be seen from the diagram, the generation framework of this invention is mainly divided into two core stages: the left side is the multi-scale scaling and cropping stage, which is responsible for transforming the spatial dimension of the high-resolution ground truth image through continuous cross-scale factors to simulate the target scale changes at different operating altitudes of the UAV, and uniformly cropping it into image patches of fixed size; the right side is the physical degradation model stage, which inputs the multi-scale image patches into a set degradation pool, and performs third-order composite degradation processing on them through three degradation operators of cascaded blur kernel, noise pool and lossy compression, so that the final output at the end of the image is a high-fidelity paired dataset that is highly consistent with the real high-altitude blind degradation characteristics.

[0060] The specific implementation method is as follows:

[0061] (1) Construct a physical perception UAV degradation model PA-UDM that strictly corresponds to the actual imaging physical process of the UAV, which includes the following three-order degradation operations:

[0062] Realistic Blur Kernel Convolution Degradation: Estimating the scene-adaptive blur kernel k(x) from real 9m-height UAV images using KernelGAN unsupervised method, and performing discrete convolution. ,in, This represents the image after blurring and degradation. Indicates the scene-adaptive fuzzy kernel. This represents a two-dimensional discrete convolution operation. This represents the input image after multi-scale affine transformation; the blur kernel integrates three types of degradation components: anisotropic motion blur caused by flight platform vibration, lens defocus circle of confusion, and atmospheric turbulence nonlinear phase distortion.

[0063] Poisson-Gaussian heterogeneous mixed noise injection: to Injection satisfies Heterogeneous mixed noise, where This indicates heterogeneous mixed noise. Represents the Poisson distribution function. Let be the expected value of the Poisson distribution. This indicates that the mean is 0 and the standard deviation is 0. The Gaussian distribution function; where the Poisson component simulates photon shot noise, and the Gaussian component N(0,σ2) simulates sensor readout thermal noise, σ∈[5,15];

[0064] Adaptive JPEG lossy compression: The compression quality factor q∼U(60,85) is uniformly and randomly sampled to obtain the final result. ;in, This represents the final low-resolution image output. This represents the lossy compression function in JPEG. U(60,85) represents the compression quality factor, and U(60,85) represents a uniform random distribution in the interval between 60 and 85.

[0065] (2) Apply PA-UDM degradation to the HR image at a height of 3m to generate a simulated LR image that is consistent with the degradation characteristics of the real high-altitude image, and construct a high-fidelity paired LR-HR dataset.

[0066] (3) Introduce a continuous random scaling factor δ∼U(0.2,0.8) to perform a multi-scale affine transformation on the HR image. Simulates cross-scale image degradation of UAVs at arbitrary flight altitudes of 3-15m; among them, Let U(0.2, 0.8) represent a continuous random scaling factor, where U(0.2, 0.8) represents a uniform random distribution in the interval between 0.2 and 0.8. This represents the scaled image output after multi-scale affine transformation. Represents a high-resolution ground truth image. This indicates a scaling operation.

[0067] (4) Input the scale-transformed image into the physical perception UAV degradation model to perform third-order physical degradation and generate a multi-scale paired dataset covering all operating heights.

[0068] The third step involves constructing a super-resolution reconstruction model for kiwifruit sunburn images. This model consists of multiple defined deep feature extraction modules, which are composed of multiple cascaded multi-scale residual hybrid attention groups and a spatial-frequency joint block. The multi-scale residual hybrid attention group enhances the model's extraction of features at different depths by aggregating receptive field features from different depths, solving the problems of complex textures in UAV agricultural images and varying scales in UAV aerial images. The spatial-frequency joint block achieves local pixel correlation and captures global image information by jointly extracting features from the spatial and frequency domains, improving the model's ability to reconstruct high-frequency subtle lesions.

[0069] (1) such as Figure 2 As shown, the super-resolution reconstruction model for kiwifruit sunburn images is designed to include a shallow feature extraction module, a deep feature extraction module, and a high-quality image reconstruction module.

[0070] (2) Set up a shallow feature extraction module;

[0071] A feature mapping layer containing 3×3 convolutional kernels is used to map the input low-resolution UAV image from the RGB color space to a high-dimensional feature space. A non-linear mapping relationship is established to extract shallow local edges and color features of the image, and a shallow feature tensor is output. ;

[0072] The shallow feature tensor It is divided into two paths: one path is input to the deep feature extraction module, and the other path is directly transmitted to the input end of the high-quality image reconstruction module through a long-span residual connection, so as to realize the reuse of low-frequency information and efficient backpropagation of gradient.

[0073] (3) Set up a deep feature extraction module;

[0074] The deep feature extraction module is used to capture the global semantics and local high-frequency texture of the image. Internally, it cascades a multi-scale residual mixing attention group and a space-frequency joint block; the specific structure is as follows:

[0075] The multi-scale residual hybrid attention group consists of K serially stacked hybrid attention modules and overlapping cross attention modules cascaded with them;

[0076] Each hybrid attention module employs a parallel dual-branch structure. The input features are first processed by layer normalization and then divided into two parallel inputs: the global branch uses a shift-window multi-head self-attention module to capture global spatial dependencies; the local branch sequentially performs convolution, activation, and pooling operations to generate a high-dimensional channel weight matrix. This weight matrix is ​​then element-wise multiplied with the original input features before layer normalization to suppress background noise and enhance key lesion features. The results from the two branches are summed and supplemented with a local residual structure before being output to the overlapping cross-attention module. When dividing the self-attention window, the overlapping cross-attention module introduces boundary overlap, constraining its sliding window stride to be strictly smaller than the window size. This allows pixels at the boundary to simultaneously participate in the attention calculation of multiple adjacent windows, eliminating feature fragmentation and boundary artifacts caused by conventional independent window division, and achieving continuous interaction of spatial feature information.

[0077] The spatial-frequency joint block is located at the end of the deep feature extraction layer and adopts a dual-branch structure for cross-domain fusion: one branch extracts high-frequency edge signals in the frequency domain through convolution and real-valued 2D Fast Fourier Transform, while the other branch extracts spatial local texture features through continuous convolution. The outputs of the two branches are concatenated in the channel dimension and then fused through 1×1 convolution for dimensionality reduction, finally outputting a deep feature tensor. .

[0078] (4) Set up a high-quality image reconstruction module.

[0079] The high-quality image reconstruction module includes a feature fusion layer and a subpixel upsampling layer;

[0080] First, the shallow feature tensor passed from the shallow feature extraction module The deep feature tensor output by the deep feature extraction module The input is fed into the feature fusion layer and added element by element: By constructing a fusion feature tensor Achieving the aggregation of shallow low-frequency contours and deep high-frequency details; subsequently, the aggregated fused feature tensor... The input is fed into a subpixel upsampling layer, where the channel-dimensional features are rearranged to the spatial dimension through a periodic shuffling operation, thereby amplifying the spatial resolution. Finally, a 3×3 reconstruction convolutional layer maps the feature map back to the RGB channels, outputting a reconstructed high-resolution image of sunburnt kiwifruit.

[0081] The fourth step is training the super-resolution reconstruction model for kiwifruit sunburn images.

[0082] This step constructs an end-to-end training mechanism based on multi-scale degradation driving and joint space-frequency optimization. It is designed to address the pain point that the edges of lesions are easily lost in the high-altitude blind degradation scenario. By using a multi-objective loss function, the network is forced to learn spatial reconstruction and frequency domain texture consistency simultaneously. This significantly broadens the model's generalization boundary for unknown and complex farmland environments while ensuring high-definition restoration of tiny sunscald lesions.

[0083] (1) Extract low-resolution images from the multi-scale paired dataset The samples were input in batches as training samples into the initialized super-resolution reconstruction model of sunburned kiwi fruit images.

[0084] (2) The shallow local edge and color features of the image are extracted by the shallow feature extraction module, and the shallow feature tensor is output and transmitted in two paths. One of the shallow feature tensors is input to the deep feature extraction module and processed by the multi-scale residual hybrid attention group with embedded overlapping cross attention module to capture the global semantics of the image and realize the continuous interaction of spatial features. Then it is sent to the space-frequency joint block to extract local high-frequency texture and output the deep feature tensor.

[0085] In the high-quality image reconstruction module, the shallow feature tensor passed through another long-span residual connection is added and fused with the deep feature tensor element by element. Then, the spatial resolution is amplified by the sub-pixel upsampling layer, and the reconstructed high-resolution image prediction value is mapped and output by the reconstruction convolutional layer.

[0086] (3) Construct a joint loss function and calculate the predicted value of the reconstructed high-resolution image and the corresponding high-resolution ground truth image in the multi-scale paired dataset. Reconstruction error between;

[0087] The error gradient of the joint loss function is calculated using the backpropagation algorithm, and the network weight parameters in the shallow feature extraction module, deep feature extraction module, and high-quality image reconstruction module are updated using the optimizer. The above forward and backpropagation processes are iterated repeatedly until the model loss tends to converge or the preset maximum number of iterations is reached, thus completing the model training and saving the optimal parameter weights.

[0088] The fifth step is the acquisition and preprocessing of drone images.

[0089] Step 6, Super-resolution reconstruction of UAV imagery: The preprocessed low-resolution UAV imagery to be detected is input into the trained kiwifruit sunburn imagery super-resolution reconstruction model, and a forward inference process is performed. The model uses the learned network weights to sequentially perform shallow feature extraction, deep spatial-frequency joint feature extraction and lesion texture enhancement on the input image, as well as shallow and deep feature fusion and sub-pixel spatial resolution magnification. Finally, the reconstructed high-resolution image of kiwifruit sunburn is output. The high-resolution image restores the high-frequency edges and detailed textures of small lesions blurred by high-altitude physical degradation. This high-resolution image is then input into the downstream target detection system for high-precision identification of kiwifruit sunburn disease and orchard damage assessment.

[0090] like Figure 4 As shown, this is a comparison chart of the reconstruction effects of the method of this invention and other existing super-resolution reconstruction methods at different acquisition heights. From Figure 4 As can be seen, at three typical operating heights of 5m, 7m, and 9m, conventional interpolation methods and bicubic downsampling super-resolution models all exhibit varying degrees of edge blurring, detail loss, or artifacts when faced with complex orchard backgrounds, making it difficult to clearly distinguish subtle kiwifruit sunscald spots. In contrast, the reconstruction method proposed in this invention demonstrates optimal visual reconstruction quality at all heights. It not only significantly eliminates block boundary artifacts and maintains smooth global spatial semantics, but also extremely accurately restores the high-frequency edges and complex texture details of tiny spots. The reconstructed image is closest to the high-resolution ground truth image at a height of 3m, fully verifying the superior performance and generalization ability of this invention in cross-scale real-world blind degradation scenarios.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for super-resolution reconstruction of kiwifruit sunburn UAV images, characterized in that, Includes the following steps: 11) Acquisition and preprocessing of UAV images of sunburn on kiwifruit; 12) Generation of multi-scale paired datasets: A dataset based on a physical perception multi-order degradation model of real imaging links and high-fidelity LR-HR paired datasets are constructed, and multi-height cross-scale affine transformation and degradation simulation are performed based on continuous random scale factors. 13) Construct a super-resolution reconstruction model for sunburn images of kiwifruit; 14) Training of a super-resolution reconstruction model for sunburn images of kiwifruit; 15) Acquisition and preprocessing of UAV imagery; 16) Super-resolution reconstruction of UAV imagery: The preprocessed low-resolution imagery of the UAV to be detected is input into the trained super-resolution reconstruction model of kiwifruit sunburn imagery, and a forward inference process is performed. The model uses the learned network weights to sequentially perform shallow feature extraction, deep spatial-frequency joint feature extraction and lesion texture enhancement on the input image, as well as shallow and deep feature fusion and sub-pixel spatial resolution magnification. Finally, the reconstructed high-resolution image of kiwifruit sunburn is output. The high-resolution image restores the high-frequency edges and detailed textures of the small lesions that are blurred due to high-altitude physical degradation.

2. The method according to claim 1, wherein, The acquisition and preprocessing of the kiwifruit sunburn drone images includes the following steps: 21) Use a multi-rotor UAV to collect visible light images of trellis-type kiwifruit orchards at four typical operating heights of 3m, 5m, 7m and 9m. The flight parameters are set as follows: flight speed 3m / s, forward overlap rate 80%, lateral overlap rate 70%, and the collection time is limited to 10:00 to 12:00 on a sunny day. 22) Perform radiometric correction, geometric correction, dehazing and uniform cropping on the acquired images in sequence to obtain a standardized image set with a size of 512×512 pixels; the images acquired at a height of 3m are used as the high-resolution ground truth (HR) and the images acquired at a height of 5-9m are used as the low-resolution validation set (LR).

3. The method of claim 1, wherein, The generation of the multi-scale pairing dataset includes the following steps: 31) Construct a physical perception UAV degradation model PA-UDM that strictly corresponds to the actual physical process of UAV imaging, which includes the following three-order degradation operations: Realistic blur kernel convolution degradation: estimating scene-adaptive blur kernel k(x) from real 9m-height UAV images in a KernelGAN unsupervised manner to perform discrete convolution wherein, denotes the blurred image, denotes the scene-adaptive blur kernel, denotes the two-dimensional discrete convolution operation, denotes the input image after multi-scale affine transformation; the blur kernel fuses three types of degradation components: anisotropic motion blur caused by flight platform vibration, lens defocus dispersion circle, and nonlinear phase distortion of atmospheric turbulence; Poisson-Gaussian heterogeneous mixed noise injection: to Injection satisfies Heterogeneous mixed noise, where This indicates heterogeneous mixed noise. Represents the Poisson distribution function. Let be the expected value of the Poisson distribution. This indicates that the mean is 0 and the standard deviation is 0. The Gaussian distribution function; where the Poisson component simulates photon shot noise, and the Gaussian component N(0,σ2) simulates sensor readout thermal noise, σ∈[5,15]; Adaptive JPEG lossy compression: The compression quality factor q∼U(60,85) is uniformly and randomly sampled to obtain the final result. ;in, This represents the final low-resolution image output. This represents the lossy compression function in JPEG. Let U(60,85) represent the compression quality factor, and U(60,85) represent a uniform random distribution in the interval between 60 and 85. This refers to image data after noise processing. 32) Apply PA-UDM degradation to HR images at a height of 3m to generate simulated LR images with degradation characteristics consistent with real high-altitude images, and construct a high-fidelity paired LR-HR dataset; 33) Introduce a continuous random scaling factor δ∼U(0.2,0.8) to perform a multi-scale affine transformation on the HR image. Simulates cross-scale image degradation of UAVs at arbitrary flight altitudes of 3-15m; among them, Let U(0.2, 0.8) represent a continuous random scaling factor, where U(0.2, 0.8) represents a uniform random distribution in the interval between 0.2 and 0.

8. This represents the scaled image output after multi-scale affine transformation. Represents a high-resolution ground truth image. Indicates scaling operation; 34) Input the scale-transformed image into the physical perception UAV degradation model to perform third-order physical degradation, generating a multi-scale paired dataset covering all operational altitudes.

4. The super-resolution reconstruction method for UAV images of kiwifruit sunburn according to claim 1, characterized in that, The construction of the super-resolution reconstruction model for sunburned kiwi fruit images includes the following steps: 41) The super-resolution reconstruction model for sunburn images of kiwifruit is designed to include a shallow feature extraction module, a deep feature extraction module, and a high-quality image reconstruction module; 42) Set up a shallow feature extraction module; A feature mapping layer containing 3×3 convolutional kernels is used to map the input low-resolution UAV image from the RGB color space to a high-dimensional feature space. A non-linear mapping relationship is established to extract shallow local edges and color features of the image, and a shallow feature tensor is output. ; The shallow feature tensor It is divided into two paths: one path is input to the deep feature extraction module, and the other path is directly transmitted to the input end of the high-quality image reconstruction module through a long-span residual connection, so as to realize the reuse of low-frequency information and efficient backpropagation of gradient; 43) Set up the deep feature extraction module; The deep feature extraction module is used to capture the global semantics and local high-frequency texture of the image. Internally, it cascades a multi-scale residual mixing attention group and a space-frequency joint block; the specific structure is as follows: The multi-scale residual hybrid attention group consists of K serially stacked hybrid attention modules and overlapping cross attention modules cascaded with them; Each hybrid attention module adopts a parallel dual-branch structure. The input features are first processed by layer normalization and then divided into two parallel inputs: the global branch uses a shift window multi-head self-attention module to capture global spatial dependencies; the local branch generates a high-dimensional channel weight matrix by sequentially performing convolution, activation and pooling operations. This weight matrix is ​​then multiplied element-wise with the original input features before layer normalization to suppress background noise and enhance key lesion features. The results of the two branches are added together and then output to the overlapping cross attention module with a local residual structure. When performing self-attention window division, the overlapping cross attention module introduces the boundary overlap part and constrains the step size of its sliding window to be strictly smaller than the window size, so that the pixels at the boundary participate in the attention calculation of multiple adjacent windows at the same time, eliminating the feature discontinuity and boundary artifacts caused by conventional independent window division, and realizing the continuous interaction of spatial feature information. The spatial-frequency joint block is located at the end of the deep feature extraction layer and adopts a dual-branch structure for cross-domain fusion: one branch extracts high-frequency edge signals in the frequency domain through convolution and real-valued 2D Fast Fourier Transform, while the other branch extracts spatial local texture features through continuous convolution. The outputs of the two branches are concatenated in the channel dimension and then fused through 1×1 convolution for dimensionality reduction, finally outputting a deep feature tensor. ; 44) Configure the high-quality image reconstruction module. The high-quality image reconstruction module includes a feature fusion layer and a subpixel upsampling layer; First, the shallow feature tensor passed from the shallow feature extraction module The deep feature tensor output by the deep feature extraction module The input is fed into the feature fusion layer and added element by element: By constructing a fusion feature tensor Achieving the aggregation of shallow low-frequency contours and deep high-frequency details; subsequently, the aggregated fused feature tensor... The input is fed into a subpixel upsampling layer, where the channel-dimensional features are rearranged to the spatial dimension through a periodic shuffling operation, thereby amplifying the spatial resolution. Finally, a 3×3 reconstruction convolutional layer maps the feature map back to the RGB channels, outputting a reconstructed high-resolution image of sunburnt kiwifruit.

5. The super-resolution reconstruction method for UAV images of kiwifruit sunburn according to claim 1, characterized in that, The training of the super-resolution reconstruction model for sunburned kiwi fruit images includes the following steps: 51) Low-resolution images from a multi-scale paired dataset The samples were input in batches as training samples into the initialized super-resolution reconstruction model of sunburned kiwi fruit images; 52) The shallow feature extraction module extracts the shallow local edge and color features of the image, outputs the shallow feature tensor and transmits it in two paths; one of the shallow feature tensors is input to the deep feature extraction module, and processed by the multi-scale residual hybrid attention group with embedded overlapping cross attention module to capture the global semantics of the image and realize the continuous interaction of spatial features. Then it is sent to the space-frequency joint block to extract local high-frequency texture and output the deep feature tensor. In the high-quality image reconstruction module, the shallow feature tensor passed through another long-span residual connection is added and fused with the deep feature tensor element by element. Then, the spatial resolution is amplified by the sub-pixel upsampling layer, and the reconstructed high-resolution image prediction value is mapped and output by the reconstruction convolutional layer. 53) Construct a joint loss function to calculate the predicted value of the reconstructed high-resolution image and the corresponding high-resolution ground truth image in the multi-scale paired dataset. Reconstruction error between; The error gradient of the joint loss function is calculated using the backpropagation algorithm, and the network weight parameters in the shallow feature extraction module, deep feature extraction module, and high-quality image reconstruction module are updated using the optimizer. The above forward and backpropagation processes are iterated repeatedly until the model loss tends to converge or the preset maximum number of iterations is reached, thus completing the model training and saving the optimal parameter weights.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, enables the super-resolution reconstruction method for UAV images of kiwifruit sunburn as described in any one of claims 1-5.

7. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can implement the super-resolution reconstruction method for UAV images of kiwifruit sunburn as described in any one of claims 1-5.

Images