A mounted aphid monitoring sensor device, monitoring method and management system for alfalfa factory inspection

Abstract

The application discloses a kind of for sprout grass factory inspection Carrying type aphid monitoring sensor device, monitoring method and management system, it is related to agricultural micro small pest detection technical field.Device includes macro image acquisition module, embedded computing module, communication module and control module;Macro image acquisition module acquires the aphid image about wheat seedling of target area, embedded computing module runs WheatAphid-YOLOv8n model trained and generates aphid detection result according to aphid image, communication module uploads result to cloud server;Control module responds to trigger signal, controls macro image acquisition module to take photograph and calls trained model to carry out aphid detection inference.The application is applicable to the image acquisition of aphid on the surface of plant in sprout grass factory, identification counting and data upload, can realize in situ monitoring and data management of wheat seedling live aphid in sprout grass factory under the premise of not relying on trap lamp or stick insect board.

Description

Technical Field

[0001] This invention relates to the field of agricultural micro-pest detection technology, and in particular to an onboard aphid monitoring sensor device, monitoring method and management system for inspection of sprout factories. Background Technology

[0002] Grass bud factories are characterized by high planting density per unit area and continuous batch production. However, due to the limitations of the tiered planting structure, the planting space is compact, requiring frequent manual inspections for pest monitoring during production. Wheat seedlings have slender, narrow leaves and overlapping branches and leaves, while aphids are tiny and mostly hidden in concealed areas such as the undersides of wheat seedling leaves, leaf sheaths, stems, and gaps between overlapping leaves, greatly increasing the difficulty of pest detection. Furthermore, there are significant visual interferences between aphid body color, environmental shadows, wheat leaf texture, and the background of the cultivation substrate, making it extremely difficult to distinguish target pests from the background. Currently, the industry mainly relies on manual, close-range, tray-by-tray inspections for pest monitoring. This method is not only extremely inefficient but also highly dependent on human experience, making it subjective and unable to generate standardized pest monitoring data that is continuous, accurately located, and fully traceable. This fails to meet the intelligent and precise pest control needs of grass bud factories.

[0003] Existing pest monitoring technologies have significant limitations in their adaptability to different scenarios. Most mainstream monitoring solutions rely on insect-attracting lamps, pheromone lures, specialized trapping devices, or sticky boards and tapes to collect pest images and count insect numbers. These solutions are only suitable for capturing and counting flying pests and counting insects on planar carriers, and cannot directly reflect the real-time insect density on the surface of live wheat seedlings. The monitoring data deviates significantly from the actual pest situation in the wheat germ factory. Furthermore, adhesive monitoring equipment is susceptible to environmental reflections, insect aggregation, insect damage, and regular equipment replacement and maintenance, resulting in poor monitoring stability and accuracy. Traditional grain pest detection technologies are designed for storage environments such as grain warehouses and grain piles. The monitoring background environment, pest survival status, and monitoring procedures are fundamentally different from the monitoring scenario of live wheat aphids in a wheat germ factory, making direct application impossible. Conventional methods for inspecting pests in field crops are mainly suitable for long-distance imaging in the field and large-scale panoramic monitoring of crop canopies. They are not suitable for the close-range macro imaging and high-precision detection of tiny pests required by plant factory inspection robots.

[0004] Existing research and applications of YOLO series pest detection algorithms mostly focus on optimizing and improving the model's own structure. They can only achieve a single pest image recognition and detection function, which cannot adapt to the full-process operation requirements of plant factory inspection robots. They are also unable to cover integrated operation links such as equipment triggering and acquisition, macro imaging control, intelligent supplemental lighting control, embedded inference, accurate pest counting, and cloud server data uploading.

[0005] Given the numerous technical shortcomings of the existing technologies, there is an urgent need in this field for an integrated technical solution that combines hardware sensing and acquisition, refined macro imaging, a specific micro-pest identification model for wheat seedling aphids, a standardized insect population counting process, and server data management. This solution would construct a complete process solution covering pest data collection, intelligent identification and counting, and cloud-based data management, thereby addressing the industry's technical pain points of low efficiency, poor accuracy, non-standardized data, and insufficient adaptability to various technical scenarios in current wheat aphid monitoring at the Yacao Factory. Summary of the Invention

[0006] In view of this, the present invention provides an onboard aphid monitoring sensor device, monitoring method and management system for inspection of wheat seedlings. Without relying on insect trapping equipment such as insect-attracting lamps or sticky traps, it is suitable for image acquisition, identification and counting and data uploading of aphids on the surface of plants in wheat seedlings factories, realizing in-situ monitoring and data management of live aphids in wheat seedlings in wheat seedlings factories.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: An onboard aphid monitoring sensor device for inspection of a grass factory includes a macro image acquisition module, an embedded computing module, a control module, and a communication module; wherein the macro image acquisition module, the embedded computing module, and the communication module are connected in sequence, and the control module is connected to the macro image acquisition module and the embedded computing module respectively; The macro image acquisition module is used to acquire images of aphids in wheat seedling leaves, leaf sheaths, stems, or locally overlapping areas in the target area; The embedded computing module is used to run the trained WheatAphid-YOLOv8n model and generate aphid detection results based on aphid images; The communication module is used to upload aphid detection results to the cloud server; The control module is used to respond to trigger signals, control the macro image acquisition module to take pictures, and call the trained WheatAphid-YOLOv8n model in the embedded computing module to perform aphid detection inference.

[0008] In the aforementioned device, optionally, the embedded computing module uses the WheatAphid-YOLOv8n model, which is based on the YOLOv8n model and combined with a high-resolution feature detection branch, a leaf background suppression attention mechanism, a fine-grained feature extraction module, a lightweight feature fusion structure, and a multi-scale feature fusion module.

[0009] The aforementioned device, optionally, includes a WheatAphid-YOLOv8n model comprising a backbone network, a neck network, and a detection head; The backbone network is used to extract shallow texture features and deep semantic features from standard wheat seedling images; the neck network is used to fuse shallow texture features and deep semantic features at different scales to obtain a wheat seedling feature fusion image; the detection head is used to identify aphids based on the wheat seedling feature fusion image and output the aphid category, location, and confidence level.

[0010] Optionally, the training process of the WheatAphid-YOLOv8n model in the aforementioned apparatus specifically includes: Data acquisition steps: Aphids of different population densities were introduced onto wheat seedlings in the experimental area, and RGB image data of wheat aphids in the experimental area were collected using an industrial camera; Data processing steps: Filter and label the RGB image data of wheat aphids, and divide the data according to a preset ratio to obtain the wheat aphid dataset; Model building steps: Build the WheatAphid-YOLOv8n model based on the YOLOv8n model to confirm the initial wheat seedling aphid recognition model; Model training steps: Based on the high-resolution input and small target enhancement training strategy, the initial wheat seedling aphid identification model was trained and tested using the wheat aphid dataset to confirm the model weights used for wheat seedling aphid identification. Model validation steps: Adjust and calibrate the initial wheat seedling aphid identification model based on the model weights to obtain a trained wheat seedling aphid identification model.

[0011] Optionally, in the training process of the WheatAphid-YOLOv8n model, the small-objective enhancement training strategy in the model training step includes at least one of image stitching, copy and paste, brightness perturbation, geometric perturbation, scale perturbation, or motion blur enhancement.

[0012] Optionally, in the above-mentioned apparatus, during the training process of the WheatAphid-YOLOv8n model, the wheat aphid dataset in the data processing step includes a training set, a validation set, and a test set; The specific steps of model training include: training the initial wheat seedling aphid identification model using the training set; adjusting parameters and selecting the model using the validation set; and evaluating the detection performance of the model using the test set. The detection performance includes at least one of the following: precision, recall, mAP50, mAP50-95, number of parameters, model size, computational cost, and inference speed.

[0013] The aforementioned device may optionally include a supplementary lighting module, a power supply module, and a storage module; the supplementary lighting module and the storage module are respectively connected to the control module, and the power supply module is respectively connected to the macro image acquisition module, the supplementary lighting module, and the embedded computing module; The supplementary lighting module is used to provide supplementary lighting when the macro image acquisition module acquires images; The power module supplies power to the macro image acquisition module, the supplementary lighting module, and the embedded computing module; The storage module is used to save aphid images, model files, or aphid detection results.

[0014] Optionally, the aphid detection results of the aforementioned device include at least one of the following: image number, acquisition time, sensor number, seedling tray number, aphid quantity, detection frame coordinates, and confidence level, and are uploaded to the server in the form of a data packet for the storage, statistics, or early warning of pest infestation during the wheat seedling growth stage in the target area.

[0015] A method for monitoring aphids on a mobile device for inspecting a grass factory, wherein the method utilizes a mobile aphid monitoring sensor device for inspecting a grass factory as described above to obtain aphid detection results in a target area.

[0016] A mounted aphid monitoring and management system for inspection of a seedling factory includes a mounted aphid monitoring sensor device for inspection of a seedling factory as described above, an inspection robot or a sliding rail inspection mechanism, and a server; the server is used to receive aphid detection results uploaded by the sensor device and to establish a pest data record according to the collection time, sensor number, robot position or seedling tray number.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention provides an onboard aphid monitoring sensor device, monitoring method and management system for inspection of grass processing plants. The present invention has the following beneficial effects: (1) In-situ real monitoring: Abandoning the traditional indirect monitoring method of trapping, the images of live aphids on wheat seedlings in the sprout factory are directly collected. This can provide real feedback on the real-time insect population status on the plant surface in high-density layered planting scenarios, realize in-situ accurate monitoring of aphids, and solve the problems of data distortion and incompatibility with live plant detection in traditional methods.

[0018] (2) Fine-grained anti-interference recognition: Based on the self-built wheat aphid RGB dataset, a WheatAphid-YOLOv8n exclusive model was constructed. To address the problems of aphids being small, having complex backgrounds, and being easy to miss, a P2 high-resolution detection branch, a lightweight attention module, a C2fAttn fine-grained feature extraction module, a LightNeck lightweight fusion structure, and a BiFPN-style multi-scale weighted fusion module were introduced to effectively improve the recognition accuracy of small aphids in complex backgrounds.

[0019] (3) Precision and Lightweight Two-Way Optimization: Through multi-module structural optimization, both detection performance and embedded deployment adaptability are taken into account. Ablation experiments show that compared with the original YOLOv8n baseline model, the precision of the E7_Final_Stable model of this invention is improved from 0.92583 to 0.93248, mAP50 from 0.95832 to 0.96064, and mAP50-95 from 0.68713 to 0.69101; at the same time, the number of parameters is reduced from 3,011,043 to 2,441,472, and the model size is compressed from 6.127 MB to 5.758 MB, with significant lightweight advantages, making it suitable for embedded terminal deployment.

[0020] (4) Full-process integrated: The functions of macro imaging acquisition, embedded intelligent recognition, aphid accurate counting and cloud data uploading are integrated into the same sensor device, opening up the entire monitoring link, solving the problem of fragmented operation in traditional solutions, and realizing the automated closed loop of insect monitoring.

[0021] (5) Highly adaptable continuous monitoring: The integrated equipment can be equipped with various inspection robots such as mobile and sliding rail robots, which are suitable for high-density, shelf-type and continuous production scenarios in the grass factory. It can realize unmanned continuous pest monitoring around the clock, effectively improving the intelligence and automation level of pest monitoring in the factory. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the structure of an onboard aphid monitoring sensor device for inspection of a grass factory, as disclosed in this invention. Figure 2 This is a schematic diagram of RGB images of wheat aphids collected according to an embodiment of the present invention; Figure 2(a) Images of wheat seedlings with aphids at the corresponding time, population density, and shooting angle, obtained based on aphid sample 1. Figure 2 (b) Images of wheat seedling aphids obtained from aphid sample 2 at the corresponding time, insect population density, and shooting angle. Figure 2 (c) Images of wheat seedling aphids obtained from aphid sample 3 at the corresponding time, insect population density, and shooting angle. Figure 2 (d) is an image of wheat seedling aphids obtained based on aphid sample 4 at the corresponding time, insect population density and shooting angle; Figure 3 This is an overall architecture diagram of the trained WheatAphid-YOLOv8n model disclosed in an embodiment of the present invention; Figure 4 These are three views of the inspection robot-mounted wheat aphid macro-identification sensor device disclosed in an embodiment of the present invention; Figure 4 (a) is a front view of the macro recognition sensor device. Figure 4 (b) is a side view of the macro recognition sensor device. Figure 4 (c) is a top view of the macro recognition sensor device; Figure 5 This is a schematic diagram of the working structure of the inspection robot-mounted wheat aphid macro-identification sensor device disclosed in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of the inspection robot-mounted wheat aphid macro recognition sensor device disclosed in this embodiment of the invention, mounted on a mobile robot and a slide rail; Figure 6 (a) A schematic diagram of a wheat aphid macro-identification sensor device mounted on a mobile robot for inspection; Figure 6 (b) A schematic diagram of the inspection robot equipped with a micro-range identification sensor for wheat aphids mounted on a slide rail; Figure 7 This is a schematic diagram illustrating the operation of the inspection robot equipped with a wheat aphid macro-identification sensor device for aphid monitoring, as disclosed in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0026] See Figure 1 As shown, this invention discloses an onboard aphid monitoring sensor device for inspection of a grass factory, comprising a macro image acquisition module, an embedded computing module, a control module, and a communication module; wherein the macro image acquisition module, the embedded computing module, and the communication module are connected in sequence, and the control module is connected to the macro image acquisition module and the embedded computing module respectively; The macro image acquisition module is used to acquire images of aphids in wheat seedling leaves, leaf sheaths, stems, or locally overlapping areas in the target area; The embedded computing module is used to run the trained WheatAphid-YOLOv8n model and generate aphid detection results based on aphid images; The communication module is used to upload aphid detection results to the cloud server; The control module is used to respond to trigger signals, control the macro image acquisition module to take pictures, and call the trained WheatAphid-YOLOv8n model in the embedded computing module to perform aphid detection inference.

[0027] Optional, Figure 1 The dashed line connection between the communication module and the cloud server indicates that the connection between the two is based on the method of data transmission or communication, specifically including wireless connection, Bluetooth connection, local area network interconnection, etc.; the solid line connection between other modules indicates that the modules are connected by standard electrical connections.

[0028] Optional, refer to Figure 1 The present invention discloses an onboard aphid monitoring sensor device for inspection of grass factories, which further includes a supplementary lighting module, a power supply module and a storage module; the supplementary lighting module and the storage module are respectively connected to the control module, and the power supply module is respectively connected to the macro image acquisition module, the supplementary lighting module and the embedded computing module.

[0029] Figure 1In the diagram, the dashed line between the supplementary lighting module and the macro image acquisition module indicates that, under the control of the control module, the supplementary lighting module provides additional illumination to the macro image acquisition module, enabling it to acquire images in a well-lit environment.

[0030] Optionally, in the embedded computing module, the WheatAphid-YOLOv8n model is constructed by combining the YOLOv8n base model with a high-resolution feature detection branch, a leaf background suppression attention mechanism, a fine-grained feature extraction module, a lightweight feature fusion structure, and a multi-scale feature fusion module.

[0031] Among them, the high-resolution feature detection branch is the P2 detection branch, which is used to improve the detection capability of tiny aphid targets by utilizing shallow high-resolution features; the leaf background suppression attention mechanism is used to enhance the feature response related to aphid targets and suppress the interference of wheat leaves, veins and shadow background; the fine-grained feature extraction module is set in the backbone network or feature fusion path to enhance the edge and local texture features of aphids; the lightweight feature fusion structure is set in the neck network to reduce redundant convolution calculations and the number of model parameters; the multi-scale feature fusion module is used to perform weighted fusion of features at different scales to improve the recognition stability of aphid targets of different sizes and postures.

[0032] Optionally, the WheatAphid-YOLOv8n model includes a backbone network, a neck network, and a detection head; The backbone network is used to extract shallow texture features and deep semantic features from standard wheat seedling images; the neck network is used to fuse shallow texture features and deep semantic features at different scales to obtain a wheat seedling feature fusion image; the detection head is used to identify aphids based on the wheat seedling feature fusion image and output the aphid category, location, and confidence level.

[0033] like Figure 3 As shown, WheatAphid-YOLOv8n is based on YOLOv8n. To address the issues of small size, dense distribution of aphids in wheat seedlings, and strong interference from the green leaf background, the model introduces a P2 high-resolution small target detection branch, which utilizes shallow high-resolution texture, edge, and local contour information.

[0034] The WheatAphid-YOLOv8n model also includes a lightweight attention module, SE / ECA, which enhances channel responses related to aphid bodies, edges, antennae, and local textures, while reducing interference from leaf textures, veins, shadows, and matrix background.

[0035] Specifically, such as Figure 3As shown, WheatAphid-YOLOv8n uses YOLOv8n as its base network, including a backbone feature extraction network, a leaf background suppression unit, a lightweight fusion network, a top-down feature transfer network, a fusion enhancement network, and a BiFPN weighted detection network. The input image size is 1280×1280×3. After feature extraction through convolutional and C2fAttn modules, features are generated at different scales (P2, P3, P4, and P5) with resolutions of 320×320, 160×160, 80×80, and 40×40, respectively.

[0036] The convolutional module consists of a two-dimensional convolutional layer, a batch normalization layer, and a SiLU activation function. The C2fAttn module includes a feature segmentation unit, at least one bottleneck feature extraction unit, and a SE / ECA lightweight attention unit, used to enhance the aphid's body edges, antennae, contours, and local texture features.

[0037] Features at scales P2, P3, P4, and P5 are input into the corresponding SE / ECA lightweight attention units, respectively. Channel response recalibration enhances aphid region features and suppresses interference from wheat leaf texture, veins, shadows, and matrix background. The attention-enhanced features at each scale are then input into the LightNeck module. The LightNeck module consists of GhostConv, depthwise separable convolutions, and 1×1 convolutions, designed to reduce the number of network parameters while preserving feature space information.

[0038] Furthermore, SE / ECA lightweight attention can be expressed as: ; in, As input features, For global average pooling, Indicates the channel interaction function. Here is the Sigmoid function, and ⊙ represents element-wise multiplication. This refers to the output characteristics after channel recalibration.

[0039] The WheatAphid-YOLOv8n model also includes a C2fAttn fine-grained feature extraction module, which enhances the expression of tiny aphid edges and local textures without changing the Detect head.

[0040] In the neck network, the WheatAphid-YOLOv8n model sets up a LightNeck lightweight fusion structure, using GhostConv and depthwise separable convolutions to replace some ordinary convolutions to reduce the number of neck parameters.

[0041] The WheatAphid-YOLOv8n model further incorporates a BiFPN-style stable fusion module, which performs learnable weighted fusion among P2, P3, and P4 multi-scale features to improve the coordination capability of multi-scale features.

[0042] In the top-down feature transfer process, low-resolution deep features are sequentially upsampled and concatenated with corresponding high-resolution shallow features, thereby transferring deep semantic information to scales P4, P3, and P2. During the fusion enhancement process, the concatenated features undergo fine-grained feature extraction via the C2fAttn fusion module and scale adjustment via convolutions with a stride of 2, forming a bottom-up feature enhancement path. This constitutes a bidirectional multi-scale feature transfer network between P2, P3, P4, and P5.

[0043] The bidirectionally fused P2, P3, P4, and P5 scale features are input into the BiFPN weight allocation unit. The BiFPN weight allocation unit uses non-negative learnable weights to normalize and weight the input features from different sources. The specific expression for its fusion process is as follows: The weighted fusion process of the BiFPN-style stable fusion module can be represented as: ; in, This represents the i-th input feature participating in the fusion of the current node. This represents the corresponding learnable fusion weights. This represents a stable term, specifically a stable term to prevent the denominator from being zero. This represents the output feature after weighted fusion. This formula is used to achieve adaptive weight allocation among features at multiple scales.

[0044] Combination Figure 3 As shown, the weighted fusion features P2, P3, P4, and P5 are input into the detection module. Specifically, the P2 detection branch is mainly used to identify tiny aphids, the P3 branch is mainly used to identify small-scale aphids, the P4 branch is used to process medium-scale targets, and the P5 branch is used to provide larger-scale targets and contextual information. The outputs of each detection branch are then aggregated to obtain the aphid detection box, confidence score, location coordinates, and number of detections.

[0045] Optionally, the training process of the WheatAphid-YOLOv8n model specifically includes: 1) Data acquisition steps: Aphids of different population densities were introduced onto wheat seedlings in the experimental area, and RGB image data of wheat aphids in the experimental area were acquired using an industrial camera; Specifically, the aphids infesting wheat seedlings include the cereal aphid and the wheat forked aphid. In training the model, this invention utilizes a SONY α7IV camera equipped with a FE 2.8 / 90 MACRO G OSS macro lens to collect RGB images of aphids at different growth stages, aphid densities, and shooting angles.

[0046] In addition, such as Figure 2 As shown, the experimental area was an ideal region selected by researchers from a sprout factory or created by simulating the sprout production environment, for planting wheat. After the wheat seedlings grew to a suitable stage, aphids were inoculated onto the seedlings to simulate the wheat seedling production conditions under different aphid distribution densities. Figure 2 (a) Figure 2 (b) Figure 2 (c) Figure 2 (d) Images of wheat seedling aphids at different growth stages, population densities, and shooting angles are shown. After the aphids formed different distribution patterns on the leaves, stems, leaf sheaths, and locally overlapping areas, high-resolution RGB images of the aphids were acquired using a SONY α7IV camera equipped with a FE 2.8 / 90 MACROG OSS macro lens. During the RGB image acquisition process, wheat seedling aphid images were collected at different growth stages, shooting angles, population densities, leaf postures, background interference, and shooting distances. The collected objects included the front of the wheat seedling leaves, leaf edges, leaf sheaths, stems, overlapping areas of leaves, and locally shaded areas to improve the dataset's coverage of the actual shooting scenes at the Sprout Factory.

[0047] 2) Data processing steps: Filter and label the RGB image data of wheat aphids, and divide the data according to a preset ratio to obtain the wheat aphid dataset; Optionally, during the training of the WheatAphid-YOLOv8n model, the wheat aphid dataset in the data processing step includes a training set, a validation set, and a test set; Similar to S2, the training of the WheatAphid-YOLOv8n model requires the removal of severely out-of-focus, abnormally exposed, or unidentifiable aphid targets from the input RGB image data of wheat aphids. LabelImg annotation software is used to annotate the aphids in the images with rectangular boxes, and the annotated RGB images and labels are divided according to the training set: validation set: test set = 8:1:1.

[0048] In the current embodiment, in a certain experiment, the collected images were screened to remove images that were severely out of focus, abnormally exposed, or whose targets were unidentifiable, ultimately forming a dataset of 3007 wheat seedling aphid images. After obtaining the dataset, the LabelImg software was used to annotate the aphids in each image with rectangular boxes to generate YOLO format label files. Subsequently, the dataset was divided into training set, validation set, and test set, and uploaded to the server for model training.

[0049] The server training environment consisted of an Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz, 502GiB of total system memory, and an NVIDIA GeForce RTX 3090 Ti GPU with driver version 595.71.05. The software environment included CUDA version 12.1, PyTorch version 2.5.1+cu121, and Ultralytics version 8.4.53. Based on this hardware and software environment, model training, validation, and ablation experiments were conducted on a self-built wheat aphid RGB image dataset to obtain an improved YOLOv8n model suitable for aphid identification in wheat sprout factories.

[0050] 3) Model building steps: Build the WheatAphid-YOLOv8n model based on the YOLOv8n model to confirm the initial wheat seedling aphid identification model; Alternatively, researchers built the WheatAphid-YOLOv8n model based on YOLOv8n, combining a high-resolution feature detection branch for identifying tiny aphids, a leaf background suppression attention mechanism, a fine-grained feature extraction module, a lightweight feature fusion structure, and a multi-scale feature fusion module.

[0051] 4) Model training steps: Based on the high-resolution input and small target enhancement training strategy, the initial wheat seedling aphid identification model was trained and tested using the wheat aphid dataset to confirm the model weights used for wheat seedling aphid identification. After constructing the WheatAphid-YOLOv8n model, the model was trained and tested using the wheat aphid dataset to obtain the model weights for identifying wheat seedling aphids, and the aphid detection boxes, confidence scores, and count results were output.

[0052] Optionally, the model training steps specifically include: training the initial wheat seedling aphid identification model using the training set; adjusting parameters and selecting the model using the validation set; and evaluating the model's detection performance using the test set, whereby the detection performance includes at least one of precision, recall, mAP50, mAP50-95, number of parameters, model size, computational cost, and inference speed.

[0053] Specifically, during the model training phase, the detection loss function can be expressed as: ; in, For bounding box regression loss, For classifying losses, For the distribution focus loss, , , These are the corresponding weighting coefficients.

[0054] Optionally, during the training of the WheatAphid-YOLOv8n model, Small-objective enhancement training strategies in the model training steps include at least one of image stitching, copy and paste, brightness perturbation, geometric perturbation, scale perturbation, or motion blur enhancement.

[0055] During model training, a small-objective augmentation training strategy was employed, including mosaic, mixup, copy-paste, and slight geometric perturbations, to improve the model's generalization ability under different insect population densities, leaf postures, lighting conditions, and local occlusion scenarios. The model retains the Ultralytics native Detect head throughout the entire process, avoiding head replacement to reduce deployment risks and inconsistencies between training and inference.

[0056] 5) Model verification step: Adjust and calibrate the initial wheat seedling aphid identification model based on the model weights to obtain a trained wheat seedling aphid identification model.

[0057] As an optional implementation, in the current embodiment of the present invention, the trained wheat seedling aphid recognition model, namely the trained WheatAphid-YOLOv8n model, is deployed in the inspection robot's onboard macro recognition sensor device. The sensor device takes close-up photos of the wheat seedlings to identify the number and location of aphids, and uploads the recognition results to the server. The sensor device will be described in detail in subsequent embodiments of the present invention.

[0058] The following experiments verify and confirm the model comparison results based on the training process of the WheatAphid-YOLOv8n model disclosed in the above embodiments of the present invention.

[0059] A dataset of 3007 wheat seedling aphid images was obtained using the above dataset construction process, and YOLO format annotation was performed using LabelImg. A YOLOv8n baseline and a series of improved models were trained on a server. The server hardware and software environment were as follows: CPU: Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz, total system memory: 502GiB; GPU: NVIDIA GeForce RTX 3090 Ti, graphics driver version: 595.71.05; Software environment: CUDA version: 12.1, PyTorch version: 2.5.1+cu121, Ultralytics version: 8.4.53.

[0060] The ablation experiments were set up from E0 to E7; E0_baseline was the original YOLOv8n; E1 introduced the P2 small objective branch; E2 added SE attention on the basis of P2; E3 introduced C2fAttn; E4 introduced a lightweight neck; E5 added BiFPN-style fusion; E6 was a candidate for Aug and WIoU enhancement strategies, but not the core innovation of this invention; E7_Final was the final stable combined model, i.e., the trained wheat seedling aphid recognition model. Table 1 lists the ablation experiment indicators most relevant to the model improvement of this invention.

[0061] Table 1 Key Results of Relevant Ablation Experiments

[0062] As can be seen from the data presented in Table 1, compared with E0_baseline, E7_Final improved mAP50 by 0.00232, mAP50-95 by 0.00388, and precision by 0.00665; the number of parameters decreased from 3.011M to approximately 2.441M, a reduction of approximately 18.92%; and the model file size decreased from 6.127MB to approximately 5.758MB, a reduction of approximately 6.02%. Therefore, the WheatAphid-YOLOv8n model used in this invention, after training and deployment, improves detection accuracy, localization quality, and false detection control capabilities while reducing the number of parameters and model size in the recognition of wheat seedling aphid images.

[0063] Furthermore, as shown in Table 2, this invention also discloses and lists the comparison results of the YOLO series models most relevant to the positioning of this invention. Table 2 does not list all models, but selects the basic model, the final model, and the representative high-precision comparison model to illustrate the technical positioning of the onboard aphid monitoring sensor device for inspection of grass processing plants disclosed in this invention.

[0064] Table 2 Comparison of WheatAphid-YOLOv8n Model with Representative YOLO Model

[0065] As shown in Table 2, some YOLO models achieve higher mAP50 or mAP50-95 values. For example, YOLOv11n and YOLOv9t have higher mAP50 performance. However, this invention does not claim that the model accuracy is superior to all YOLO models. Instead, it emphasizes the advantages of this invention in scene adaptability, micro-pest identification process, dataset construction method, parameter quantity and model volume control, counting process and device integration in the micro-aphid macro recognition sensor for wheat seedlings in the Yacao Factory.

[0066] Optionally, the aphid identification results in S3 include at least one of the following: image number, acquisition time, sensor number, seedling tray number, aphid quantity, detection box coordinates, and confidence level. These results are uploaded to the server in the form of a data packet for storing, counting, or issuing early warnings about pest infestations during the wheat seedling growth stage in the target area.

[0067] Specifically, such as Figure 7 As shown, the operation process of the aphid monitoring sensor device for inspection of a grass factory disclosed in this invention can be regarded as a server data uploading and pest management process. (Refer to...) Figure 7 The detection results are packaged into JSON format; the JSON data includes at least the image number, shooting time, robot position, seedling tray number, aphid count, average confidence level, detection box coordinates, original image path, visualization result path, sensor number, and model version number.

[0068] After the test results are packaged, the JSON data is uploaded to the server. Upon receiving the data, the server establishes a pest database based on the seedling tray number, robot location, and time. The server can generate pest population change curves for the same seedling tray at different times, as well as pest population distribution records for different seedling trays within the same inspection batch. When the aphid population exceeds a preset threshold, or when the pest population growth rate exceeds a preset threshold during consecutive inspections, the server outputs a pest alert.

[0069] based on Figure 7 As shown, an embodiment of a wheat aphid macro recognition sensor device mounted on an inspection robot disclosed in this invention includes a robot mounting interface module, a macro image acquisition module, a supplementary lighting module, an embedded computing module, a control module, a communication module, a power supply module, a storage module, and a housing; wherein, the macro image acquisition module, the supplementary lighting module, the embedded computing module, the control module, the communication module, the power supply module, and the storage module are disposed in the housing; The robot mounting interface module is located on the outer shell and is used to mount the sensor device to the mobile inspection robot, the sliding rail inspection mechanism or the fixed inspection bracket. The macro image acquisition module is used to acquire images of aphids on wheat seedling leaves, leaf sheaths, stems, or locally overlapping areas; the supplementary lighting module is used to provide supplementary lighting during image acquisition; and the embedded computing module is used to run the WheatAphid-YOLOv8n model and generate aphid detection results. The control module is used to receive trigger signals, control the taking of pictures, control the supplementary lighting, and call the model inference; the communication module is used to upload the detection results to the server; the power supply module is used to supply power to the macro image acquisition module, the supplementary lighting module, and the embedded computing module; and the storage module is used to save images, model files, or detection results.

[0070] The control module is also used to assess the quality of the acquired aphid images. Specifically, the control module obtains a sharpness evaluation value by calculating the Laplacian variance of the image transmitted by the image acquisition module, and obtains an exposure evaluation value by calculating the average grayscale value of the image. When either the sharpness evaluation value or the exposure evaluation value does not reach the preset range, the control module re-acquires the image. When the image quality meets the preset requirements, the image is input into the aphid recognition model.

[0071] Optionally, the robot mounting interface module adopts a universal quick-release mounting plate and is equipped with M3 / M4 screw holes, oblong holes, T-slots or round rod clamping structures, and pitch angle adjustment structures. Since the structures of different inspection robots or rail-mounted robots may not be consistent, this invention does not limit the robot body model equipped with the sensor device. Instead, it ensures that the sensor can be adapted to different robot platforms through universal multi-hole positions, quick-release plates, and pitch angle adjustment components.

[0072] Optionally, the macro image acquisition module includes an image sensor and an adjustable-focus macro lens. The image sensor is a SONY IMX477 12.3MP image sensor or an equivalent 12MP or higher image sensor, and the adjustable-focus macro lens is a 16mm macro lens or a macro lens suitable for working distances of 15-35cm.

[0073] Optionally, the robot mounting interface module includes at least one of the following: quick-release mounting plate, multi-hole connecting plate, elongated adjustment hole, slide rail connector, round rod clamping component, or pitch angle adjustment component, to adapt to different models of inspection robots or slide rail inspection mechanisms.

[0074] Optionally, the supplementary lighting module is a ring-shaped LED supplementary lighting module or a double-sided strip LED supplementary lighting module, and the illumination area of ​​the supplementary lighting module covers the imaging field of view of the macro image acquisition module.

[0075] The hardware structure of the sensor device disclosed in the above embodiments will be further described below.

[0076] In a preferred embodiment, such as Figure 4 The image shows a three-view drawing of a wheat aphid macro-identification sensor device mounted on an inspection robot, as disclosed in this embodiment of the invention. Figure 4 (a) is a front view of the macro recognition sensor device. Figure 4 (b) is a side view of the macro recognition sensor device. Figure 4 (c) is a top view of the macro recognition sensor device. (Refer to...) Figure 4 The device uses a small rectangular housing with a camera window and a 6500K white ring LED at the front.

[0077] Specifically, the image sensor is a 12MP or higher image sensor, and in a preferred embodiment, an internally fixed Sony IMX477 12.3MP image sensor is used; the macro lens is an adjustable-focus macro lens suitable for a working distance of 15-35 cm, and in a preferred embodiment, a 16 mm C / CS interface adjustable-focus macro lens is used, with an imaging field of view of 5cm×4cm to 8cm×6cm; this field of view can ensure that aphids on wheat seedlings have sufficient pixel size in the image, while covering local leaf, leaf sheath and stem areas.

[0078] The supplemental lighting module uses a 4500K-6500K white LED supplemental lighting module. In a preferred embodiment, a 6500K white ring LED is used, and the brightness is adjusted by PWM and turned on synchronously at the moment of taking a picture. This design is used to reduce the effects of light flicker from plant factory lights, shelf obstruction, leaf reflection, and shadows caused by robot movement.

[0079] The embedded computing module uses a Jetson Orin Nano 8GB processor. After receiving a trigger signal from the inspection robot, the control module controls the camera to take pictures, controls the supplementary lighting, calls the model for inference, and generates the detection results. The communication module preferably uses Wi-Fi 6 or Gigabit Ethernet, and is also compatible with 4G / 5G communication. The power module receives a 12V or 24V power input from the inspection robot, which is converted via DC-DC to 5V / 3A and 12V LED power, respectively powering the camera, computing module, and supplementary lighting module. The casing has pre-drilled ventilation holes or a fan to reduce the temperature of the embedded computing module during continuous inference.

[0080] The process of taking pictures and recognizing images using the sensor transposition disclosed in the above embodiments on the inspection robot will continue to be explained.

[0081] like Figure 5 and Figure 6 As shown, the inspection robot moves along the path of the seedling trays or planting racks in the Sprout Factory. Figure 6 (a) Figure 6(b) are schematic diagrams of the inspection robot-mounted wheat aphid macro recognition sensor device mounted on the mobile robot and the slide rail. When the inspection robot reaches the preset collection position, the robot control system sends a trigger signal to the control module of the sensor device. After receiving the trigger signal, the control module controls the supplementary lighting module to turn on and controls the Sony IMX477 macro camera to collect local images of wheat seedlings at a working distance of about 20 cm.

[0082] The embedded computing module preprocesses the image and then inputs the preprocessed image into the WheatAphid-YOLOv8n model. The model outputs aphid detection boxes, confidence scores, and location coordinates. The control module removes low-confidence detection boxes based on a preset confidence threshold, removes duplicate boxes based on the intersection-over-union ratio (IoU) or non-maximum suppression (NMS) strategy, and obtains the aphid count result for the current image based on the number of filtered detection boxes.

[0083] This invention also discloses a mounted aphid monitoring method for inspection of a grass sprout factory, which uses a mounted aphid monitoring sensor device for inspection of a grass sprout factory as described in any of the above embodiments to obtain aphid detection results in a target area; the monitoring method includes the following steps: Image acquisition is performed using a macro image acquisition module. During image acquisition, the control module adjusts the supplementary lighting module to provide adaptive lighting, ensuring that the macro image acquisition module can acquire images of wheat seedling aphids in a well-lit environment. After acquisition, the images are input into an embedded computing module. The pre-trained WheatAphid-YOLOv8n model embedded in the module monitors aphids using the input images and outputs aphid detection results. The detection results output by the model are uploaded to the server via a communication module, and the server provides corresponding pest warnings based on the detection results.

[0084] Optionally, in the mounted aphid monitoring method, images of wheat seedlings in the target area are acquired using a SONY α7IV camera equipped with a FE 2.8 / 90 MACRO GOSS macro lens.

[0085] Optionally, the image recognition requirements in S2 specifically include shooting focal length standards, image exposure standards, and image sharpness requirements; S2 specifically includes the following steps: S21 Removal Step: Based on the shooting focal length standard, image exposure standard, and image sharpness requirements, remove images of wheat seedlings that are out of focus, abnormally exposed, or have unidentifiable aphid targets to obtain identifiable wheat seedling images. S22 annotation steps: Use LabelImg annotation software to annotate the aphids in the identifiable wheat seedling image with rectangular boxes to obtain a standard wheat seedling image.

[0086] Furthermore, the present invention also discloses a pest data management system for wheat aphids in a seedling factory, including a sensor device, an inspection robot or a sliding rail inspection mechanism, and a server based on the above embodiments; the server is used to receive the detection results uploaded by the sensor device, and establish pest data records according to the collection time, sensor number, robot position or seedling tray number; the server is also used to generate pest population change records based on the number of aphids obtained from continuous inspections, and output early warning information when the number of aphids or the pest population growth rate reaches a preset threshold.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. 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 the invention. Therefore, the invention 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. An onboard aphid monitoring sensor device for inspection of a grass processing plant, characterized in that, It includes a macro image acquisition module, an embedded computing module, a control module, and a communication module; wherein the macro image acquisition module, the embedded computing module, and the communication module are connected in sequence, and the control module is connected to the macro image acquisition module and the embedded computing module respectively; The macro image acquisition module is used to acquire images of aphids in wheat seedling leaves, leaf sheaths, stems, or locally overlapping areas in the target area; The embedded computing module is used to run the trained WheatAphid-YOLOv8n model and generate aphid detection results based on aphid images; The communication module is used to upload aphid detection results to the cloud server; The control module is used to respond to trigger signals, control the macro image acquisition module to take pictures, and call the trained WheatAphid-YOLOv8n model in the embedded computing module to perform aphid detection inference.

2. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 1, characterized in that, In the embedded computing module, the WheatAphid-YOLOv8n model is built by combining the YOLOv8n base model with a high-resolution feature detection branch, a leaf background suppression attention mechanism, a fine-grained feature extraction module, a lightweight feature fusion structure, and a multi-scale feature fusion module.

3. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 1, characterized in that, The WheatAphid-YOLOv8n model consists of a backbone network, a neck network, and a detection head. The backbone network is used to extract shallow texture features and deep semantic features from standard wheat seedling images; The neck network is used to fuse shallow texture features and deep semantic features at different scales to obtain a fused image of wheat seedling features; the detection head is used to identify aphids based on the fused image of wheat seedling features and output the aphid category, location and confidence level.

4. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 1, characterized in that, The training process of the WheatAphid-YOLOv8n model specifically includes: Data acquisition steps: Aphids of different population densities were introduced onto wheat seedlings in the experimental area, and RGB image data of wheat aphids in the experimental area were collected using an industrial camera; Data processing steps: Filter and label the RGB image data of wheat aphids, and divide the data according to a preset ratio to obtain the wheat aphid dataset; Model building steps: Build the WheatAphid-YOLOv8n model based on the YOLOv8n model to confirm the initial wheat seedling aphid recognition model; Model training steps: Based on the high-resolution input and small target enhancement training strategy, the initial wheat seedling aphid identification model was trained and tested using the wheat aphid dataset to confirm the model weights used for wheat seedling aphid identification. Model validation steps: Adjust and calibrate the initial wheat seedling aphid identification model based on the model weights to obtain a trained wheat seedling aphid identification model.

5. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 4, characterized in that, During the training of the WheatAphid-YOLOv8n model, the small-objective enhancement training strategies in the model training steps include at least one of image stitching, copy and paste, brightness perturbation, geometric perturbation, scale perturbation, or motion blur enhancement.

6. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 4, characterized in that, During the training of the WheatAphid-YOLOv8n model, the wheat aphid dataset in the data processing step includes a training set, a validation set, and a test set. The specific steps of model training include: training the initial wheat seedling aphid identification model using the training set; adjusting parameters and selecting the model using the validation set; and evaluating the detection performance of the model using the test set. The detection performance includes at least one of the following: precision, recall, mAP50, mAP50-95, number of parameters, model size, computational cost, and inference speed.

7. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 1, characterized in that, It also includes a lighting module, a power module, and a storage module; the lighting module and the storage module are respectively connected to the control module, and the power module is respectively connected to the macro image acquisition module, the lighting module, and the embedded computing module; The supplementary lighting module is used to provide supplementary lighting when the macro image acquisition module acquires images; The power module supplies power to the macro image acquisition module, the supplementary lighting module, and the embedded computing module; The storage module is used to save aphid images, model files, or aphid detection results.

8. The mounted aphid monitoring sensor device for inspection of a grass processing plant according to claim 1, characterized in that, The aphid detection results include at least one of the following: image number, collection time, sensor number, seedling tray number, aphid count, detection frame coordinates, and confidence level. These results are uploaded to the server in the form of a data packet for storing, statistically analyzing, or issuing early warnings about aphid infestations during the wheat seedling growth stage in the target area.

9. A mounted aphid monitoring method for inspection of a grass processing plant, characterized in that, The aphid detection results of the target area are obtained by using the onboard aphid monitoring sensor device for inspection of a grass factory as described in any one of claims 1-8.

10. A mounted aphid monitoring and management system for inspection of a grass processing plant, characterized in that, The invention includes a mounted aphid monitoring sensor device, an inspection robot or a sliding rail inspection mechanism, and a server for inspecting a seedling factory, as described in any one of claims 1 to 8; the server is used to receive aphid detection results uploaded by the sensor device and to establish a pest data record according to the collection time, sensor number, robot position or seedling tray number.

Images