A mosquito multi-form identification model, a training method, an identification method and a system

Abstract

This invention discloses a mosquito multimorphic recognition model, training method, recognition method, and system. This application provides a complete method and system from data acquisition and model training to mobile deployment, enabling high-precision multi-attribute joint recognition of mosquitoes to run in real time on mobile terminals, greatly promoting the practical application of the technology in vector monitoring, public health, and other fields. The recognition method involves the following steps: inputting static images or video stream data of the mosquito to be identified into a terminal device deployed with a trained mosquito multimorphic recognition model; the terminal device performing feature extraction, multi-scale feature fusion, and target detection on the input image / video frame data using the mosquito multimorphic recognition model to complete the joint recognition of the mosquito target's species, sex, and developmental stage; and the terminal device outputting the mosquito target's location information and the corresponding species, sex, and developmental stage recognition results in real time.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent recognition technology, specifically relating to a technical solution for mosquito multimorphic recognition based on an improved convolutional neural network. Background Technology

[0002] Mosquitoes are important vectors for many infectious diseases, with species such as Aedes, Culex, and Anopheles widely involved in the transmission of diseases like dengue fever, malaria, and Zika virus. Accurate and rapid mosquito identification is a crucial aspect of disease control and vector surveillance. Current mosquito identification methods primarily rely on manual morphological identification, which requires specialized personnel, is labor-intensive, inefficient, and difficult to scale up and operate in real-time. Furthermore, different mosquitoes exhibit high morphological similarity, especially between adult males and females and during the larval stage, making traditional manual identification methods prone to misjudgment.

[0003] With the development of computer vision and deep learning technologies, image-based mosquito identification methods have gradually attracted attention. However, existing research mostly focuses on the identification of single mosquito morphology or single attribute, lacking a unified model for the multimorphic and multi-attribute characteristics of mosquitoes. In addition, the uneven distribution of mosquito samples among different categories limits the generalization ability of the model, making it difficult to meet the needs of practical applications.

[0004] Therefore, there is an urgent need for an intelligent identification method and system that can optimize the network structure based on the multimorphic characteristics of mosquitoes and can be directly applied to real-world scenarios. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems in the prior art, the present invention provides a mosquito multimorphic recognition model, method, system and training method with high recognition accuracy, strong generalization ability and deployable on mobile terminals.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The present invention provides a mosquito multimorphic recognition model, which is a convolutional neural network model based on an improved target detection convolutional neural network, and is adapted to the differences in target size, appearance and texture features between adult mosquitoes and larvae. The model includes a backbone network, a feature fusion network and a detection head. The backbone network is configured with a Stem module, a multi-layer HGBlock and DWConv module, and an SPPF module. The Stem module is used to perform initial feature extraction and downsampling on the input image. The HGBlock and DWConv modules are used to extract semantic features at different levels step by step. The SPPF module is used to expand the receptive field and aggregate multi-scale contextual information. The feature fusion network is configured with an Upsample module, a Concat module, and a Rep-Ghost module. The Upsample and Concat operations are used to achieve top-down and bottom-up fusion of multi-scale features, and the Rep-Ghost module is used for efficient feature generation and fusion. The detection head sets up corresponding detection branches for multi-scale outputs of P3, P4, and P5, and is configured with an SCDH structure and a ConvGN module. The SCDH structure is used to uniformly process the detection features, and the ConvGN module is used to improve the stability of model training. The detection head finally outputs the mosquito target location and the corresponding species, sex, and developmental stage information. The improved model maintains high detection accuracy while reducing the number of parameters and computational load, enabling joint identification of multiple attributes such as mosquito species, sex, and developmental stage. By adopting the above technical solution—a specific modular combination of backbone network, feature fusion network, and detection head—it specifically models the scale, appearance, and texture differences of mosquito targets, achieving efficient extraction and fusion of multi-scale features. This allows for the simultaneous location of mosquitoes and identification of their multi-dimensional attributes in a single forward inference.

[0007] This invention also provides a training method for a mosquito multimorphic recognition model, comprising the following steps: Step S1, mosquito image data acquisition: Collect mosquito image data, including the adult male and female morphology of Aedes mosquitoes, Culex mosquitoes and Anopheles mosquitoes, as well as the larval morphology of Aedes mosquitoes and Culex mosquitoes; Step S2, Image Preprocessing and Data Augmentation: The collected mosquito images are preprocessed and data augmentation is performed by rotating, mirroring, adjusting brightness and adding noise perturbation to improve the sample imbalance problem between mosquito categories. Step S3, Data annotation and dataset construction: The enhanced mosquito images are annotated in target detection format. The annotation content includes at least the location of the mosquito and the corresponding species, sex and developmental stage information, and a mosquito multimorphic recognition dataset is constructed. Step S4, Model Training: Input the mosquito multimorphic recognition dataset into the above-mentioned mosquito multimorphic recognition model for training, so that the model has the ability to jointly recognize multiple attributes such as mosquito species, sex and developmental stage, and obtain the trained mosquito multimorphic recognition model.

[0008] By adopting the above technical solution, using a dedicated dataset containing multi-morphological and multi-attribute annotations, and combining it with targeted data augmentation strategies, the class imbalance problem is effectively alleviated, providing sufficient data support for the model to learn the diverse features of mosquitoes, and ensuring that the trained model has excellent generalization performance.

[0009] Preferably, the preprocessing in step S2 further includes Gaussian blur and motion blur processing. Gaussian blur is used to simulate the loss of imaging details, and motion blur is used to simulate the motion blur caused by the relative motion of the target or camera during the acquisition process. By adopting the above technical solution, and by introducing various image degradation operations that simulate real-world scenarios in the preprocessing, the robustness of the model to complex imaging conditions (such as inaccurate focusing and target movement) is enhanced, and the recognition stability of the model in practical applications is improved.

[0010] Preferably, the target detection format in step S3 is the YOLO format. By adopting the above technical solution—using the YOLO format for annotation and training—it is easier to leverage the mature YOLO framework ecosystem for efficient model development, training, and deployment, thus improving the engineering efficiency of the technical solution.

[0011] The present invention provides a mosquito multimorphic recognition method based on the aforementioned model, comprising the following steps: Step A, Image / Video Data Input: Input the static image or video stream data of the mosquito to be identified into the terminal device that has been deployed with the trained mosquito multimorphic recognition model; Step B, Model Reasoning and Recognition: The terminal device uses a mosquito multimorphic recognition model to extract features, fuse multi-scale features, and detect targets from the input image / video frame data, thereby completing the joint recognition of multiple attributes of the mosquito target, including species, sex, and developmental stage. Step C, Identification Result Output: The terminal device outputs the location information of the mosquito target and the corresponding species, sex, and developmental stage identification results in real time.

[0012] By adopting the above technical solution, a complete end-to-end identification process is provided. Users only need to input images or videos to quickly obtain identification results containing location and multiple attributes, which greatly improves the automation level and convenience of mosquito monitoring.

[0013] Preferably, the model deployment step is also included: converting the PyTorch format mosquito multimorphic recognition model file obtained from the training end into the ONNX universal inference format, and then into the NCNN format model file that can be loaded by the mobile inference engine; completing the project configuration and dependency integration in the Android Studio development environment, compiling and packaging to generate an application installation package, and deploying the NCNN format model file on the terminal device. By adopting the above technical solution, through a standardized model conversion process, complex deep learning models can be efficiently deployed to mobile terminals with limited computing resources, achieving offline and real-time mosquito recognition capabilities and eliminating dependence on cloud servers.

[0014] Preferably, the dependency integration includes the introduction of the Ncnn-Android-Vulkan mobile inference framework and the OpenCV-Mobile mobile image processing library. Ncnn-Android-Vulkan is used for model loading and inference computation, while OpenCV-Mobile is used for reading image / video frames, preprocessing, and visualizing recognition results. By adopting the above technical solution, and utilizing industry-leading, efficient, and mature mobile AI inference and image processing libraries, the smoothness of the model's operation on mobile devices and the real-time performance of recognition are ensured, while reducing the complexity of application development.

[0015] Preferably, the terminal device is a mobile terminal, and the mosquito multimorphic recognition model runs on the mobile terminal as an application, supporting both static image recognition mode and video stream recognition mode. By adopting the above technical solution, integrating recognition capabilities into mobile terminals such as smartphones, on-site staff or ordinary users can conveniently and quickly identify and monitor mosquitoes anytime, anywhere, greatly expanding the application scenarios of the technology.

[0016] The present invention provides a mosquito multimorphic identification system, comprising a data acquisition module, a data processing module, a model training module, a model deployment module, and an identification application module; The data acquisition module is used to acquire mosquito image data of adult male and female morphology of Aedes, Culex, and Anopheles mosquitoes, as well as larval morphology of Aedes and Culex mosquitoes. The data processing module is used to preprocess and augment the collected mosquito images, and then to annotate the augmented images with target detection format to construct a mosquito multimorphic recognition dataset containing information on mosquito location, species, sex and developmental stage. The model training module is used to construct the above-mentioned mosquito multimorphic recognition model and train the model using the mosquito multimorphic recognition dataset to obtain the trained mosquito multimorphic recognition model. The model deployment module is used to complete the format conversion of the mosquito multimorphic recognition model and its deployment on the terminal device, so as to realize the inference operation of the model on the terminal device. The identification application module is used to receive mosquito image / video data to be identified, call the deployed model to complete inference identification, and output the mosquito identification results in real time.

[0017] By adopting the above technical solutions, a complete chain system covering data preparation, model training and optimization, terminal deployment and application has been built, organically combining algorithms, data and engineering implementation to form a complete and practical intelligent mosquito identification solution.

[0018] Preferably, the identification application module is deployed on a mobile terminal, supporting both static image recognition and video stream recognition modes, and can simultaneously output the location information of mosquito targets as well as multi-attribute identification results such as species, sex, and developmental stage. By adopting the above technical solution, the final output form and deployment carrier of the system are clarified, highlighting the practicality, portability, and real-time interactive capabilities of this technical solution.

[0019] Therefore, the present invention has the following beneficial effects: (1) Through the specially designed backbone network, feature fusion network and detection head structure, it realizes accurate modeling and efficient fusion of multi-scale and multi-morphological features of mosquitoes, reducing model complexity while ensuring high recognition accuracy; (2) Through data augmentation strategies that include multiple simulated real scenarios and processing for sample imbalance, it significantly improves the generalization ability and robustness of the model in actual complex environments; (3) It provides a complete method and system from data collection, model training to mobile terminal deployment, enabling high-precision mosquito multi-attribute joint recognition capabilities to run in real time on mobile terminals, greatly promoting the practical application of the technology in vector monitoring, public health and other fields; (4) The recognition results simultaneously include target location, species, sex and developmental stage information, providing comprehensive data support for refined and intelligent mosquito ecological research and control decision-making. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the mosquito multimorphic recognition model of the present invention.

[0021] Figure 2 This is a schematic diagram of the model recognition process of the present invention.

[0022] Figure 3 This is a schematic diagram of the mobile deployment process of the model of the present invention.

[0023] Figure 4 The images show samples of various mosquito species, including male and female samples of the three main disease-carrying mosquitoes, as well as the larval morphology of Aedes and Culex mosquitoes.

[0024] Figure 5 This is a flowchart of a training method for a mosquito multimorphic recognition model according to the present invention.

[0025] Figure 6 This is a flowchart of a mosquito multimorphic identification method according to the present invention.

[0026] Figure 7 This is a system block diagram of the mosquito multimorphic recognition system of the present invention. Detailed Implementation

[0027] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0028] like Figure 1 The mosquito multimorphic recognition model shown is a convolutional neural network model improved from a target detection convolutional neural network. It adapts to the differences in target scale, appearance, and texture features between adult mosquitoes and larvae. The model includes a backbone network, a feature fusion network, and a detection head. In this structure, the backbone network is responsible for extracting multi-level features from the input image, ranging from low-level details to high-level semantics. The feature fusion network efficiently fuses features from different levels, enabling the network to simultaneously perceive details of small targets (such as mosquito antennae and larvae) and understand the context of overall targets (such as the outline of adult mosquitoes). At the same time, through the multi-scale design and attribute classification branches of the detection head, it achieves accurate localization of mosquito targets of different sizes in the image and outputs multi-dimensional information such as their location, species, sex, and developmental stage at once. This avoids the efficiency loss and error accumulation caused by the need to connect multiple single-task models in traditional solutions, achieving a good balance between accuracy and speed.

[0029] The backbone network consists of a Stem module, multi-layer HGBlock and DWConv modules, and an SPPF module. The Stem module performs initial feature extraction and rapid downsampling on the input image. The HGBlock and DWConv modules extract semantic features at different levels step by step. The SPPF module expands the receptive field and aggregates multi-scale contextual information. In this structure, the Stem module, as the network's entry point, can quickly focus on the core region of the image, reducing subsequent computation. The HGBlock module enhances feature representation through an efficient gating mechanism, and the DWConv module further compresses the number of parameters using depthwise separable convolution. Simultaneously, the spatial pyramid pooling operation of the SPPF module aggregates contextual features at different scales without significantly increasing the computational burden. This enables the model to have robust feature extraction capabilities for both adult mosquitoes and larvae targets with drastic scale changes. This combination effectively resolves the contradiction between large target scale differences and lightweight model in mosquito identification.

[0030] The feature fusion network consists of an Upsample module, a Concat module, and a Rep-Ghost module. Upsample and Concat operations enable top-down and bottom-up fusion of multi-scale features, while the Rep-Ghost module handles efficient feature generation and fusion. In this architecture, the Upsample module amplifies deep semantic feature maps, and the Concat module concatenates them with shallow detail feature maps of the same scale, allowing high-level semantic information to guide lower-level features. Simultaneously, the introduction of the Rep-Ghost module replaces traditional standard convolutions for feature generation and fusion. This module offers richer feature transformation paths during training and can be reparameterized into an efficient, lightweight structure during inference. This significantly reduces the computational overhead and memory usage of the feature fusion process with almost no loss of accuracy, making the entire model more suitable for deployment on resource-constrained mobile devices.

[0031] The detection head sets corresponding detection branches for multi-scale outputs of P3, P4, and P5, and is configured with an SCDH structure and a ConvGN module. The SCDH structure is used to uniformly process the detection features, and the ConvGN module is used to improve the stability of model training. The detection head finally outputs the mosquito target location and the corresponding species, sex, and developmental stage information. In this structure, the P3, P4, and P5 branches correspond to small, medium, and large target scales, respectively, ensuring the detection sensitivity for mosquitoes of different body sizes (such as larvae and adults). The SCDH structure uniformly regularizes and recalibrates the detection features from different branches, enhancing feature consistency. At the same time, the Conv_GN module (convolution combined with batch normalization) replaces the traditional Conv-BN (convolution combined with batch normalization) structure, effectively alleviating the problem of inaccurate statistical estimation during small-batch training, improving the stability and convergence speed of the model training process, and ultimately enabling the detection head to robustly and accurately complete the joint task of localization and multi-attribute classification.

[0032] Key test and verification results of the mosquito multimorphic recognition model provided in the embodiments of this application. 1. Ablation experiment verification (module effectiveness) By introducing the HGNetv2 backbone network, Rep-Ghost feature fusion module, and SCDH weight-sharing detection head one by one, the impact of each module on model performance was verified: HGNetv2 backbone network: 2.35M parameters, 6.9 GFLOPs, which reduces computational cost compared to YOLOv8n (3.0M / 8.1GFLOPs), while achieving P=98.1%, R=97.0%, and mAP50=98.7%, demonstrating superior small target feature extraction capabilities; Rep-Ghost feature fusion module: After replacing the C2f structure of YOLOv8, P=97.6%, R=97.1%, mAP50-95=83.4%, the feature representation capability is significantly improved while the number of parameters (2.6M) and the computational cost (7.3GFLOPs) remain unchanged. SCDH weight-sharing detector head: When using GN normalization, P=97.8%, R=97.7%, mAP50=99.0%, the number of parameters is reduced to 2.4M, and the computational load is 6.5GFLOPs, effectively reducing detector head redundancy; Overall model (this model): After integrating the three modules, the number of parameters decreased from 3.0M to 1.2M (a reduction of 60%), and GFLOPs decreased from 8.1 to 4.4 (a reduction of 45%); the accuracy indicators P=97.4%, R=97.2%, mAP50=98.7%, and mAP50-95=82.4%, which are 1.1%, 2.1%, 0.7%, and 1.9% higher than YOLOv8, respectively, achieving a balance between accuracy and lightweight design.

[0033] 2. Loss Function Comparison and Validation Comparing the performance of the four loss functions GIoU, CIoU, EIoU, and SIoU in this model: The CIoU performance was optimal: mAP50=98.7%, mAP50-95=82.4%, P=97.4%, R=97.2%; GIoU is second best: P=97.8%, R=97.4%, which makes gradient propagation more stable in scenarios with small targets and weak overlap. EIoU and SIoU performed slightly worse, with mAP50-95 both below 80%, indicating that in the detection of small mosquito targets, a loss function that comprehensively considers the center point distance and aspect ratio is more suitable.

[0034] 3. Comparative experiments to verify (model superiority) Comparison with classic models such as YOLOv3, YOLOv5, YOLOv11, and SSD: Leading accuracy: The mAP50 (98.7%) of this model is higher than that of YOLOv8 (98.0%), YOLOv11 (98.0%), and YOLOv5 (97.4%), and the mAP50-95 (82.4%) is 1.9 percentage points higher than that of YOLOv8; Significant advantages of lightweight design: The number of parameters (1.2M) is only 40% of that of YOLOv8 and 1.9% of that of YOLOv3. GFLOPs (4.4) are reduced by 45% compared to YOLOv8 and are better than YOLOv11 (2.7M / 6.5GFLOPs). Faster convergence speed: mAP improves rapidly in the early stages of training, and its stability is better than other models in the later stages, achieving stable performance within 100 epochs.

[0035] 4. Validation of Feature Extraction Effectiveness (Visualization of Grad-CAM) Comparison using Grad-CAM heatmaps: The activation areas of this model are more concentrated and complete, and are closely aligned with the mosquito's body (key feature areas such as wings, antennae, and abdomen); Compared to YOLOv8, this model effectively suppresses background noise activation and has a stronger ability to focus on the core features of small targets, thus verifying the rationality of feature extraction.

[0036] A method for training a mosquito multimorphic recognition model includes the following steps: Step S1, mosquito image data acquisition: Collect mosquito image data, including the adult male and female morphology of Aedes, Culex and Anopheles mosquitoes, as well as the larval morphology of Aedes and Culex mosquitoes.

[0037] Step S2, Image Preprocessing and Data Augmentation: The collected mosquito images are preprocessed and data augmentation is performed by rotating, mirroring, adjusting brightness, and adding noise perturbation to improve the sample imbalance problem between mosquito categories.

[0038] Step S3, Data Labeling and Dataset Construction: The enhanced mosquito images are labeled in target detection format. The labeling content includes at least the location of the mosquito and its corresponding species, sex and developmental stage information, and a mosquito multimorphic recognition dataset is constructed.

[0039] Step S4, Model Training: Input the mosquito multimorphic recognition dataset into the above mosquito multimorphic recognition model for training, so that the model has the ability to jointly recognize multiple attributes such as mosquito species, sex and developmental stage, and obtain the trained mosquito multimorphic recognition model.

[0040] This method constructs a comprehensive dataset covering multiple species, sexes, and developmental stages, providing a comprehensive knowledge source for model learning. At the same time, by comprehensively utilizing various data augmentation techniques, it not only expands the amount of data but, more importantly, simulates the lighting changes, pose changes, and image quality degradation that may be encountered in real-world application scenarios. This allows the model to be exposed to and adapt to these complex situations from the training phase, thereby achieving excellent generalization performance and practical deployment results.

[0041] In some embodiments, the preprocessing in step S2 further includes Gaussian blur and motion blur. Gaussian blur is used to simulate the loss of imaging details caused by inaccurate focusing or shallow depth of field, while motion blur is used to simulate target blur caused by mosquito flight or camera shake during acquisition. This structure, by introducing these two image degradation simulations that closely resemble actual acquisition conditions, forces the model to learn more robust feature representations rather than over-relying on sharp image details. Simultaneously, this targeted data augmentation effectively compensates for the lack of high-quality labeled data, especially for mosquito samples that are most difficult to capture clearly in reality, such as those in flight or motion, improving the model's ability to identify "difficult samples"—an effect that is difficult to achieve through simple geometric transformations.

[0042] In some embodiments, the object detection format in step S3 is YOLO format. For example, the annotation file is in .txt text format, with each line containing the object category index and the normalized bounding box coordinates. This structure, using the YOLO format, facilitates the direct use of the training framework and evaluation tools of the YOLO series of algorithms, simplifying the data processing flow; at the same time, its compact text format also facilitates storage and transmission, providing convenience for the construction and management of large-scale datasets.

[0043] A method for identifying the multimorphic characteristics of mosquitoes, targeting insects such as... Figure 4 The image shows samples of various mosquito categories. It includes male and female samples of the three main disease-carrying mosquitoes, as well as larval morphology of Aedes and Culex mosquitoes. Data augmentation was used to improve class imbalance and enhance the model's generalization ability, including the following steps: Step A, Image / Video Data Input: Input the static image or video stream data of the mosquito to be identified into the terminal device that has been deployed with a trained mosquito multimorphic recognition model.

[0044] Step B, Model Inference and Recognition: The terminal device uses a mosquito multimorphic recognition model to extract features, fuse multi-scale features, and detect targets from the input image / video frame data, thereby completing the joint recognition of multiple attributes of the mosquito target, including species, sex, and developmental stage.

[0045] Step C, Identification Result Output: The terminal device outputs the location information of the mosquito target and the corresponding species, sex, and developmental stage identification results in real time.

[0046] In this method, users acquire images or videos of mosquitoes on-site through the camera of their mobile terminals. The identification process is completed entirely locally on the device without the need for a network connection, ensuring data privacy and real-time identification. At the same time, a complete result containing bounding boxes and multiple classification labels can be obtained through a single inference. The result is intuitive and clear, and can be directly used for on-site judgment or uploaded to the management platform as structured data, greatly improving work efficiency.

[0047] In some embodiments, the method further includes a model deployment step: converting the PyTorch format mosquito multimorphic recognition model file obtained from the training end into the ONNX universal inference format, and then into the NCNN format model file that can be loaded by the mobile inference engine; completing the project configuration and dependency integration in the Android Studio development environment, compiling and packaging to generate an application installation package, and deploying the NCNN format model file on the terminal device. In this structure, ONNX serves as an intermediate representation, bridging the gap between the training framework and various inference engines; NCNN is a high-performance neural network inference framework optimized for mobile devices, and its conversion process performs specific optimizations on the model operators; at the same time, this standardized deployment process ensures that the complexly trained deep learning model can run efficiently and stably on Android mobile devices, truly endowing advanced recognition capabilities to terminal applications.

[0048] In some embodiments, the dependency integration includes the introduction of the Ncnn-Android-Vulkan mobile inference framework and the OpenCV-Mobile mobile image processing library. Ncnn-Android-Vulkan is used for model loading and GPU-accelerated inference computation based on the Vulkan API, while OpenCV-Mobile is used for preprocessing such as reading image / video frames, scaling, and color space conversion, as well as visualization operations such as drawing recognition results (e.g., bounding boxes, label text) onto the image. In this architecture, Ncnn-Android-Vulkan fully leverages the parallel computing capabilities of mobile device GPUs to achieve rapid model inference; simultaneously, OpenCV-Mobile handles image I / O and rendering tasks, working collaboratively with NCNN to form the core technology stack of mobile recognition applications, ensuring overall application performance and user experience.

[0049] In some embodiments, the terminal device is a mobile terminal, and the mosquito multimorphic recognition model runs on the mobile terminal as an application, supporting both static image recognition mode and video stream recognition mode. In this architecture, the static image recognition mode is suitable for detailed analysis of collected sample photos; the video stream recognition mode supports real-time preview and identification of mosquitoes in the image via a camera, enabling dynamic monitoring. Seamless switching between the two modes allows this technical solution to be used for both laboratory sample identification and real-time field surveys, making its application scenarios extremely flexible.

[0050] Referring to the accompanying drawings, the working principle of this invention is as follows: First, through a convolutional neural network model architecture, this model efficiently extracts multi-level features from mosquito images using a backbone network, and fuses semantic and detailed information at different scales through a feature fusion network. In the detection head part, the model simultaneously performs target localization and fine-grained classification based on species, sex, and developmental stage. To train this model, a high-quality dataset covering various mosquito morphologies and attributes needs to be constructed, and data augmentation strategies simulating real-world scenarios are employed to improve the model's robustness. After training, the model is optimized and format-converted, making it deployable on mobile terminals such as smartphones. In practical applications, the terminal device calls the model to perform inference on input image or video frames, quickly outputting recognition results containing location and multiple attributes. Therefore, this invention, through the collaborative design of innovative model structure, optimized training strategies, and efficient mobile deployment, ultimately achieves an intelligent solution capable of accurately, quickly, and conveniently performing joint recognition of multiple mosquito morphologies and attributes in real-world scenarios.

[0051] Although specific embodiments of the invention have been described in detail herein, they are provided for illustrative purposes only and should not be construed as limiting the scope of the invention. Various substitutions, alterations, and modifications can be conceived without departing from the spirit and scope of the invention. The above description is merely a preferred embodiment of the invention and is not intended to limit the invention in any way. While the invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the invention, without departing from the scope of the invention, shall still fall within the scope of the invention.

Claims

1. A mosquito multi-modal recognition model, characterized in that, The model is a convolutional neural network model based on the improved object detection convolutional neural network, which is adapted to the differences in target size, appearance and texture features between adult mosquitoes and larvae. The model includes a backbone network, a feature fusion network and a detection head. The backbone network is configured with a Stem module, a multi-layer HGBlock and DWConv module, and an SPPF module. The Stem module is used to perform initial feature extraction and downsampling on the input image. The HGBlock and DWConv modules are used to extract semantic features at different levels step by step. The SPPF module is used to expand the receptive field and aggregate multi-scale contextual information. The feature fusion network is configured with an Upsample module, a Concat module, and a Rep-Ghost module. The Upsample and Concat operations are used to achieve top-down and bottom-up fusion of multi-scale features, and the Rep-Ghost module is used for efficient feature generation and fusion. The detection head sets up corresponding detection branches for multi-scale outputs of P3, P4, and P5, and is configured with an SCDH structure and a Conv_GN module. The SCDH structure is used to uniformly process the detection features, and the Conv_GN module is used to improve the stability of model training. The detection head finally outputs the mosquito target location and the corresponding species, sex, and developmental stage information. The improved model maintains high detection accuracy while reducing the number of parameters and computational load, and can achieve joint identification of multiple attributes such as mosquito species, sex and developmental stage.

2. A training method of a mosquito multi-form identification model, characterized in that, Includes the following steps: Step S1, mosquito image data acquisition: Collect mosquito image data, including the adult male and female morphology of Aedes mosquitoes, Culex mosquitoes and Anopheles mosquitoes, as well as the larval morphology of Aedes mosquitoes and Culex mosquitoes; Step S2, Image Preprocessing and Data Augmentation: The collected mosquito images are preprocessed and data augmentation is performed by rotating, mirroring, adjusting brightness and adding noise perturbation to improve the sample imbalance problem between mosquito categories. Step S3, Data annotation and dataset construction: The enhanced mosquito images are annotated in target detection format. The annotation content includes at least the location of the mosquito and the corresponding species, sex and developmental stage information, and a mosquito multimorphic recognition dataset is constructed. Step S4, Model Training: Input the mosquito multimorphic recognition dataset into the mosquito multimorphic recognition model for training, so that the model has the ability to jointly recognize multiple attributes such as mosquito species, sex and developmental stage, and obtain the trained mosquito multimorphic recognition model.

3. The method of claim 2, wherein the plurality of morphologies of the mosquito comprises a plurality of morphologies of a female mosquito and a plurality of morphologies of a male mosquito. The preprocessing described in step S2 also includes Gaussian blur and motion blur processing. Gaussian blur is used to simulate the loss of imaging details, and motion blur is used to simulate the motion blur caused by the relative motion of the target or camera during the acquisition process.

4. The method of claim 2, wherein the plurality of morphologies of the mosquito comprises a resting state, a flying state, a landing state, a feeding state, and a mating state. The target detection format described in step S3 is the YOLO format.

5. A method for identifying the multimorphic features of mosquitoes, characterized in that, Includes the following steps: Step A, Image / Video Data Input: Input the static image or video stream data of the mosquito to be identified into the terminal device that has been deployed with a trained mosquito multimorphic recognition model; Step B, Model Reasoning and Recognition: The terminal device uses a mosquito multimorphic recognition model to extract features, fuse multi-scale features, and detect targets from the input image / video frame data, thereby completing the joint recognition of multiple attributes of the mosquito target, including species, sex, and developmental stage. Step C, Identification Result Output: The terminal device outputs the location information of the mosquito target and the corresponding species, sex, and developmental stage identification results in real time.

6. The mosquito polymorphic recognition method according to claim 5, wherein, It also includes model deployment steps: converting the PyTorch format mosquito multimorphic recognition model file obtained from the training end into the ONNX general inference format, and then into the NCNN format model file that can be loaded by the mobile inference engine; completing the project configuration and dependency integration in the Android Studio development environment, compiling and packaging to generate the application installation package, and deploying the NCNN format model file on the terminal device.

7. The mosquito polymorphic recognition method according to claim 6, characterized in that, The dependency integration includes the introduction of the Ncnn-Android-Vulkan mobile inference framework and the OpenCV-Mobile mobile image processing library. Ncnn-Android-Vulkan is used to implement model loading and inference calculation, while OpenCV-Mobile is used to complete the reading, preprocessing, and visualization of recognition results of image / video frames.

8. The mosquito polymorphic recognition method according to any one of claims 5-7, wherein, The terminal device is a mobile terminal, and the mosquito multimorphic recognition model runs on the mobile terminal as an application, supporting static image recognition mode and video stream recognition mode.

9. A mosquito polymorphic recognition system characterized in that, It includes a data acquisition module, a data processing module, a model training module, a model deployment module, and a recognition application module; The data acquisition module is used to acquire mosquito image data of adult male and female morphology of Aedes, Culex, and Anopheles mosquitoes, as well as larval morphology of Aedes and Culex mosquitoes. The data processing module is used to preprocess and augment the collected mosquito images, and then to annotate the augmented images with target detection format to construct a mosquito multimorphic recognition dataset containing information on mosquito location, species, sex and developmental stage. The model training module is used to construct the mosquito multimorphic recognition model and train the model using the mosquito multimorphic recognition dataset to obtain the trained mosquito multimorphic recognition model. The model deployment module is used to complete the format conversion of the mosquito multimorphic recognition model and its deployment on the terminal device, so as to realize the inference operation of the model on the terminal device. The identification application module is used to receive mosquito image / video data to be identified, call the deployed model to complete inference identification, and output the mosquito identification results in real time.

10. The mosquito polymorphic identification system of claim 9, wherein, The identification application module is deployed on a mobile terminal, supports static image recognition mode and video stream recognition mode, and can simultaneously output the location information of mosquito targets and multi-attribute identification results such as species, sex, and developmental stage.

Images