Aphid honeydew discharge behavior identification method based on spatio-temporal feature adaptive fusion

Abstract

The invention discloses an aphid honeydew discharge behavior identification method based on spatio-temporal feature adaptive fusion. The method comprises the following steps: collecting video data of aphids in a natural feeding environment; performing spatio-temporal feature extraction on the original video data by adopting a time sequence weighted frame difference adaptive fusion framework to obtain a motion feature map, inputting the motion feature map into the trained FGC-YOLO small target fine behavior detection model for detection, and outputting a box-selected aphid behavior category result identifier; and carrying out post-processing on an output result of the FGC-YOLO small target fine behavior detection model, and realizing honeydew discharge behavior identification through a dynamic double-cache interpolation mechanism and a composite behavior dynamic judgment algorithm. The method is high in detection precision and good in real-time performance, effectively solves the problems of low aphid honey dew analysis efficiency, poor real-time performance, complex operation and the like through spatial-temporal feature fusion and multi-module cooperation, and is expected to provide technical support for aphid feeding behavior monitoring and pest prevention and control.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and agricultural pest monitoring, and in particular to an intelligent recognition method for aphid honeydew discharge behavior based on target detection and adaptive fusion of spatiotemporal features. Background Art

[0002] Aphids are important pests of agricultural and forestry crops and one of the most important vectors for spreading plant viruses. Their honeydew-excreting behavior during feeding contains important information about the interaction between insects and plants, providing an important basis for research on plant resistance mechanisms and dynamic monitoring of pest occurrence.

[0003] Currently, the characterization of aphid honeydew secretion behavior mainly relies on manual visual statistics or chemical analysis. These methods are inefficient and have poor real-time performance. The development of machine vision and artificial intelligence technology in recent years has provided technical support for the non-destructive and real-time monitoring of honeydew secretion behavior.

[0004] Most behavioral recognition research focuses on larger targets such as humans and livestock. There is relatively little research on the behavioral recognition of small insects, mainly including the larval wriggling and grooming behavior of fly insects. However, there are currently no relevant literature reports on the behavioral recognition of aphids.

[0005] Aphids are small, their honeydew excretion behavior is subtle, and the target features in video data are weak, making small targets difficult to detect. Furthermore, the translucency of aphids, the similarity of leaves and backgrounds, and the clumping of aphids further complicate behavioral recognition. Furthermore, traditional object detection algorithms, such as the YOLO series of models, have limited ability to represent fine-grained actions in video temporal action detection, making it difficult to effectively capture the subtle changes and temporal dynamics of honeydew excretion behavior. This makes existing behavior recognition methods face numerous challenges in detecting aphid honeydew excretion behavior. Summary of the Invention

[0006] Aiming at the problems of low aphid honeydew detection efficiency, poor real-time performance, and difficulty in identifying small targets in the existing technology, the present invention proposes a method for identifying aphid honeydew excretion behavior based on adaptive fusion of spatiotemporal features.

[0007] The object of the present invention is achieved in the following manner: A method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features, the method comprising: S1: Collect video data of aphids in their natural feeding environment; S2: Use the time-series weighted frame difference adaptive fusion framework to extract spatiotemporal features from the original video data to obtain a motion feature map, including: S21: Processing continuous video frames using a sliding time window mechanism, calculating the time-order weighted frame difference to obtain a frame difference image; S22: Perform weighted superposition on all frame differences within the time window to obtain an initial motion map; S23: Adaptive threshold processing based on statistical features to obtain a motion saliency map; S24: performing Gaussian filtering to remove noise from the motion saliency map, and adaptively fusing it with the original image to obtain a motion feature map; S3: Combined with prior information on aphid behavior categories, the original video data after spatiotemporal feature extraction is used to mark the aphid's location and behavior type, and a fine-grained behavior dataset containing three behavior categories: movement, kicking, and honeydew discharge; Input the dataset into the FGC-YOLO small target fine behavior detection model for training and cross-validation, and output the trained FGC-YOLO small target fine behavior detection model; The FGC-YOLO small target fine behavior detection model uses the YOLOv11 model as the baseline network, including the backbone network, neck network and head network. The FGC-YOLO small target fine behavior detection model uses the fine-grained bidirectional feature pyramid network FGBiFPN as the neck network. The neck network introduces the global to local spatial aggregation module GLSA. The local window attention mechanism CLWA module integrated with C3k2 is used to replace C3k2 as the node processing unit. S4: Input the motion feature map into the trained FGC-YOLO small target fine behavior detection model for detection, and output the result identification of the aphid behavior category selected by the frame; S5: Post-process the output results of the FGC-YOLO small target fine behavior detection model, and realize the recognition of honeydew discharge behavior through the dynamic double buffer interpolation mechanism and the composite behavior dynamic judgment algorithm.

[0008] The S2 specifically includes: Temporal motion feature extraction: Using a sliding time window mechanism to process continuous video frames, let the input sequence be , the sliding window size is T, then the output frame difference It can be calculated by formula (1); In the formula is the time series weight coefficient; Motion saliency calculation: perform weighted superposition of all frame differences within the time window to obtain the initial motion map , as shown in formula (2); (2) In order to effectively suppress background noise, an adaptive threshold based on statistical characteristics is introduced , motion saliency map Calculated by formula (3); In the formula , and Mapping The mean and standard deviation of Feature optimization fusion: To obtain a smooth feature representation, the saliency map Gaussian filtering is performed to obtain the denoised saliency map, which is adaptively fused with the original image I and calculated by formula (4); In the formula is the final output motion feature map, is a color representation of motion saliency, is the adaptive mixing coefficient based on motion intensity, .

[0009] The fine-grained bidirectional feature pyramid network (FGBiFPN) in S3 improves the neck network structure in YOLOv11 by replacing C3k2 with the CLWA module as the node processing unit. It enhances local feature extraction capabilities through three cyclic stacking operations and standardizes the number of channels in each feature layer. The feature fusion process of FGBiFPN is expressed by formula (5); In the formula Y represents the output features, For the input features, For the corresponding weight, by initial learnable parameters conduct Activate to get; is a correction factor used to ensure the stability of the value.

[0010] The global to local spatial aggregation module GLSA in S3 adopts a dual-branch design, including a local spatial attention branch and a global spatial attention branch, which are used to extract local detail features and global structural information respectively, and then the outputs of the two branches are spliced ​​and fused along the channel dimension.

[0011] The CLWA module in S3 integrates LWA into the parallel convolution path of C3k2 and replaces the original Bottleneck structure with the LWABottleneck structure. In each parallel branch of C3k2, the LWABottleneck module adaptively adjusts the feature weights through the LWA attention mechanism, making the model more focused on key information.

[0012] The dynamic double buffer interpolation mechanism in S5 adopts a double buffer structure of odd-even frame division. When the number of consecutive undetected times is less than 2, a linear interpolation method is used to generate the current frame detection result.

[0013] The dynamic determination algorithm of the composite behavior in S5 includes: Behavior association determination: When kicking behavior is detected within the set time window, it is preliminarily determined that there is a possibility of honeydew discharge; Spatial proximity determination: After confirming that the aphid is kicking, the space around the aphid is further inspected to determine whether there is a "honeydew ball" target located within the aphid's kicking behavior frame; Identification of composite behaviors: When both the behavioral association and spatial proximity conditions are met, the aphid individual and the honeydew ball are combined into the honeydew excretion category to characterize this composite movement behavior of the aphid; Result storage and visualization: After the determination of aphid honeydew discharge behavior is completed, the video and images containing the behavior information are saved.

[0014] Beneficial effects of the present invention: The present invention constructs a small-target fine behavior detection model to achieve intelligent recognition of aphid honeydew excretion behavior, which has the following advantages: 1. High detection accuracy: The proposed TWFDAFF framework effectively extracts the spatiotemporal motion characteristics of aphids. The FGC-YOLO model enhances small target detection capabilities through multi-module optimization, achieving an average accuracy of 81.5% and an accuracy of 99.9% for identifying honeydew discharge behavior. 2. Good real-time performance: The average detection speed is 65 frames per second, meeting the needs of real-time monitoring. The parameter volume is 18.8M and the floating-point calculation volume is 84.3G, taking into account both accuracy and efficiency.

[0015] 3. Strong adaptability: Through the fusion of spatiotemporal features and multi-module collaboration, it effectively solves the detection problems of small aphid targets, translucency, cluster adhesion, etc., and performs well in scenes with different brightness and aphid density.

[0016] 4. High application value: It lays the foundation for a large-scale intelligent monitoring system for aphid feeding behavior, provides technical support for intelligent prediction and early warning of pests, and can shorten the research cycle of crop insect resistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a real-life rendering of the video data of peach aphid behavior.

[0018] Figure 2 It is a temporal weighted frame difference adaptive fusion framework.

[0019] Figure 3It is a visual graph of motion feature extraction under different time windows.

[0020] Figure 4 It is the feature map after TWFDAFF processing.

[0021] Figure 5 are the labeled aphid behaviors in the dataset.

[0022] Figure 6 It is the FGC-YOLO overall detection framework.

[0023] Figure 7 This is the GLSA implementation flow chart.

[0024] Figure 8 This is the CLWA architecture diagram.

[0025] Figure 9 It is the behavior detection result diagram.

[0026] Figure 10 It is the loss curve of training set and validation set.

[0027] Figure 11 This is the visualization result of the ablation experiment.

[0028] Figure 12 This is a diagram of the honey dew wrong kick behavior. DETAILED DESCRIPTION

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meanings as those commonly understood by those skilled in the art to which this application belongs. 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.

[0031] A method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features, the method comprising: S1: Collect video data of aphids in their natural feeding environment; The aphid behavior videos in this study were taken from the greenhouse of the Key Laboratory of New Pesticide Creation and Application, College of Plant Protection, Henan Agricultural University. The subjects were wingless peach aphids (Myzus persicae), which were raised on healthy tobacco leaves (Zhongyan No. 1 variety) for a long time, covering aphid individuals at different developmental stages. A Sony camera (model ILCE-7RM2) was used as the shooting device. In order to clearly capture the movements and behaviors of aphids, a macro lens (Laowa camera lens 25mm f / 2.8 ULTRAMACRO 2.5-5X) was required. The video image resolution was 1920×1080 pixels and the frame rate was 30FPS. A total of 32 sets of video data were collected in the experiment, and each video was about 30 minutes long. The actual shooting effect of the aphid behavior video is as follows: Figure 1 shown.

[0032] S2: Temporal weighted frame difference adaptive fusion framework (TWFDAFF) is used to extract spatiotemporal features of the original video data to obtain a motion feature map; The frame difference method (FDM) is a commonly used method for motion target detection. It identifies moving objects in a scene by calculating the difference between corresponding pixels in adjacent frames. Due to its simple computational simplicity, strong real-time performance, low sensitivity to scene changes such as lighting, and excellent stability, it is therefore used for extracting aphid behavioral features. However, the FDM relies on a selected inter-frame interval and is insensitive to small movements or gradual changes. Traditional two-frame and three-frame difference methods are prone to false targets, so improvements were made to this method during the experiment.

[0033] In order to simultaneously obtain the appearance and behavior characteristics of aphids, this paper proposes a temporal weighted frame difference adaptive fusion framework (TWFDAFF) to detect aphid behavior. Figure 2 shown.

[0034] S21: Processing continuous video frames using a sliding time window mechanism, calculating the time-series weighted frame difference to obtain a frame difference image; To accurately capture the behavioral characteristics of aphids expelling honeydew, the experiment used a sliding temporal window mechanism to process continuous video frames. The length of the temporal window significantly affects the extraction of motion features. In this study, the image resolution was kept fixed at 2K to explore the effects of different temporal window sizes. The results of motion feature extraction are shown in Figure 3.

[0035] Through the comparative analysis in the figure, it can be seen that the motion feature image generated by the time window size of 10 frames achieves a good balance between capturing low-frequency motion information and retaining detail features. The highlight part is high in clarity and there is no obvious tailing. Therefore, the window length is finally determined to be 10 frames.

[0036] Assume the input sequence is , the sliding window size is T, then the output frame difference It can be calculated by formula (1); In the formula is the temporal weight coefficient; in order to highlight the correlation between adjacent frames, an exponentially increasing weight is adopted. This design makes the weight distribution more consistent with the characteristics of the biological visual system that is sensitive to short-term temporal information, while also ensuring the stability of the feature scale.

[0037] S22: Perform weighted superposition on all frame differences within the time window to obtain an initial motion map; Perform weighted superposition of all frame differences within the time window to obtain the initial motion map , as shown in formula (2).

[0038] S23: Adaptive threshold processing based on statistical features to obtain a motion saliency map; In order to effectively suppress background noise, an adaptive threshold based on statistical characteristics is introduced , motion saliency map Calculated by formula (3).

[0039] In the formula , and Mapping The threshold design takes into account the overall level of background noise and reflects the dynamic characteristics of the scene.

[0040] S24: Perform Gaussian filtering and denoising on the motion saliency map, and adaptively fuse it with the original image to obtain a motion feature map.

[0041] In order to obtain a smooth feature representation, the saliency map Gaussian filtering is performed to obtain the denoised saliency map , and the original image Perform adaptive fusion and calculate it by formula (4); In the formula is the final output motion feature map, is the color representation of motion saliency (the yellow channel is used in the experiment), is the adaptive mixing coefficient based on motion intensity, ,This adaptive fusion strategy not only retains the detail information of the ,original image, but also moderately enhances the motion ,features.

[0042] Figure 4 The process of generating feature map visualization using TWFDAFF is shown. The algorithm first calculates the weighted difference between consecutive frames, and obtains Figure 4 (b) The motion map is shown. Then, through adaptive threshold processing, we can get Figure 4 (c) The motion saliency map is shown. Then, the motion saliency map is subjected to noise reduction using Gaussian filtering. The result is shown in Figure 4 Finally, the denoised motion features are compared with Figure 4 The original image in (a) is adaptively fused, and the fusion result is as follows Figure 4 As shown in (e), the yellow pixel group represents the area where the aphid's action behavior occurs. Compared with the traditional frame difference method, this method significantly reduces noise interference and provides more reliable feature input for subsequent behavior recognition.

[0043] S3: Build the FGC-YOLO small target fine behavior detection model: Behavioral labeling and dataset partitioning: Based on prior information on aphid behavior categories, we annotated the aphid locations and behavior types in the original video data after spatiotemporal feature extraction, constructing a fine-grained behavioral dataset encompassing three behavioral categories: movement, kicking, and honeydew discharge. In behavioral recognition research, accurate behavioral labeling and reasonable data set division play a decisive role in model performance. The original video is processed by the TWFDAFF feature extraction method, and key frames are extracted according to the rule of every 5 frames. The labelimg software is used to select and mark the location and behavior type of aphids. The recorded behavior types include three typical behaviors: movement, kicking, and honeydew discharge (such as Figure 5 As shown in the figure, pedipalping and honeydew excretion together constitute honeydew excretion behavior. These behavioral categories will provide a basis for subsequent research on feeding behavior. The behavioral recognition dataset in this experiment consists of 7,149 images, capturing typical behavioral patterns exhibited by aphids under natural conditions. To ensure good generalization performance of the model, the dataset was divided into training, validation, and test sets, with a ratio of 70%, 15%, and 15% respectively. The training set was used for model parameter learning, the validation set was used to adjust model hyperparameters during training to avoid overfitting, and the test set was used for final evaluation of model performance.

[0044] The dataset is input into the FGC-YOLO small target fine behavior detection model for training and cross-validation, and the trained FGC-YOLO small target fine behavior detection model is output.

[0045] In order to achieve real-time monitoring of aphids' honeydew discharge behavior, the network structure of YOLOv11 was adjusted to improve the detection ability of small targets and motion information features, and the FGC-YOLO model was proposed. It adopts FG-BiFPN, GLSA, and an efficient channel attention mechanism CLWA module that integrates the parallel convolution structure of C3k2 and the local window attention mechanism (LWA) to better capture the key features of aphids' honeydew discharge behavior; at the same time, a post-processing optimization framework based on double-buffered smooth interpolation and dynamic category judgment is proposed in the output post-processing, aiming to improve the continuity of target detection results and the ability to recognize complex behaviors. The overall detection framework is as follows: Figure 6 shown.

[0046] S31: Use fine-grained bidirectional feature pyramid network (FGBiFPN) as the neck network; FGBiFPN introduces finer feature maps and designs a feature fusion path, significantly enhancing its ability to perceive small objects. At the network node processing level, the CLWA module replaces the C3k2 module as the core node processing unit. Through a three-way loop stacking operation, it provides enhanced local feature extraction capabilities. Furthermore, FGBiFPN standardizes the number of channels at each feature layer to ensure the coherence and stability of information flow throughout the network. This simplifies the overall network architecture, reduces computational complexity, and improves feature fusion efficiency.

[0047] FGBiFPN inherits the fast normalization weighting mechanism of BiFPN, and the feature fusion process is expressed by formula (5); In the formula represents the output features, For the input features, For the corresponding weight, by initial learnable parameters conduct Activate to get; is a correction factor used to ensure the stability of the value.

[0048] S32: Introducing the global to local spatial aggregation module (GLSA); GLSA, whose core structure architecture adopts a dual-branch design, focuses on local spatial attention (LSA) and global spatial attention (GSA), aiming to aggregate local detail features and global structural information in parallel to optimize the performance of action recognition.

[0049] The implementation process of the GLSA module is as follows Figure 7 As shown in Figure 2, the input feature map X is split into two parts, X0 and X1, according to the channel dimension. The two parts enter the LSA branch and the GSA branch for processing respectively.

[0050] The core goal of the LSA branch is to extract local detail features. The input X0 first passes through a scale-1 convolutional layer to achieve channel dimensionality reduction and feature transformation, thereby obtaining intermediate features. This intermediate feature is then fed into the local feature enhancement module, which further extracts and enhances local spatial attention features, accurately mining local detail information within the image.

[0051] The GSA branch focuses on capturing global structural information. The input X1 also first passes through a scale-1 convolutional layer to perform channel dimensionality reduction and feature transformation. This feature is then fed into the global context modeling module, which captures global context information and long-range dependencies, grasping the overall characteristics of the image from a macro perspective.

[0052] Finally, the outputs of the LSA and GSA branches are concatenated along the channel dimension and passed through a scale-1 convolutional layer for feature fusion and channel number adjustment to generate the final output features. GLSA, by parallelizing and fusing local and global spatial information, balances local refinement and global consistency in the image, improving the model's ability to recognize large and small objects and key details in tasks such as action recognition.

[0053] S33: Design the local window attention mechanism (CLWA) module integrated with C3k2; The CLWA module combines the parallel convolutional architecture of C3k2 with the efficient LWA channel attention mechanism. The original Bottleneck structure is replaced by the LWA Bottleneck structure. Within each parallel branch of C3k2, the LWA Bottleneck module adaptively adjusts feature weights through the LWA attention mechanism, making the model more focused on key information. The CLWA architecture details are as follows: Figure 8 CLWA combines the efficient computing of C3k2 and the refined feature extraction of LWA, improving the target detection performance of the target detection algorithm in complex scenarios.

[0054] S4: Input the motion feature map into the trained FGC-YOLO small target fine behavior detection model for detection, and output the result identification of the aphid behavior category selected by the frame.

[0055] S5: Post-process the output results of the FGC-YOLO small target fine behavior detection model, and realize the recognition of honeydew discharge behavior through the dynamic double buffer interpolation mechanism and the composite behavior dynamic judgment algorithm.

[0056] The honeydew excretion behavior of aphids consists of continuous kicking movements and honeydew balls. To improve the accuracy of composite behavior detection, a post-processing optimization framework based on double-buffered smooth interpolation and dynamic category determination was proposed.

[0057] 1. Dynamic double buffer interpolation mechanism Based on the assumption of temporal continuity, a double buffer structure with odd-even frame division is constructed. The two groups of buffers store odd frames and even frames respectively, and fault tolerance control is achieved by introducing a missing counter Count; when the number of consecutive undetected times Count < 2, the linear interpolation method is used to generate the current frame detection result , which can be calculated by formula (6).

[0058] Where, is the interpolation result of the current frame, and Represent the detection results of the previous and next frames respectively, is the interpolation weight coefficient. This mechanism effectively solves the problem of detection interruption caused by motion mutation.

[0059] 2. Composite Behavior Dynamic Judgment Algorithm To determine the honeydew-excretion behavior of aphids, this study used a dynamic compound behavior determination algorithm based on semantic relationships. The algorithm is implemented in the following four specific steps.

[0060] 1) Behavioral Correlation Determination. If aphids are detected kicking within a very short time window (set to 2 seconds in this study), based on relevant research and experience in aphid behavior, it can be preliminarily determined that honeydew discharge is possible. This is because there is a certain temporal correlation between kicking and honeydew discharge in aphid behavioral patterns.

[0061] 2) Spatial Proximity Determination. After confirming that an aphid is kicking, the surrounding space is further examined to determine whether a "honeydew ball" is located within the aphid's kicking behavior. This step aims to confirm the spatial proximity between the "honeydew ball" and the aphid's kicking behavior, providing a spatial dimension for subsequent behavioral determination.

[0062] 3) Identification of composite behaviors. When both of the above conditions are met, the individual aphid and its adjacent honeydew ball are combined into a single detection result and classified as "honeydew discharge," characterizing this composite aphid movement. This classification clearly reflects the intrinsic connection between aphid kicking and honeydew secretion, enabling accurate identification of honeydew discharge.

[0063] 4) Result storage and visualization: After determining the aphid's honeydew excretion behavior, the video and images containing this behavioral information are saved for subsequent analysis and research.

[0064] 3. Test results and analysis: 1) Identification effect of aphid honeydew discharge behavior: The FGC-YOLO model was used in the experiment to detect three behaviors: movement, kicking, and honeydew discharge. On a 4090 GPU device, it can achieve an average processing speed of 65 frames per second, with 18.8M parameters and 84.3G floating-point calculations. The average accuracy of the detection frame reached 81.5% after 209 training cycles, and the accuracy of honeydew discharge recognition was 99.9%. The behavior detection results are as follows: Figure 9 shown.

[0065] The changes in loss on the training set and validation set during model training are as follows: Figure 10 As shown in the figure. The losses on the training and validation sets include localization loss (Box Loss), classification loss (Cls Loss), and dynamic feature loss (DFLLoss). Figure (a) shows that the training loss values ​​all show a downward trend with the increase in training rounds, indicating that the model is continuously learning and optimizing parameters. Specifically, in the early stages of training (rounds 0-50), the values ​​of the three loss functions all decrease at a relatively fast rate, indicating that the model can quickly adapt to the training dataset. In the later stages of training after 50 rounds, the rate of loss decline gradually slows, eventually stabilizing at approximately 1.0 for localization loss, approximately 0.6 for classification loss, and approximately 0.5 for dynamic feature loss, indicating that model training is gradually converging. Notably, the dynamic feature loss remains higher than the other two losses throughout the entire training cycle. It is speculated that its impact on model performance is more sensitive, possibly due to the high response of this loss function to changes in model parameters.

[0066] During the evaluation of the validation set, the loss function showed a similar trend to that of the training set. The localization loss and dynamic feature loss experienced a rapid decline at the beginning of training before gradually stabilizing. The classification loss converged rapidly within the first 20 rounds of training, and the validation classification loss even dropped to a negative value. This result strongly demonstrates that the model achieved high classification accuracy on the validation set. However, the subsequent classification loss showed a slight upward trend, suggesting that the model may have begun to show a certain degree of overfitting. Overall, the loss trend on the validation set was largely consistent with that on the training set, confirming the model's good performance in generalization and convergence stability.

[0067] 2) Ablation test results To validate the effectiveness of each module, this study conducted ablation tests using YOLOv11n and YOLOv11m as baseline models. The results are shown in Table 1. Experimental observations show that the basic YOLOv11n model has low detection accuracy. In comparison, YOLOv11m has moderate complexity and was selected as the baseline model. During the experiments, the FGBiFPN, GLSA, and CLWA modules were gradually introduced, and the model's average detection precision increased to 80.2%, 80.9%, and 81.5%, respectively. However, this improvement in accuracy also comes with an increase in model complexity, including increased parameters and floating-point computation. Notably, the model achieved significant performance improvements in the key behavior honeydew detection task, with the detection index rising from 71.1 to 77.3. These experimental results demonstrate that the FGC-YOLO model significantly enhances its ability to detect key behaviors of small objects, while sacrificing computational effort to a certain extent, achieving a balanced performance.

[0068] This study also conducted visualization analysis in the ablation experiment under three different scenarios (low brightness and high aphid density, low brightness and low aphid density, and normal brightness and high aphid density). Figure 11 As shown in the figure, the detection performance of each model combination in different scenarios is intuitively presented through heat maps, and the differences in detection results are clearly shown in a visual way.

[0069] In the first set of experimental scenarios, heatmap analysis of the "Kicking" target box showed that after optimization with FGBiFPN and GLSA, the activation area was significantly expanded and more concentrated; however, after the introduction of CLWA, the activation area decreased. This visualization result is consistent with the mAP data for the "Kicking" category in Table 1, intuitively confirming the differential impact of improvements in different modules on detection performance.

[0070] In the second set of scenarios, the activation areas of the three Moving category targets diffused to the non-target area after applying FGBiFPN, retracted and concentrated after applying GLSA, and became smaller after applying CLWA, which is consistent with the Moving category mAP scores shown in Table 1.

[0071] In the second set of experimental scenarios, observing three moving objects, we found that after introducing FGBiFPN, the activation area in the heatmap spread to non-target areas outside the target box; however, after adding the GLSA module, the activation area effectively converged and became more centralized; and after further introducing CLWA, the activation area was significantly reduced. This series of visual changes is consistent with the mAP data for the moving category in Table 1, fully verifying the optimization effect of each module on object detection performance.

[0072] Given the tiny size of honeydew targets, heatmaps struggle to accurately capture their characteristics. Confidence analysis was conducted on the honeydew targets in a third set of scenarios with varying aphid densities. The results showed that the introduction of FGBiFPN resulted in the most significant fluctuations in detection performance across density conditions. The addition of the GLSA module resulted in a decline in overall detection performance. However, the integration of CLWA significantly improved and optimized detection performance, achieving significant improvements.

[0073] The ablation test results clearly demonstrate the contribution of each module, highlighting the advantages of FGC-YOLO in improving the ability to detect specific key behaviors of small objects.

[0074] 3) Comparison of different target detection algorithms In this experiment, a comparative study of different object detection algorithms was conducted on aphid honeydew excretion behavior. On the same training and test sets, mainstream object detection algorithms, including YOLOv5m, YOLOv7, YOLOv8m, YOLOv10m, YOLOv11m, YOLOv12m, and RT-DETR-r18, were used for recognition. The experimental results are shown in Table 2. The experimental results show that the proposed FGC-YOLO model improves precision, recall, and average precision by 3.7%, 4.3%, and 6.2%, respectively, compared to the original YOLOv11m model, while reducing the number of parameters by 1.2M. Compared to the YOLOv8m model, while maintaining the same precision, the recall and average precision improve by 24.2% and 5.1%, respectively, while reducing the number of parameters by 7M. Compared to the other algorithms, the FGC-YOLO model significantly outperforms the other compared algorithms in both precision and average precision. At the same time, the model maintains comparable performance to similar algorithms in terms of both the number of model parameters and the amount of computation required. Therefore, the FGC-YOLO model can provide technical support for real-time detection of aphid honeydew discharge and lay a theoretical foundation for subsequent research on the feeding behavior of piercing-sucking insects.

[0075] 4) Analysis of test results As shown in the above figures, the proposed FGC-YOLO model achieves a high precision of 99.9% when identifying aphid honeydew excretion, but its recall is only 54.6%. This demonstrates that the FGC-YOLO model accurately identifies honeydew excretion in videos once detected, with virtually no false positives. However, the model suffers from significant underdetection. The low recall is primarily due to three factors: 1. The sample size of honeydew excretion data is significantly different from that of movement data, resulting in a significant class imbalance in the dataset. This can be addressed in future research by acquiring larger datasets. 2. In 1K resolution images, small objects have a pixel size of only 45×45, occupying a very small proportion of the entire image. Due to their weak feature information and low discernibility, they are easily missed by the algorithm during object detection, resulting in a significantly lower recall than expected. 3. The entire process from honeydew excretion to being kicked off the aphid's body is extremely brief. Furthermore, the original video dataset was derived from a real aphid feeding scene, where multiple aphids are present simultaneously. This leads to a series of complex situations, such as mutual obstruction between aphids, misjudgment of honeydew kicks, and honeydew splashing due to external factors. For example, in an experimental example, two aphids excrete honeydew simultaneously, as shown in Figure 12(a). The larger aphid, due to its proximity, kicked the honeydew from the smaller aphid onto the third, stationary aphid. See Figures 12(b), (c), and (d) for details. Such complex situations can easily lead to missed detections or misjudgments during automatic identification and detection. These issues can be addressed by adding constraints to the honeydew location (limited to the anus at the tail end of the aphid's abdomen) in the algorithm's judgment.

[0076] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several changes and improvements can be made without departing from the overall concept of the present invention, and these should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features, characterized by: The method comprises: S1: Collect video data of aphids in their natural feeding environment; S2: Use the time-series weighted frame difference adaptive fusion framework to extract spatiotemporal features from the original video data to obtain a motion feature map, including: S21: Processing continuous video frames using a sliding time window mechanism, calculating the time-series weighted frame difference to obtain a frame difference image; S22: Perform weighted superposition on all frame differences within the time window to obtain an initial motion map; S23: Adaptive threshold processing based on statistical features to obtain a motion saliency map; S24: performing Gaussian filtering to remove noise from the motion saliency map, and adaptively fusing it with the original image to obtain a motion feature map; S3: Combined with prior information on aphid behavior categories, the original video data after spatiotemporal feature extraction is used to mark the aphid's location and behavior type, and a fine-grained behavior dataset containing three behavior categories: movement, kicking, and honeydew discharge; Input the dataset into the FGC-YOLO small target fine behavior detection model for training and cross-validation, and output the trained FGC-YOLO small target fine behavior detection model; The FGC-YOLO small target fine behavior detection model uses the YOLOv11 model as the baseline network, including the backbone network, neck network and head network. The FGC-YOLO small target fine behavior detection model uses the fine-grained bidirectional feature pyramid network FGBiFPN as the neck network. The neck network introduces the global to local spatial aggregation module GLSA. The local window attention mechanism CLWA module integrated with C3k2 is used to replace C3k2 as the node processing unit. S4: Input the motion feature map into the trained FGC-YOLO small target fine behavior detection model for detection, and output the result identification of the aphid behavior category selected by the frame; S5: Post-process the output results of the FGC-YOLO small target fine behavior detection model, and realize the recognition of honeydew discharge behavior through the dynamic double buffer interpolation mechanism and the composite behavior dynamic judgment algorithm.

2. The method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features according to claim 1, characterized in that: The S2 specifically includes: Temporal motion feature extraction: Using a sliding time window mechanism to process continuous video frames, let the input sequence be , the sliding window size is T, then the output frame difference It can be calculated by formula (1); In the formula is the time series weight coefficient; Motion saliency calculation: perform weighted superposition of all frame differences within the time window to obtain the initial motion map , as shown in formula (2); （2） In order to effectively suppress background noise, an adaptive threshold based on statistical characteristics is introduced , motion saliency map Calculated by formula (3); In the formula , and Mapping The mean and standard deviation of Feature optimization fusion: To obtain a smooth feature representation, the saliency map Gaussian filtering is performed to obtain the denoised saliency map , is adaptively fused with the original image I and calculated by formula (4); In the formula is the final output motion feature map, is a color representation of motion saliency, is the adaptive mixing coefficient based on motion intensity, .

3. The method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features according to claim 1, characterized in that: The fine-grained bidirectional feature pyramid network (FGBiFPN) in S3 improves the neck network structure in YOLOv11 by replacing C3k2 with the CLWA module as the node processing unit. It enhances local feature extraction capabilities through three cyclic stacking operations and standardizes the number of channels in each feature layer. The feature fusion process of FGBiFPN is expressed by formula (5); In the formula represents the output features, For the input features, For the corresponding weight, by initial learnable parameters conduct Activate to get; is a correction factor used to ensure the stability of the value.

4. The method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features according to claim 1, characterized in that: The global to local spatial aggregation module GLSA in S3 adopts a dual-branch design, including a local spatial attention branch and a global spatial attention branch, which are used to extract local detail features and global structural information respectively, and then the outputs of the two branches are spliced ​​and fused along the channel dimension.

5. The method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features according to claim 1, characterized in that: The CLWA module in S3 integrates LWA into the parallel convolution path of C3k2 and replaces the original Bottleneck structure with the LWABottleneck structure. In each parallel branch of C3k2, the LWABottleneck module adaptively adjusts the feature weights through the LWA attention mechanism, making the model more focused on key information.

6. The method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features according to claim 1, characterized in that: The dynamic double buffer interpolation mechanism in S5 adopts a double buffer structure of odd-even frame division. When the number of consecutive undetected times is less than 2, a linear interpolation method is used to generate the current frame detection result.

7. The method for identifying aphid honeydew discharge behavior based on adaptive fusion of spatiotemporal features according to claim 1, characterized in that: The dynamic determination algorithm of the composite behavior in S5 includes: Behavior association determination: When kicking behavior is detected within the set time window, it is preliminarily determined that there is a possibility of honeydew discharge; Spatial proximity determination: After confirming that an aphid is kicking, the space around the aphid is further inspected to determine whether a "honeydew ball" target is located within the aphid's kicking behavior frame. Identification of composite behaviors: When both the behavioral association and spatial proximity conditions are met, the aphid individual and the honeydew ball are combined into the honeydew excretion category to characterize this composite movement behavior of the aphid; Result storage and visualization: After the determination of aphid honeydew discharge behavior is completed, the video and images containing the behavior information are saved.

Images