Natural Background Detection Method for Citrus Huanglongbing Based on Improved Cascade RCNN Network

Abstract

The present invention discloses a method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network, including: extracting features of the citrus image to be detected by using the backbone network improved from the Cascade RCNN network, and at least one layer in the improved backbone network adopts deformable convolution; using the region feature extraction network improved from the Cascade RCNN network to perform high-level semantic feature extraction on the features output by the backbone network, and the improved region feature extraction network includes a bottom-up convolution fusion process; using the cascade detector of the Cascade RCNN network to perform a multi-stage cascade object recognition and detection process on the high-level semantic features output by the region feature extraction network, and obtaining the detection result of citrus huanglongbing in the citrus image to be detected. The present invention effectively improves the accuracy of detecting citrus huanglongbing.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network. Background Art

[0002] Most of the disease characteristics of citrus huanglongbing are manifested through its leaves. Traditional prevention and control of citrus huanglongbing is that experienced fruit farmers make corresponding disease judgments by observing the leaves of citrus. However, this work not only takes a lot of time, but also is prone to misjudgment by manual judgment. In recent years, with the rapid development of computer vision and deep learning, convolutional neural networks have made great progress in the classification and detection of crop disease images. Using convolutional neural networks to detect citrus huanglongbing is a feasible solution.

[0003] A large number of existing algorithms have been widely applied in the agricultural field, including the detection of citrus huanglongbing. However, there are still problems in the detection of citrus huanglongbing in the existing technology, mainly including:

[0004] (1) Since the images are taken in natural scenes, there are a large number of interfering factors such as healthy leaves, other diseased leaves, and weeds in the image background, resulting in low detection accuracy of existing algorithms;

[0005] (2) The mutual occlusion between leaves and the distance lead to large changes in the morphology and size of huanglongbing leaves, and it is very easy to miss small-sized targets;

[0006] (3) The color and texture features of huanglongbing leaves are very similar to those of other citrus diseases, and there is an easy problem of false detection. Summary of the Invention

[0007] In view of the above problems existing in the prior art, the present invention provides a method, system, device and computer-readable storage medium for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network, which effectively improves the detection accuracy of citrus huanglongbing. The technical solution includes:

[0008] In the first aspect, a method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network is provided, including the following steps:

[0009] Obtain a citrus image to be detected;

[0010] Extract features from the citrus image to be detected by using the backbone network improved by the Cascade RCNN network, and at least one layer in the improved backbone network adopts deformable convolution;

[0011] The improved region feature extraction network using the Cascade RCNN network extracts high-level semantic features from the features output by the backbone network. The improved region feature extraction network includes a bottom-up convolutional fusion process added to the right side during the top-down fusion process of the FPN network;

[0012] The cascade detector of the Cascade RCNN network performs a multi-stage cascade object recognition and detection process on the high-level semantic features output by the region feature extraction network to obtain the detection result of huanglongbing for the citrus image to be detected.

[0013] In a possible implementation manner, the backbone network is implemented based on ResNetXt101.

[0014] In a possible implementation manner, based on the concept of the DCNv2 model, the backbone network adopts deformable convolutions for the last 3 layers of the backbone network.

[0015] In a possible implementation manner, the region feature extraction network is improved based on the FPN network and includes:

[0016] Image pyramid of the bottom-up process: formed based on the multi-scale feature maps output by the last 4 layers of the backbone network;

[0017] Top-down fusion process: The fused feature pyramid is denoted as C4, C5, C6, and C7 from top to bottom respectively;

[0018] Bottom-up convolutional fusion process: C7 is used as the starting feature map P1, and during the bottom-up convolutional process, after each layer is convolved, it is fused with the corresponding layer in the left top-down process to form the upper layer feature map.

[0019] In a possible implementation manner, during the bottom-up convolutional process, the convolutional parameters use 3×3 convolution followed by 2-fold downsampling.

[0020] In a possible implementation manner, in the cascade detector of the Cascade RCNN network, the three detection IOU values are 0.5, 0.6, 0.7 in sequence or 0.6, 0.7, 0.8 in sequence.

[0021] In a possible implementation manner, before detecting the citrus image to be detected, it further includes:

[0022] The process of training the improved Cascade RCNN network based on the huanglongbing citrus sample image set. The sample image set includes multiple sample images and is divided into a training set and a validation set;

[0023] The process of training the improved Cascade RCNN network based on the citrus huanglongbing sample image set further includes:

[0024] Using the methods of cutmix splicing and CLAHE to perform data augmentation on the sample image set.

[0025] In a second aspect, a natural background citrus huanglongbing detection system based on an improved Cascade RCNN network is provided, including:

[0026] An image to be detected acquisition module, configured to acquire a citrus image to be detected;

[0027] A feature extraction module, which uses the backbone network improved by the Cascade RCNN network to extract features from the citrus image to be detected, and at least one subsequent layer in the improved backbone network uses deformable convolution;

[0028] A feature bidirectional fusion module, which uses the region feature extraction network improved by the Cascade RCNN network to perform high-level semantic feature extraction on the features output by the backbone network. The improved region feature extraction network includes a bottom-up convolutional fusion process added to the right side during the top-down fusion process of the FPN network;

[0029] A huanglongbing detection module, which uses the cascade detector of the Cascade RCNN network to perform a multi-stage cascaded object recognition and detection process on the high-level semantic features output by the region feature extraction network, and obtains the huanglongbing detection result of the citrus image to be detected.

[0030] In a third aspect, a citrus huanglongbing detection device is provided, and the device includes:

[0031] A processor;

[0032] A memory for storing instructions executable by the processor;

[0033] Wherein, the processor realizes the method described in the first aspect above by running the executable instructions.

[0034] In a fourth aspect, a computer-readable storage medium is provided, on which computer instructions are stored, and characterized in that when the instructions are executed by a processor, the steps of the method described in the first aspect above are realized.

[0035] An improved Cascade RCNN network-based natural background citrus huanglongbing detection method of the present invention has the following beneficial effects: Aiming at the large changes in the morphology and size of huanglongbing-infected leaves caused by mutual occlusion between leaves and the distance, which easily leads to the missed detection of small-sized targets, and the color and texture features of huanglongbing-infected leaves are very similar to those of other citrus diseases, which is prone to false detection problems. In this application, deformable convolutions are used for at least one layer in the backbone network, and at the same time, a convolution-based bottom-up fusion process is newly added to the regional feature extraction network. By adding deformable convolutions in the backbone network, it adapts to the geometric deformation of huanglongbing-infected citrus leaves, adaptively changes local sampling points according to the shape of huanglongbing-infected leaves, enhances the network's ability to model the geometric deformation of the target, effectively extracts target features and improves the detection effect of citrus huanglongbing. At the same time, high-level semantic feature extraction is performed based on the FPN network in the regional feature extraction network. On the basis of the FPN network, a bottom-up convolution fusion process is added. Using the bidirectional fusion process can not only enrich the position information of each layer of feature maps, but also retain the semantic information of each layer of feature maps, enhancing the model's multi-scale information fusion ability, which is beneficial to the detection of small-sized and occluded targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of the improved Cascade RCNN network-based natural background citrus huanglongbing detection method in an embodiment of the present application;

[0037] Figure 2 is a structural diagram of the improved Cascade RCNN network in an embodiment of the present application;

[0038] Figure 3 Structural diagram of the improved Cascade RCNN network-based natural background citrus huanglongbing detection system in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be construed as limiting the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0040] As Figure 1 and 2 shown, an embodiment of the present application provides an improved Cascade RCNN network-based natural background citrus huanglongbing detection method, including the following steps:

[0041] Obtain the citrus image to be detected;

[0042] The backbone network improved by the Cascade RCNN network is used to extract features from the citrus image to be detected, and at least one layer in the improved backbone network adopts deformable convolution;

[0043] The region feature extraction network improved by the Cascade RCNN network is used to extract high-level semantic features from the features output by the backbone network. The improved region feature extraction network includes a bottom-up convolutional fusion process added to the right side during the top-down fusion process of the FPN network;

[0044] The cascade detector of the Cascade RCNN network is used to perform a multi-stage cascade target recognition and detection process on the high-level semantic features output by the region feature extraction network, and obtain the detection result of huanglongbing for the citrus image to be detected.

[0045] In the embodiment of the present application, the image to be detected is first input into the improved backbone network for feature extraction, and then the extracted feature map is input into the bidirectional fusion feature pyramid for multi-scale fusion of features. Finally, the fused feature map is input into the cascade detector to obtain the classification score and the detection box.

[0046] Aiming at the problems that the morphological and size changes of huanglongbing leaves are large due to the mutual occlusion between leaves and the distance, which easily leads to the missed detection of small-size targets, and the color and texture features of huanglongbing leaves are very similar to those of other citrus diseases, which easily leads to misdetection. In the present application, by adopting deformable convolution for at least one layer in the backbone network, and at the same time adding a bottom-up fusion process based on convolution to the region feature extraction network. By adding deformable convolution in the backbone network, it adapts to the geometric deformation of huanglongbing leaves, adaptively changes the local sampling points according to the shape of huanglongbing leaves, enhances the network's ability to model the geometric deformation of the target, effectively extracts the target features and improves the detection effect of citrus huanglongbing. At the same time, high-level semantic feature extraction is performed based on the FPN network in the region feature extraction network. On the basis of the FPN network, a bottom-up convolutional fusion process is added. By using the bidirectional fusion process, it can not only enrich the position information of each layer of the feature map, but also retain the semantic information of each layer of the feature map, enhance the multi-scale information fusion ability of the model, and is beneficial to the detection of small-size and occluded targets.

[0047] In addition, by using the cascade detector of the Cascade RCNN network, the purpose of continuously optimizing the prediction result is achieved by cascading several detection networks with different IOU thresholds.

[0048] Furthermore, the above backbone network is implemented based on ResNetXt101.

[0049] The backbone network can be MobileNet V1, MobileNet V2, Vgg16, DenseNet, etc. In the embodiments of this application, the ResNet series of models are adopted, preferably ResNet101, and further preferably ResNetXt101. The essential difference between ResNetXt101 and ResNet101 is the adoption of channel grouping. Taking Conv2 as an example, ResNet101 has 64 1×1 convolutional kernels, while ResNetXt101 has 32 groups, with each group having four 1×1 convolutional kernels. Compared with ResNet101, ResNetXt101 has a simpler network structure, is more modular, has fewer hyperparameters that need to be manually adjusted, and has better feature extraction capabilities.

[0050] Furthermore, based on the concept of the DCNv2 model, deformable convolutions are applied to the last three layers of the backbone network.

[0051] In the embodiments of this application, the backbone network not only has at least one layer of deformable convolution. The basic idea of deformable convolution is to use sampling with offsets to replace fixed-position sampling, and adjust the size and position of the deformable convolution kernel according to the learned offsets. The sampling points of the convolution kernels at different positions will be adaptively adjusted according to the position and size of the detection target, so as to adapt to the geometric deformation of citrus huanglongbing leaves. In this application, for the setting of the distribution of the number of deformable convolution layers, at least one layer can be selected from any layer in the backbone network to adopt deformable convolution, and the results after adding deformable convolution at different positions are compared to obtain the optimal setting. In this application, drawing on the concept of the DCNv2 model, deformable convolutions are applied to the last three layers of the backbone network, and during the deformable convolution process, not only the offset parameters of the sampling points are learned, but also a weight coefficient is added to each sampling point to adaptively change the local sampling points in the detection of citrus huanglongbing, so as to obtain better detection results. Further, the settings of the DCNv2 model can be directly used. The backbone network adopts the ResNetXt101 network, and the deformable convolution calculations of Conv3 - Conv5 in the backbone network of DCNv2 are adopted in the Conv3, Conv4, and Conv5 convolutional layers.

[0052] The above regional feature extraction network is improved based on the FPN network and includes:

[0053] The image pyramid in the bottom-up process: formed based on the multi-scale feature maps output by the last four layers of the backbone network;

[0054] The top-down fusion process: The fused feature pyramid is denoted as C4, C5, C6, and C7 from top to bottom respectively;

[0055] Bottom-up convolution fusion process: C7 is used as the starting feature map P1. During the bottom-up convolution process, after each layer of convolution, it is fused with the corresponding layer in the top-down process on the left to form the feature map of the upper layer.

[0056] Specifically, the region feature extraction network in the embodiment of the present application adds a bottom-up feature fusion process on the basis of the FPN network. Specifically, this fusion process includes:

[0057] The output feature map C7 of the original FPN network is used as the starting feature map P1, and P1 Figure 3 After a 3×3 convolution, it is downsampled by a factor of 2 and directly added to C6 to fuse and generate the feature map P2. P2 Figure 3 After a 3×3 convolution, it is downsampled by a factor of 2 and directly added to C5 to fuse and generate the feature map P3. P3 Figure 3 After a 3×3 convolution, it is downsampled by a factor of 2 and directly added to C4 to fuse and generate the feature map P4.

[0058] Next, the connection relationship between the region feature extraction network and the cascade detector module of the Cascade RCNN network in the embodiment of the present application will be described.

[0059] First, the cascade detector module of the Cascade RCNN network includes: 1 RPN network. The output of the RPN network is respectively connected to 3 region of interest extraction networks with different IoU thresholds. The output of each region of interest extraction network is respectively connected to a fully connected layer. Each fully connected layer respectively outputs the corresponding classification score and the candidate box region parameters. Denote the 3 region of interest extraction networks as ROI Align1, ROI Align2, and ROI Align3 respectively, and denote the fully connected layers connected to the 3 regions of interest as the first fully connected layer, the second fully connected layer, and the third fully connected layer. The candidate box region parameters output by the first fully connected layer are simultaneously input to ROI Align2, and the candidate box region parameters output by the second fully connected layer are simultaneously input to ROI Align3;

[0060] The P1, P2, P3, and P4 feature maps output by the region feature extraction network are input to the RPN network. At the same time, the P4 feature map is input to ROI Align1.

[0061] As Figure 2 shown, the detection process of the cascade detector module includes:

[0062] The feature maps P1, P2, P3, and P4 are input into the RPN together to obtain some candidate boxes. Then, the feature map P4 and these candidate boxes are input into the one-stage ROI Align1 together. Next, the result output by ROI Align1 is input into the one-stage fully connected layer FC1 to obtain the one-stage classification score S1 and the candidate box B1. The candidate boxes obtained by the RPN and B1 are input into the two-stage ROI Align2 together. Then, the result output by ROI Align2 is input into the two-stage FC2 to obtain the two-stage classification score S2 and the candidate box B2. Finally, the candidate boxes obtained by the RPN and B2 are input into ROI Align3 together. Next, the result output by ROI Align is input into the three-stage FC3 to obtain the final classification score S3 and the bounding box B3.

[0063] Furthermore, in the above bottom-up convolution process, the convolution parameters are 3×3 convolution followed by 2-fold downsampling.

[0064] In the embodiments of the present application, the last four layers Conv2, Conv3, Conv4, and Conv5 of the backbone network output the feature maps C0, C1, C2, and C3 in sequence. C0, C1, C2, and C3 form an image pyramid of the FPN network from bottom to top. In the top-down fusion process, C3 is used as the starting feature map C4. The result of 2-fold upsampling of the feature map C4 is added and fused with the result of 1*1 convolution of C2 to generate the feature map C5. The result of 2-fold upsampling of the feature map C5 is added and fused with the result of 1*1 convolution of C1 to generate the feature map C6. The result of 2-fold upsampling of the feature map C5 is added and fused with the result of 1*1 convolution of C0 to generate the feature map C7;

[0065] In the present application, in the top-down fusion process, 2-fold upsampling is used to increase the resolution of the feature map so that it is consistent with the resolutions of C0, C1, C2, and C3 in the image pyramid. At the same time, 1*1 convolution is performed on C0, C1, C2, and C3 so that the number of channels of C0, C1, C2, and C3 is consistent with that of C4, C5, C6, and C7, realizing the addition fusion of deep feature maps and shallow feature maps. Matching the top-down fusion process, in the bottom-up convolution process, 2-fold downsampling is used to reduce the resolution of the feature map so that it is consistent with the resolutions of C4, C5, C6, and C7. At the same time, 3×3 convolution is used to eliminate the feature distribution differences from C4, C5, C6, and C7, thereby ensuring the stability of features.

[0066] The three detection IOU values in the cascade detector of the above Cascade RCNN network are 0.5, 0.6, 0.7 in sequence or 0.6, 0.7, 0.8 in sequence.

[0067] Generally, the three detection IOU values of the Cascade RCNN network are set to 0.5, 0.6, and 0.7 in sequence. In the embodiments of the present application, for the improved Cascade RCNN network, the three detection IOU values can also be set to 0.6, 0.7, and 0.8, which can achieve better detection effects and will not reduce the detection performance of the Cascade RCNN network.

[0068] Before detecting the citrus images to be detected, it also includes the process of constructing the improved Cascade RCNN network and training the improved Cascade RCNN network based on the citrus huanglongbing sample image set. The sample image set includes multiple sample images and is divided into a training set and a validation set.

[0069] In the embodiments of the present application, for the citrus images taken in natural scenes, there are a large number of interference factors such as healthy leaves, other diseased leaves, and weeds in the image background, resulting in low detection accuracy. During the model training process, it also includes: performing data augmentation and data enhancement processing on the sample image set through the methods of cutmix splicing and CLAHE to reduce the interference of the background.

[0070] Specifically, the cutmix splicing method includes:

[0071] Obtain the citrus huanglongbing leaves, other diseased leaves, and healthy leaves in the citrus huanglongbing sample image set;

[0072] Cut out the huanglongbing area leaves in the huanglongbing leaves and splice them onto other diseased leaves and / or healthy leaves.

[0073] Through the above cutmix splicing, the number and diversity of detection targets in the training set images are increased. Because in the actual environment, images with multiple diseases mixed are everywhere, the images after cutmix splicing provide image samples with multiple diseases coexisting for the training process.

[0074] Among them, performing data enhancement processing on the sample image set through CLAHE includes:

[0075] Extract the components of the sample image data, and separate the first component image and the second component image through logarithmic calculation;

[0076] Perform frequency domain conversion on the image based on the discrete cosine transform;

[0077] Perform filtering processing on the data after frequency domain conversion based on a preset filter, and the filter is: where D(x, y) is the distance from the pixel point (x, y) to the filter center, and a is an adjustable parameter;

[0078] Perform pixel value correction on pixel points of the first component image after filtering processing by using the CLAHE method, and perform linear transformation on the second component image at the same time;

[0079] Perform inverse discrete cosine transform on the corrected first component image and the transformed second component image to obtain enhanced sample image data.

[0080] The embodiment of the present application also provides a natural background citrus huanglongbing detection system based on an improved Cascade RCNN network. This citrus huanglongbing detection system can adopt software modules or hardware modules, or a combination of both to become a part of a computer device. Specifically, the system includes:

[0081] An image to be detected acquisition module, configured to acquire an image of citrus to be detected;

[0082] A feature extraction module, which uses the backbone network improved by the Cascade RCNN network to extract features from the image of citrus to be detected. At least one layer at the back of the improved backbone network adopts deformable convolution;

[0083] A feature bidirectional fusion module, which uses the region feature extraction network improved by the Cascade RCNN network to perform high-level semantic feature extraction on the features output by the backbone network. The improved region feature extraction network includes a bottom-up convolution fusion process added on the right side during the top-down fusion process of the FPN network;

[0084] A huanglongbing detection module, which uses the cascade detector of the Cascade RCNN network to perform a multi-stage cascade target recognition detection process on the high-level semantic features output by the region feature extraction network to obtain the huanglongbing detection result of the image of citrus to be detected.

[0085] For the specific limitations of the citrus huanglongbing detection system, reference can be made to the limitations of the citrus huanglongbing detection method in the above text, which will not be elaborated here. Each module in the above citrus huanglongbing detection system can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0086] The embodiment of the present application also provides a citrus huanglongbing detection device, which includes:

[0087] A processor;

[0088] A memory for storing instructions executable by the processor;

[0089] Among them, the processor realizes the above-mentioned natural background citrus huanglongbing detection method by running the executable instructions.

[0090] The citrus huanglongbing detection device provided by the embodiment of the present application may be a server. The device includes a processor, a memory, and a network interface connected through a system bus. The memory may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The memory can store data to support the operation of the device terminal. For example, it can store an operating system, application programs, a database, etc. The processor may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logic blocks, modules, and circuits described in connection with the disclosure of the present application. The processor may also be a combination that realizes computing functions, such as a combination including one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0091] The embodiment of the present application also provides a computer-readable storage medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the steps of the above-mentioned natural background citrus huanglongbing detection method are realized.

[0092] In the embodiment of the present application, the computer-readable medium may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination of the above.

[0093] The present invention is not limited to the above specific embodiments. Those of ordinary skill in the art starting from the above concepts and making various transformations without creative labor fall within the protection scope of the present invention.

Claims

1. A method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network, characterized in that It includes the following steps: Obtain the citrus image to be detected; Use the backbone network improved by the Cascade RCNN network to extract features from the citrus image to be detected. At least one layer in the improved backbone network uses deformable convolution; Use the region feature extraction network improved by the Cascade RCNN network to perform high-level semantic feature extraction on the features output by the backbone network. The improved region feature extraction network includes a bottom-up convolutional fusion process added to the right side during the top-down fusion process of the FPN network; The region feature extraction network is improved based on the FPN network and includes: an image pyramid in the bottom-up process: formed based on the multi-scale feature maps output by the last 4 layers of the backbone network; Top-down fusion process: The fused feature pyramid is denoted as C4, C5, C6, and C7 from top to bottom respectively; Bottom-up convolutional fusion process: C7 is used as the starting feature map P1, and after each layer of convolution in the bottom-up convolution process, it is fused with the corresponding layer in the left top-down process to form the upper layer feature map; Use the cascade detector of the Cascade RCNN network to perform a multi-stage cascade target recognition and detection process on the high-level semantic features output by the region feature extraction network, and obtain the detection result of huanglongbing for the citrus image to be detected.

2. The method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network according to claim 1, characterized in that The backbone network is implemented based on ResNetXt101.

3. The method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network according to any one of claims 1 or 2, characterized in that Based on the concept of the DCNv2 model, the last 3 layers of the backbone network use deformable convolution.

4. The method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network according to claim 1, characterized in that In the bottom-up convolution process, the convolution parameters use 3×3 convolution followed by 2-fold downsampling.

5. The method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network according to claim 1, characterized in that In the cascade detector of the Cascade RCNN network, the three detection IOU values are 0.5, 0.6, 0.7 in sequence or 0.6, 0.7, 0.8 in sequence.

6. The method for detecting citrus huanglongbing in natural background based on an improved Cascade RCNN network according to claim 1, characterized in that Before detecting the citrus image to be detected, it also includes: The process of training the improved Cascade RCNN network based on the huanglongbing citrus sample image set. The sample image set includes multiple sample images and is divided into a training set and a validation set; The process of training the improved Cascade RCNN network based on the huanglongbing citrus sample image set also includes: Using the methods of cutmix stitching and CLAHE to perform data augmentation on the sample image set.

7. A detection system for citrus huanglongbing in natural background based on an improved Cascade RCNN network, characterized in that It includes: An image acquisition module to be detected, used to obtain the citrus image to be detected; A feature extraction module that uses the backbone network improved by the Cascade RCNN network to extract features from the citrus image to be detected. At least one layer at the back of the improved backbone network uses deformable convolution; A feature bidirectional fusion module that uses the region feature extraction network improved by the Cascade RCNN network to perform high-level semantic feature extraction on the features output by the backbone network. The improved region feature extraction network includes a bottom-up convolutional fusion process added to the right side during the top-down fusion process of the FPN network; The region feature extraction network is improved based on the FPN network and includes: an image pyramid in the bottom-up process: formed based on the multi-scale feature maps output by the last 4 layers of the backbone network; Top-down fusion process: The fused feature pyramids are denoted as C4, C5, C6, and C7 from top to bottom; Bottom-up convolutional fusion process: C7 is used as the starting feature map P1. During the bottom-up convolutional process, after each layer is convolved, it is fused with the corresponding layer in the left top-down process to form the feature map of the upper layer; The huanglongbing detection module uses the cascade detector of the Cascade RCNN network to perform a multi-stage cascaded object recognition and detection process on the high-level semantic features output by the region feature extraction network, and obtains the huanglongbing detection result of the citrus image to be detected.

8. A citrus huanglongbing detection device, characterized in that The device includes: A processor; A memory for storing processor-executable instructions; Wherein, the processor realizes the method according to any one of claims 1-6 by running the executable instructions.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the instructions are executed by the processor, the steps of the method according to any one of claims 1-6 are realized.

Images