An unmanned aerial vehicle data fusion deep learning fish disease detection method, system, device and medium

Abstract

The application provides a kind of unmanned aerial vehicle data fusion deep learning fish disease detection method, system, equipment and medium, the method comprises the following steps: fish data is obtained by unmanned aerial vehicle, and the fish data is annotated processing;The fish data after annotation processing is input into artificial intelligence detection model for feature fusion to obtain target fish data;The target fish data is detected by the constructed gate heat conduction attention fusion model;The present method collects data by unmanned aerial vehicle, analyzes data, provides a large amount of basic data for future fish disease prevention and treatment industrialization and fish disease research, trains a special fish disease detection model, is low in cost and simple in operation, and is extremely convenient for popularization to vast rural areas;At the same time, the mixed algorithm is used to greatly improve the accuracy of detecting fish disease;The system has the same beneficial effects.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a deep learning method, system, device and medium for fish disease detection based on UAV data fusion. Background Technology

[0002] Spring viremia in carp and hemorrhagic disease in grass carp can threaten the aquaculture industry of major freshwater fish species, including black carp, grass carp, common carp, and crucian carp. These are strictly monitored Class II diseases, both causing symptoms such as hemorrhage and tissue ulceration in fish. However, current disease monitoring technologies mainly rely on viral PCR and fish histological morphology testing, requiring sophisticated instruments and professionally trained personnel. This makes them difficult to implement in vast rural areas, and large-scale monitoring is extremely costly.

[0003] Current publicly available deep learning detection methods for hemorrhagic diseases mainly extract images from videos, which suffers from problems such as limited sample data. Furthermore, analyzing images extracted from videos results in the loss of time-series information in the videos, including the difference in continuous swimming patterns between sick and healthy fish. As a result, existing artificial intelligence methods have low accuracy in detecting fish diseases.

[0004] Therefore, how to improve the accuracy of fish disease detection while reducing costs is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a deep learning-based fish disease detection method, system, device, and medium based on UAV data fusion. This method utilizes UAVs to collect and analyze data, providing a large amount of fundamental data for future fish disease prevention and control industrialization and fish disease research. It trains a dedicated fish disease detection model, is cost-effective and easy to operate, and is extremely convenient for widespread application in rural areas. Furthermore, the use of a hybrid algorithm significantly improves the accuracy of fish disease detection.

[0006] The first objective of this invention is to provide a deep learning-based fish disease detection method based on UAV data fusion;

[0007] The technical solution provided by this invention is as follows:

[0008] A deep learning-based fish disease detection method based on UAV data fusion includes the following steps:

[0009] Fish data is acquired using drones, and the fish data is then labeled.

[0010] The labeled fish data is input into an artificial intelligence detection model for feature fusion to obtain target fish data;

[0011] The target fish data were detected using a gated heat conduction attention fusion model.

[0012] Preferably, the step of acquiring fish data via drone and annotating the fish data specifically includes:

[0013] The drone collects fish image data and fish video data;

[0014] The fish image data and fish video data are labeled using a preset labeling strategy.

[0015] Preferably, the step of inputting the labeled fish data into an artificial intelligence detection model for feature fusion to obtain target fish data specifically includes:

[0016] The labeled fish data is input into the artificial intelligence detection model for training to extract 2D appearance feature data and 3D spatiotemporal feature data;

[0017] The 2D appearance feature data and the 3D spatiotemporal feature data are fused according to the decoupling-fusion strategy to obtain target fish data.

[0018] Preferably, the step of inputting the labeled fish data into the artificial intelligence detection model for training to extract 2D appearance feature data and 3D spatiotemporal feature data specifically includes:

[0019] The labeled fish data is input into the artificial intelligence detection model, and the 2D appearance feature data is output through the ConvNeXtV2 backbone network.

[0020] Preferably, the step of inputting the labeled fish data into the artificial intelligence detection model for training to extract 2D appearance feature data and 3D spatiotemporal feature data further includes:

[0021] The labeled fish data is input into the artificial intelligence detection model and the 3D spatiotemporal feature data is output through the Video SwinTransformer network.

[0022] Preferably, the step of fusing the 2D appearance feature data and the 3D spatiotemporal feature data according to the decoupling-fusion strategy to obtain target fish data specifically includes:

[0023] The 2D appearance feature data is decoupled into a classification-biased 2D appearance feature data and a regression-biased 2D appearance feature data by passing two independent convolution heads.

[0024] The classification-biased 2D appearance feature data, the regression-biased 2D appearance feature data, and the 3D spatiotemporal feature data are fused together to obtain target fish data.

[0025] Preferably, the gated heat conduction attention fusion model is constructed as follows:

[0026] A gated heat conduction attention fusion model is constructed by combining content-adaptive gating mechanisms with physics-inspired global information.

[0027] The second objective of this invention is to provide a deep learning-based fish disease detection system based on UAV data fusion;

[0028] The technical solution provided by this invention is as follows:

[0029] A deep learning-based fish disease detection system based on UAV data fusion includes: a labeling module, an acquisition module, and a detection module;

[0030] The annotation module is used to acquire fish data through a drone and to annotate the fish data.

[0031] The acquisition module is used to input the labeled fish data into an artificial intelligence detection model for feature fusion in order to obtain target fish data;

[0032] The detection module is used to detect the target fish data using a constructed gated heat conduction attention fusion model.

[0033] The third objective of this invention is to provide an electronic device;

[0034] The technical solution provided by this invention is as follows:

[0035] An electronic device, comprising:

[0036] At least one processor; and

[0037] A memory communicatively connected to the at least one processor, the memory storing a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method steps of any one of the deep learning fish disease detection methods based on UAV data fusion.

[0038] A fourth objective of this invention is to provide a computer-readable storage medium;

[0039] The technical solution provided by this invention is as follows:

[0040] A computer-readable storage medium for storing a computer program for causing a computer to perform the steps of any one of the deep learning fish disease detection methods based on UAV data fusion.

[0041] This invention provides a deep learning-based fish disease detection method using UAV data fusion, comprising the following steps: acquiring fish data via UAV and labeling the fish data; inputting the labeled fish data into an artificial intelligence detection model for feature fusion to obtain target fish data; and detecting the target fish data using a constructed gated heat conduction attention fusion model. This method uses UAVs to collect and analyze data, providing a large amount of basic data for future fish disease prevention and control industrialization and fish disease research, training a dedicated fish disease detection model, and is low-cost and simple to operate, making it extremely convenient for widespread application in rural areas. Simultaneously, the use of a hybrid algorithm significantly improves the accuracy of fish disease detection.

[0042] The present invention also provides a deep learning fish disease detection system based on UAV data fusion. Since this system and the deep learning fish disease detection method based on UAV data fusion solve the same technical problem and belong to the same technical concept, they should have the same beneficial effects, and will not be described in detail here. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating a deep learning-based fish disease detection method using UAV data fusion, as described in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the video and image fusion dual-backbone spatiotemporal detection model architecture in an embodiment of the present invention;

[0046] Figure 3 This is a structural diagram of an ELAN-based PA-FPN network in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of video data feature extraction in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram comparing the detection accuracy of various models in the embodiments of the present invention;

[0049] Figure 6 This is a structural diagram of a deep learning fish disease detection system based on UAV data fusion, as described in an embodiment of the present invention.

[0050] Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0051] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] like Figure 1 As shown in the figure, this embodiment of the invention provides a deep learning-based fish disease detection method based on UAV data fusion, comprising the following steps:

[0053] S1. Acquire fish data using a drone and then label the fish data;

[0054] S2. Input the labeled fish data into an artificial intelligence detection model for feature fusion to obtain target fish data;

[0055] S3. The target fish data is detected using a gated heat conduction attention fusion model.

[0056] In steps S1 to S3, fish data is collected using a drone, and then labeled according to a labeling strategy to balance efficiency and accuracy. The labeled fish data is then input into an artificial intelligence detection model for feature fusion to obtain target fish data. Finally, a gated heat conduction attention fusion model is used to detect the target fish data. Data collection and analysis using drones provide a large amount of basic data for future fish disease prevention and control industrialization and fish disease research. Training a dedicated fish disease detection model is cost-effective and simple to operate, making it extremely convenient for widespread application in rural areas. Simultaneously, the use of hybrid algorithms significantly improves the accuracy of fish disease detection.

[0057] Preferably, the step of acquiring fish data via drone and annotating the fish data specifically includes:

[0058] The drone collects fish image data and fish video data;

[0059] The fish image data and fish video data are labeled using a preset labeling strategy.

[0060] In practical application, fish image data and fish video data are collected by drones. For video data, in order to ensure the labeling quality of the FishAction-2 dataset, a rigorous multi-stage labeling and quality control process was designed and implemented. First, clear definitions were established for the two core categories of "healthy" and "sick". "Healthy" individuals are defined as active swimmers with smooth bodies, while "sick" individuals are defined as: (1) abnormal trajectory; (2) bleeding features. Then, a semi-automatic strategy of "large model pre-labeling - manual refinement" was adopted to balance efficiency and accuracy. First, a large language model was used to automatically generate the initial bounding boxes of all fish on the image with "fish" as the text prompt, which improved the efficiency of initial localization. Then, manual review and classification were carried out: the images with pre-labeled boxes were submitted to a team of three experts with experience in fish disease diagnosis for review. The labelers reviewed the accuracy of each bounding box one by one and assigned the final label of "healthy" or "sick" to the target in the box according to the above biological standards. Data scaling: After the annotation task was completed, all images and their corresponding annotation files were scaled down to a resolution of 320×240. This was done to reduce the storage size of the dataset and decrease the computational cost during subsequent model training.

[0061] Preferably, the step of inputting the labeled fish data into an artificial intelligence detection model for feature fusion to obtain target fish data specifically includes:

[0062] The labeled fish data is input into the artificial intelligence detection model for training to extract 2D appearance feature data and 3D spatiotemporal feature data;

[0063] The 2D appearance feature data and the 3D spatiotemporal feature data are fused according to the decoupling-fusion strategy to obtain target fish data.

[0064] In practical applications, annotated fish data is input into an AI detection model for training. This model employs a decoupling-fusion strategy to maximize the synergistic effect of 2D appearance features and 3D spatiotemporal features. The final detection task comprises two related yet distinct sub-tasks: classification (identifying healthy or diseased animals) and regression (precisely locating bounding boxes). The classification task relies more on subtle behavioral patterns throughout the video segment, while the regression task requires more precise edge and contour information. Therefore, as... Figure 2As shown, instead of directly adding the 3D features to the fused 2D feature map, the multi-scale feature map (PA-FPN) output from the 2D backbone is first decoupled into a "classification-biased" 2D feature and a "regression-biased" 2D feature through two independent convolutional heads. Then, the extracted 3D spatiotemporal features are fused with these two decoupled feature maps respectively to obtain the target fish data. This targeted fusion allows different parts of the model to focus on their respective strengths, thereby improving overall detection performance.

[0065] Preferably, the step of inputting the labeled fish data into the artificial intelligence detection model for training to extract 2D appearance feature data and 3D spatiotemporal feature data specifically includes:

[0066] The labeled fish data is input into the artificial intelligence detection model, and the 2D appearance feature data is output through the ConvNeXtV2 backbone network.

[0067] In practical applications, the interface for processing 2D image data uses an ELAN-based PA-FPN network. When designing an efficient network, the effectiveness of the gradient path must be considered. The E-ELAN (Extended Efficient Layer Aggregation Networks) structure can continuously enhance the network's learning ability without disrupting the original gradient path. Specifically, for example… Figure 3 As shown, Figure 3 (a) PA-FPN Neck structure diagram; Figure 3 (b) shows the E-ELAN architecture diagram; feature maps (C3, C4, C5) from stages 3, 4, and 5 of the ConvNeXtV2 backbone network were extracted. These feature maps formed a pyramid with dimensions of H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32. Subsequently, these three layers of feature maps were fed into the ELAN-based PA-FPN, where semantic fusion and localization information enhancement were performed to generate a set of 2D image appearance feature maps rich in multi-scale contextual information and taking into account both semantics and details. These were then fused with the spatiotemporal features of the 3D video.

[0068] While maintaining the core stacked block design of ConvNeXtV1, ConvNeXtV2 introduces a novel Global Response Normalization (GRN) layer to replace the original LayerScale. The introduction of GRN aims to enhance feature contrast and selectivity by increasing competition among neurons. GRN enhances competition and contrast between channels through a simple three-step process—global feature aggregation, feature normalization, and feature calibration—thereby improving feature diversity. Its core computation process is as follows:

[0069] ;

[0070] Represents the first element in the input feature map. Characteristics of each channel; Indicates the first L2 norm of each channel feature; This represents the total number of channels in the feature map; This represents the average L2 norm of all channels; This represents a very small constant (epsilon) used to prevent the denominator from being zero and to ensure the stability of numerical calculations; and This represents two learnable parameters, used to scale and shift the normalized features, respectively. This indicates the number after calibration. The output features of each channel. This embodiment adopts the "backbone + neck" paradigm of ConvNeXtV2 and PA-FPN, which can effectively balance feature representation capability and computational efficiency. ConvNeXtV2 exhibits excellent local feature capture capability while maintaining computational efficiency, while PA-FPN can efficiently integrate multi-scale features, which is crucial for detecting fish of different sizes and distances in complex aquaculture environments.

[0071] Preferably, the step of inputting the labeled fish data into the artificial intelligence detection model for training to extract 2D appearance feature data and 3D spatiotemporal feature data further includes:

[0072] The labeled fish data is input into the artificial intelligence detection model and the 3D spatiotemporal feature data is output through the Video SwinTransformer network.

[0073] In practical applications, the interface for processing 3D video data uses the Transformer architecture as the 3D backbone of the model, replacing traditional 3D convolutional neural networks. Directly applying the global self-attention mechanism of the standard Transformer to video results in high computational and memory costs because the introduction of the temporal dimension into video data dramatically increases the computational load. The Video Swin Transformer achieves a balance between performance and efficiency by introducing the effective inductive bias of spatiotemporal locality, confining self-attention computation to non-overlapping 3D windows. This design significantly reduces computational complexity. To achieve cross-window information interaction while maintaining efficient computation, a clever 3D Shifted Window based MSA (3D SW-MSA) mechanism is employed.

[0074] For a given Transformer module, it employs a conventional windowing strategy (3D W-MSA) to partition the input T'×H'×W' 3D token tensor into... T′ / P × H′ / M × W′ / M A non-overlapping 3D window of size P×M×M is used, and self-attention is computed within it. In the next module, the window's mesh is shifted along the time, height, and width axes by a distance of (P / 2, M / 2, M / 2) tokens (3D SW-MSA). For example, with an 8×8×8 token input and a 4×4×4 window... Figure 4 As shown, Figure 4 (a) is a case study for feature extraction from video data. Figure 4(b) is a flowchart illustrating the video data feature extraction process. Corresponding to the calculation formula below, the conventional partitioning of data tokens using the original deep learning model would result in 2×2×2=8 windows. In this application, after the window is moved by (2,2,2) tokens, the new window partitioning will cross the original window boundaries, theoretically forming 3×3×3=27 windows. This shift operation allows tokens in different windows in the previous layer to be partitioned into the same window in the new layer, thus achieving cross-window information flow. It is worth noting that although the number of windows appears to increase after shifting, through an efficient cyclic shift and masking mechanism, the final number of windows used for calculation remains 8. The successive Transformer modules alternate between conventional window (3D-W-MSA) and shifted window (3D-SW-MSA) configurations, and their calculation process can be summarized as follows:

[0075] ;

[0076] in, It is the first The output features of each module are represented by LN, which stands for layer normalization, and FFN, which stands for feedforward network. This design is not only theoretically efficient, but its effectiveness has also been verified on several large-scale video recognition benchmarks.

[0077] Preferably, the gated heat conduction attention fusion model is constructed as follows:

[0078] A gated heat conduction attention fusion model is constructed by combining content-adaptive gating mechanisms with physics-inspired global information.

[0079] In practical applications, after high-quality 2D appearance features and 3D spatiotemporal features, a novel attention fusion model called GateHCO (Gate-Controlled Heat Conduction Attention Fusion) is proposed to achieve deep and efficient fusion of these two heterogeneous features with different origins and properties. This model is designed based on a core inductive bias: the propagation and association of visual semantic information can be simulated by the heat conduction process in the physical world. By introducing the physical principle of heat conduction into feature fusion and combining it with an efficient gating mechanism, GateHCO can achieve global, adaptive deep fusion of heterogeneous features. Theoretical analysis shows that the computational complexity of HCO is O(N¹·kJ / L). 5 This makes it theoretically more efficient than self-attention, while in practice it also demonstrates a broader global perspective and greater flexibility than standard convolution.

[0080] This process can be transformed into a heat conduction operator (HCO) that can be used in deep learning. Specifically, the input feature map... The initial "temperature distribution" is considered. First, it is transformed to the frequency domain using a computationally efficient two-dimensional discrete cosine transform (DCT-2D), which is equivalent to a Fourier transform and applies Neumann boundary conditions. Then, in the frequency domain, the features are subjected to an exponentially decaying filter. Multiply them. Finally, transform them back to the spatial domain using the inverse discrete cosine transform (IDCT-2D). The entire process can be summarized by the following equation:

[0081] ;

[0082] in It is the input feature map. This is the output feature map after information fusion. Essentially, it is an adaptive filter: it suppresses different modes (such as high-frequency edge textures and low-frequency smooth regions) to varying degrees, thereby simulating the smoothing and balancing effect brought about by heat diffusion.

[0083] However, a robust fusion module not only needs to efficiently "mix" global information, but also must be able to selectively "preserve" original local details. Therefore, instead of directly using HCO, it is embedded as a core component within an efficient gating structure, forming the complete GateHCO model.

[0084] In this model, the input feature tensor first passes through a fully connected layer, and is then split into three independent branches: the content path, The features of this branch will be fed into the HCO module of the design for global information mixing. (Retained branch (Identity Path)) This branch retains its characteristics directly, carrying the original, more localized information without modification. (Gate Path) This branch uses an activation function to generate a dynamic gating signal to control the information flow.

[0085] Ultimately, the model's output is a dynamic combination of content branches and retained branches using gating signals:

[0086] ;

[0087] ;

[0088] in, Represents the input feature tensor; Indicates a fully connected layer; This represents a regularization technique that, during training, sets the output of the entire main path to zero with a certain probability. This represents the Sigmoid activation function; This represents a two-dimensional heat conduction operator. This represents the output feature tensor.

[0089] This design gives the model great flexibility. Gating signals. Based on the input content, the system can learn to decide whether to rely more on features diffused through global information or on original local features in the final fused features. This content-adaptive gating mechanism, combined with physically inspired global information mixing, constitutes the GateHCO model. Finally, after concatenating 2D appearance features and 3D spatiotemporal features, a deep fusion is performed through a GateHCO model, which can accurately identify diseased fish and provide high-quality, information-rich feature representations for subsequent localization and classification tasks.

[0090] Compared to existing image processing models, although models like YOLOv7 achieve a high mAP@50 of 91.2% due to their powerful static feature extraction capabilities, all 2D detectors failed to surpass the model presented in this application. These results were achieved under the favorable conditions of higher input resolution and richer spatial details observed by 2D static detectors. Even so, the model presented in this application still outperforms due to its deep understanding of temporal information. This strongly demonstrates that many key cues for judging the health status of fish (such as abnormal swimming posture and struggling) are inherently dynamic and cannot be fully captured by isolated single-frame images. The model presented in this application successfully overcomes this limitation by explicitly modeling temporal dependencies. After establishing the necessity of temporal modeling, this application further directly compares the model with the YOWO series, which is also a single-stage spatiotemporal detection framework. Early versions of YOWO, due to their simple feature fusion strategy, only achieved 53.86% mAP@50 on the dataset presented in this application. Its improved version YOWO-v2

[24] has significantly improved performance to 84.94%, demonstrating the significant progress of its architecture. Nevertheless, the model proposed in this application has achieved a comprehensive improvement over YOWO-v2 in all metrics, with mAP@50 being nearly 9 percentage points higher. Table 1 shows the general 2D detection model, while Table 2 shows YOWOv1 and YOWOv2 as 3D spatiotemporal detection models. In comparison, the AP value of this method has a significant advantage. At the same time, the model size is significantly reduced compared to the 3D spatiotemporal model. This method is 156.16MB, which is 48-220MB smaller than other 3D models; it can also be achieved through Figure 5 As shown in the figure, the size of the circle represents the number of model parameters. The number of parameters in this model is similar to that of other detection methods. The horizontal axis represents the computational cost, which is lower than that of other methods. The vertical axis represents the average detection accuracy, which is higher than that of other methods.

[0091] Table 1 Comparison of Accuracy of General 2D Detection Models

[0092]

[0093] Table 2 Comparison of Accuracy of 3D Spatiotemporal Detection Models

[0094]

[0095] like Figure 6 As shown, this embodiment of the invention provides a deep learning fish disease detection system based on UAV data fusion, including: a labeling module, an acquisition module, and a detection module;

[0096] The annotation module is used to acquire fish data through a drone and to annotate the fish data.

[0097] The acquisition module is used to input the labeled fish data into an artificial intelligence detection model for feature fusion in order to obtain target fish data;

[0098] The detection module is used to detect the target fish data using a constructed gated heat conduction attention fusion model.

[0099] In practical applications, the deep learning-based fish disease detection system based on UAV data fusion includes a labeling module, an acquisition module, and a detection module. The acquisition module is connected to both the labeling and detection modules. The labeling module acquires fish data from the UAV, labels the data, and then transmits the labeled data to the acquisition module. The acquisition module inputs the labeled fish data into an AI detection model for feature fusion to obtain target fish data, which is then transmitted to the detection module. The detection module uses a constructed gated heat conduction attention fusion model to detect the target fish data. This system analyzes the data collected by the UAV through the labeling, acquisition, and detection modules, providing a large amount of basic data for future fish disease prevention and control industrialization and fish disease research. It trains a dedicated fish disease detection model, which is low-cost and easy to operate, making it extremely convenient for widespread application in rural areas. Furthermore, the use of a hybrid algorithm significantly improves the accuracy of fish disease detection.

[0100] Furthermore, embodiments of this application also disclose an electronic device, Figure 7This is a structural diagram of an electronic device according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0101] Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of this application. The electronic device 20 specifically includes: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the deep learning fish disease detection method based on UAV data fusion disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment can specifically be an electronic computer.

[0102] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a deep learning fish disease detection channel for UAV data fusion between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0103] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0104] The operating system 221 manages and controls the various hardware devices on the electronic device 20 and the computer program 222 to enable the processor 21 to perform operations and processing on the data 223 in the memory 22. It can be Windows Server, Netware, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the deep learning fish disease detection method based on UAV data fusion executed by the electronic device 20 as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the deep learning fish disease detection device based on UAV data fusion from external devices, and may also include data collected by its own input / output interface 25.

[0105] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0106] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned deep learning fish disease detection method based on UAV data fusion. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0107] It should be understood that the use of terms such as "method," "apparatus," "unit," and / or "module" in this application is merely to distinguish one method of different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0108] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0109] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0110] If a flowchart is used in this application, it is used to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0111] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning fish disease detection method for unmanned aerial vehicle data fusion, characterized in that, The method comprises the following steps: acquiring fish data by a UAV and performing annotation processing on the fish data; inputting the fish data after the annotation processing into an artificial intelligence detection model for feature fusion to obtain target fish data; detecting the target fish data by a constructed gated heat conduction attention fusion model; the inputting of the fish data after the annotation processing into the artificial intelligence detection model for feature fusion to obtain target fish data specifically comprises: training the fish data after the annotation processing in the artificial intelligence detection model to extract 2D appearance feature data and 3D spatiotemporal feature data, specifically comprising: outputting the 2D appearance feature data by the ConvNeXtV2 backbone network of the artificial intelligence detection model to which the fish data after the annotation processing is inputted; outputting the 3D spatiotemporal feature data by the Video SwinTransformer network of the artificial intelligence detection model to which the fish data after the annotation processing is inputted; fusing the 2D appearance feature data and the 3D spatiotemporal feature data according to a decoupling-fusion strategy to obtain target fish data.

2. The method according to claim 1, wherein, the acquiring of fish data by a UAV and the performing of annotation processing on the fish data specifically comprises: collecting fish picture data and fish video data by the UAV; performing annotation processing on the fish picture data and the fish video data by a preset annotation strategy. 3.The unmanned aerial vehicle data fusion deep learning fish disease detection method of claim 1, wherein, the fusing of the 2D appearance feature data and the 3D spatiotemporal feature data according to a decoupling-fusion strategy to obtain target fish data specifically comprises: decoupling the 2D appearance feature data into a classification-biased 2D appearance feature data and a regression-biased 2D appearance feature data through two independent convolution heads; fusing the classification-biased 2D appearance feature data, the regression-biased 2D appearance feature data and the 3D spatiotemporal feature data to obtain target fish data. 4.The unmanned aerial vehicle data fusion deep learning fish disease detection method of claim 1, wherein, the construction mode of the gated heat conduction attention fusion model is: constructing the gated heat conduction attention fusion model through a content-adaptive gating mechanism and a physically-inspired global information.

5. A deep learning fish disease detection system for unmanned aerial vehicle data fusion, characterized in that, comprises: an annotation module, an acquisition module and a detection module; the annotation module is used for acquiring fish data by a UAV and performing annotation processing on the fish data; the acquisition module is used for inputting the fish data after the annotation processing into an artificial intelligence detection model for feature fusion to obtain target fish data; the detection module is used for detecting the target fish data by a constructed gated heat conduction attention fusion model; the acquisition module is specifically used for: training the fish data after the annotation processing in the artificial intelligence detection model to extract 2D appearance feature data and 3D spatiotemporal feature data, specifically comprising: outputting the 2D appearance feature data by the ConvNeXtV2 backbone network of the artificial intelligence detection model to which the fish data after the annotation processing is inputted; The fish data processed by the marking is input into the artificial intelligence detection model, and the 3D space-time feature data is output through a Video SwinTransformer network. The 2D appearance feature data and the 3D space-time feature data are fused according to a decoupling-fusion strategy to obtain target fish data.

6. An electronic device, comprising: Comprise: At least one processor; And The memory is in communication connection with the at least one processor, and the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the method of any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The storage medium is used to store a computer program, and the computer program is used to enable a computer to execute the method of any one of claims 1 to 4.

Images