EIPE-CSWFPN enhanced YOLO identification method and system for low-altitude inspection of cold region agriculture

Abstract

The application provides an EIPE-CSWFPN enhanced YOLO identification method and system for low-altitude agricultural inspection in cold regions, and belongs to the field of low-altitude unmanned aerial vehicle plant protection inspection. In order to solve the problems that the existing method is prone to misjudgment and missed detection due to the small size of the pest target, the weak boundary and texture features, and the interference of complex leaf texture and light disturbance, and the traditional detection network is insufficient in utilizing small target information in the feature extraction and multi-scale fusion link, the EIPE feature extraction module is introduced in the existing YOLO detection framework, and the cross-stage weighted feature pyramid CSWFPN is used to replace the traditional Neck structure, so that the key technical links of agricultural pest identification in the low-altitude inspection scene are systematically optimized without changing the unmanned aerial vehicle operation process and the detection output form. The application realizes more stable and reliable automatic identification of the pest target, and takes into account the detection accuracy, robustness and engineering feasibility, so that the risk of misjudgment and missed detection to plant protection decision can be effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of low-altitude unmanned aerial vehicle (UAV) plant protection inspection technology, and more specifically, to an EIPE-CSWFPN enhanced YOLO recognition method and system for low-altitude inspection of cold-region agriculture. Background Technology

[0002] In low-altitude drone-based plant protection and inspection applications, existing pest identification primarily relies on general target detection models. These models depend on aerial images to directly output pest categories and locations, which users then use to make control decisions. However, in cold-region agriculture, low-altitude flights are significantly affected by wind disturbances, rapid changes in light intensity, and crop canopy obstruction. Aerial images often exhibit insufficient contrast, dense textures, and localized blurring. Furthermore, pest targets, viewed from above, typically occupy only a few pixels and are highly similar to background details such as leaf veins, spots, and branch shadows, making detection results highly sensitive to environmental conditions. If the model misdetects or misses targets in real-world operations, it can easily lead to unnecessary pesticide application, missing the optimal control window, or over-treatment, thereby amplifying ecological and ethical risks and raising the barrier to entry for low-altitude technology in farmers' operations.

[0003] From a network representation perspective, traditional YOLO-based object detection algorithms typically employ general convolutional or multi-layer feature stacking modules focused on high-level semantic aggregation during the backbone feature extraction stage. Their design prioritizes enhancing semantic discrimination capabilities, while lacking targeted and explicit enhancement mechanisms for low-level edge structures, contour information, and local geometric details—crucial for fine target localization. "Small targets" specifically refer to pests in aerial images of farmland or woodland, whose pixel area occupies less than 0.1% of the entire image. These targets are typically characterized by extremely small scale, significant morphological variations, subtle texture features, and easy blending with the background.

[0004] When existing YOLO models are applied to the aforementioned small-scale pest detection tasks, especially in complex natural scenes characterized by small target size, weak texture details, and strong background interference, the network is prone to over-smoothing of target boundary information during multiple downsampling and feature fusion processes. This leads to problems such as incomplete target contour representation, unstable candidate box regression, and jittering of detection box positions. In real-world application environments with dense leaf surface textures, drastic lighting changes, or significant shadow occlusion, the model tends to learn local correlation features of background textures. This can result in false detections, misidentifying background textures as pest targets, or false detections, misclassifying real pest targets as background areas. Consequently, it is difficult to achieve stable and reliable identification of small-scale pest targets under complex background conditions.

[0005] At the multi-scale fusion level, existing methods generally adopt traditional feature pyramids or PAN / FPN structures, which aggregate features by upsampling / downsampling at each level and then directly adding or concatenating them at the fusion node. In low-altitude inspection scenarios with large scale spans, complex background textures, and varying effective scales due to different heights and viewpoints, this approach is prone to two prominent problems: First, different scales are treated "equally." When a branch at a certain scale has stronger noise or more prominent background texture, the fusion result is dominated by noise, weakening the contribution of the truly effective scale. Second, relying solely on cross-layer transmission through step-by-step sampling is prone to information attenuation, making it difficult to stably compensate high-level semantics to the fine-resolution level, thus affecting the fine-grained discrimination and regression accuracy of small targets.

[0006] To address the issue of insufficient detection accuracy for small-scale pest targets in complex backgrounds, existing improvement approaches typically revolve around increasing model size or computational complexity. Examples include increasing input resolution, deepening or widening the network structure, adding general attention modules to the backbone or neck network, or upgrading from a lightweight model to a larger-scale detector framework to improve detection accuracy. However, these methods often come at the cost of significantly increasing the number of model parameters, computational load, and inference latency, making it difficult to meet the low-power and real-time application requirements of UAVs. For instance, YOLOv10s, as a lightweight detection model that balances accuracy and efficiency, typically has around 7–8M parameters, making it well-suited for embedded or airborne computing platforms. However, in practical applications, if the detection capability for small targets is enhanced by deepening the network layers, widening the channel width, or introducing additional attention modules, the model size often rapidly evolves to a medium-to-large structure, with the number of parameters increasing to over 20M, resulting in a significant increase in computational power consumption and GPU memory usage. Furthermore, increasing the input resolution from the commonly used 640×640 to 960×960 or 1024×1024, while not directly increasing the number of model parameters, significantly amplifies the feature map computation scale, further burdening the edge-side inference. In actual UAV operation scenarios, the onboard computing unit, in addition to performing target detection inference, also needs to undertake multiple real-time tasks such as flight control, attitude stabilization, path planning, and communication. When the detection model is significantly "burdened" on top of YOLOv10s, it can easily lead to increased overall system power consumption, decreased inference frame rate, or intensified resource contention, resulting in problems such as shortened flight time, insufficient real-time performance, and increased edge-side deployment complexity. Ultimately, this creates an application gap where "the model is usable in the experimental environment, but it is difficult to run stably for a long time in complex field conditions."

[0007] Meanwhile, some existing methods attempt to compensate for the lack of clarity in small targets by image enhancement or high-frequency detail emphasis. However, without a mechanism to suppress noise amplification and a training stability design, the enhancement operation often amplifies the high-frequency texture of the background under conditions such as wind field jitter, compressed noise, rain and fog, making the network learning process more unstable and the generalization more fragile. Ultimately, this manifests as significant fluctuations in accuracy between different plots and different lighting periods, strong threshold dependence, and the need for frequent manual parameter tuning.

[0008] Furthermore, pest identification is often accompanied by long-tailed categories and sample imbalance. General training processes are insufficient for learning small sample categories, easily leading to unstable recall and higher misclassification rates for rare categories. When the identification results are directly used for plant protection decisions, this instability further amplifies the risks in actual operations. In summary, existing technologies for low-altitude pest detection still generally suffer from the following shortcomings: insufficient representation of small target boundaries in complex backgrounds, lack of scene adaptation in multi-scale fusion, difficulty in balancing edge deployment and robustness, and insufficient stability in long-tailed category identification. These shortcomings make it difficult to form a unified solution that is easy for farmers to use while balancing accuracy and risk control. Summary of the Invention

[0009] The technical problem to be solved by this invention is:

[0010] To address the issues of small-scale pest targets, weak boundary and texture features, and susceptibility to interference from complex leaf textures and lighting disturbances in low-altitude drone agricultural inspection scenarios, leading to false positives and false negatives, as well as the insufficient utilization of small target information by traditional detection networks in feature extraction and multi-scale fusion stages.

[0011] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0012] This invention provides an EIPE-CSWFPN-enhanced YOLO identification method for low-altitude inspection in cold-region agriculture, comprising the following steps:

[0013] S100: Using aerial images collected by drones as input, the C2F unit in YOLOv10 is replaced with the EIPE edge information perception enhancement module in the backbone network feature extraction stage.

[0014] S200. After completing the backbone feature extraction in step S100, the Neck structure in YOLOv10 is replaced with the Cross-Stage Weighted Feature Pyramid (CSWFPN). Multi-scale features from each stage of the backbone network are incorporated into the same fusion framework. Cross-stage information transmission is completed using a bidirectional fusion link that is both bottom-up and top-down. A weighted fusion mechanism is introduced during the bidirectional fusion process to adjust the contribution of features from different stages and scales.

[0015] S300. After completing the EIPE enhanced feature extraction in step S100 and the CSWFPN cross-stage weighted fusion in step S200, the multi-scale feature input of the fused product is used to perform target classification and position regression output, thereby achieving reliable localization and identification of small-scale pest targets.

[0016] Further, in step S100, the EIPE edge information perception enhancement module uses the input feature map Starting from the first branch, multiple parallel multi-scale branches are constructed to cover different receptive fields. Each branch uses adaptive average pooling to reshape features into a fixed spatial grid, thus obtaining coarse-to-fine contextual representations at the same level. The branch features after adaptive average pooling are fed into an edge information extractor to generate edge enhancement terms, and residuals are superimposed with the branch features to obtain the enhanced branch output. The enhancement results of all scale branches are concatenated in the channel dimension, while retaining one channel from the original input. The direct bypass path to the splicing end; the spliced ​​features are accessed via... Convolution is used for fusion and channel compression to obtain output features. .

[0017] Further, step S100 specifically includes,

[0018] Input features for any scale branch First through convolution Complete channel alignment and redundancy compression, then... convolution The local structural response is extracted and then upsampled to a spatial scale consistent with the branch features to obtain the intermediate features. :

[0019]

[0020] To explicitly separate high-frequency components at the edges, intermediate features are introduced. Scale-consistent local average pooling smoothing term :

[0021]

[0022] in, The scale index represents the multi-scale branch; Indicates the first The intermediate feature maps obtained after convolution transformation and upsampling of each scale branch, and,

[0023]

[0024] in, , and These represent intermediate feature maps. The number of channels, height, and width; Indicates the core size is Local average pooling operation with a step size of 1 and a fill width of 1;

[0025] High-frequency residuals are generated through differential methods. :

[0026]

[0027] high frequency residual Send in Convolution followed by sigmoid mapping Obtain edge response terms :

[0028]

[0029] At the branch level of the EIPE edge information perception enhancement module, the edge response item As an edge enhancement term, it is injected back into the scale branch features in a residual manner to form an enhanced output. :

[0030]

[0031] At the multi-scale fusion level, the enhanced output of each scale branch After channel concatenation and fusion with the original input features from the bypass, a joint representation, i.e., the output feature, is obtained, which combines fine-grained boundary sensitivity with global context consistency. .

[0032] Furthermore, in step S200, in the cross-stage weighted feature pyramid CSWFPN, the features of four stages, P2-4, P3-8, P4-16, and P5-32, are used as inputs. Each path first completes channel shaping through convolution, and then aligns features of different resolutions to the target scale through upsampling and downsampling. The bottom-up path gradually gathers low-level details and structural information into deeper feature representations, while the top-down path feeds back high-level semantic information to higher resolution levels, realizing semantic compensation for details and details constraining semantics. At the same time, cross-stage bypass inputs are introduced at key nodes, enabling the fusion node to simultaneously receive multi-source feature streams transmitted from different stages and directions.

[0033] Furthermore, in step S200, the fusion process of the cross-stage weighted feature pyramid (CSWFPN) is completed by two types of structural units. The first type is the Concat fusion unit: the aligned features are concatenated at the same scale, and feature reshaping and nonlinear refinement are performed by the C2f unit to fully couple the semantic information of the upsampled backflow with the detailed information at the same scale, forming a structured multi-scale aggregated representation. The second type is the Fusion1, Fusion2, and Fusion3 weighted fusion units: Conv1×1 is introduced before these nodes to align the channels of each input, so that features from different sources enter the same fusion space. Then, learnable weights are introduced in the fusion nodes to perform normalized weighted fusion of two or three inputs. The fusion result is then connected to the C2f unit and passed to a deeper Fusion layer through downsampling, so that the weighting calibration is continuously effective at multiple scales. The final output forms a three-layer pyramid feature, which is directly connected to the detection head.

[0034] For the fusion of two inputs, let the feature maps of the two inputs be respectively... and The corresponding learnable weights are respectively and Two-way fused output feature map Defined as:

[0035]

[0036] For the fusion of three inputs, let the feature maps of the three inputs be respectively... , and The corresponding learnable weights are respectively , and Three-way fusion output feature map Defined as:

[0037]

[0038] Among them, the upper right corner mark Indicates feature fusion from two paths, indicated by the upper right corner label. This indicates the fusion of three features; and This represents the feature maps from different sources of the two input fusion nodes. , and This represents the feature maps from different sources of the input three-way fusion nodes; , and Indicates the source or scale index of each input feature; , , , and These are learnable weights that correspond one-to-one with the corresponding input features; It is a preset positive constant.

[0039] An EIPE-CSWFPN enhanced YOLO recognition system for low-altitude inspection of cold-region agriculture is provided. This system has program modules corresponding to the above steps and executes the steps in the above-described EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection of cold-region agriculture during operation.

[0040] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of an EIPE-CSWFPN enhanced YOLO identification method for low-altitude inspection in cold-region agriculture.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention systematically optimizes key technical aspects of agricultural pest identification in low-altitude inspection scenarios by introducing an EIPE feature extraction module into the existing YOLO detection framework and replacing the traditional neck structure with a cross-stage weighted feature pyramid (CSWFPN). This optimization is achieved without altering the UAV operation process or detection output format, and offers the following advantages:

[0043] First, this invention introduces an EIPE module into the backbone network to replace the traditional feature extraction unit, enabling the network to generate a more stable response to the local structure, edge contours, and fine-grained texture changes of pest targets during the feature extraction stage. Compared to feature extraction methods that rely solely on conventional convolution stacking, this method effectively alleviates the problem of small-scale target detail information being easily weakened during downsampling, improves the distinguishability of pest targets in shallow and mid-level features, and reduces the interference of complex leaf textures and lighting changes on the recognition results from the source, providing a more reliable basic feature representation for subsequent feature fusion and detection.

[0044] Secondly, the cross-stage weighted feature pyramid (CSWFPN) employed in this invention incorporates multi-scale features from different stages of the backbone network into a unified modeling system through a bidirectional fusion link, achieving effective integration of detailed and semantic information. By introducing a weighting mechanism during the fusion process, the contribution of features from different stages and scales is dynamically adjusted, enabling the network to adaptively highlight feature levels more valuable for pest identification in low-altitude inspection scenarios with complex background textures or strong noise. This avoids the noise dominance and information attenuation problems caused by the simple superposition of multi-scale features in traditional Neck structures, thereby significantly improving the stability and consistency of multi-scale feature fusion.

[0045] In terms of overall structural design, this invention enhances the two key stages of feature extraction and fusion in a targeted manner, rather than simply relying on deepening the network or increasing the input resolution to improve accuracy. This improves the accuracy and robustness of small-scale pest identification while effectively controlling the model parameter scale and computational overhead, meeting the engineering requirements of low-altitude UAVs for real-time performance, power consumption, and stable operation, and avoiding the problem of "improved accuracy but difficult deployment".

[0046] In summary, this invention achieves more stable and reliable automatic identification of pest targets in low-altitude agricultural inspection scenarios with complex natural backgrounds and significant multi-scale changes. It balances detection accuracy, robustness, and engineering feasibility, effectively reducing the risks of false detections and missed detections to plant protection decisions, and has positive significance for enhancing the practical value of low-altitude UAV intelligent plant protection systems. Attached Figure Description

[0047] Figure 1 This is a network structure diagram of an EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture, as described in an embodiment of the present invention.

[0048] Figure 2 This is a network architecture diagram of the edge information perception enhancement module in an embodiment of the present invention;

[0049] Figure 3 This is a network architecture diagram of the cross-stage weighted feature pyramid network module in an embodiment of the present invention;

[0050] Figure 4 This is a diagram showing the recognition effect of the YOLOv10s method in an embodiment of the present invention.

[0051] Figure 5 This is a diagram illustrating the recognition effect achieved using this method in an embodiment of the present invention. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] With the routine application of low-altitude drones in plant protection inspections in cold-region agriculture, the scale of aerial monitoring data is growing rapidly. Farmers need to make pest identification and treatment decisions within a limited time. However, the overhead imaging of low-altitude inspections makes pest targets appear as small-scale features in the images. In addition, field operations are generally affected by factors such as image shake caused by wind disturbances, changes in lighting and shadow interference, and high-frequency crop leaf textures with dense background details. The visual differences between pest targets and easily confused areas such as leaf vein textures, spots, and branch shadows are significantly compressed, making existing general detection models prone to false positives and false negatives in real-world operations. If the identification results directly drive pesticide application and control actions, misjudgments will lead to consequences such as over-application or delayed control, further amplifying environmental and ethical risks and raising the barrier for farmers to use low-altitude technology.

[0054] Existing YOLO models typically focus on semantic aggregation during feature extraction, lacking explicit and controllable enhancement mechanisms for the crucial boundary and contour details in small target localization. When faced with the challenge of detecting pests characterized by small target scale, fine texture, and strong background interference, edge responses in features are easily smoothed, leading to unstable bounding box regression and insufficient discriminative ability against complex backgrounds. The challenge lies in how to enhance the backbone feature extraction stage with stronger edge perception and multi-scale representation capabilities without significantly increasing end-side overhead, while maintaining semantic integrity during enhancement and avoiding noise amplification and training instability caused by differential enhancement, thereby providing a more discriminative feature foundation for subsequent detection heads.

[0055] Furthermore, pests in low-altitude inspections exhibit a wide scale range that dynamically changes with flight altitude, shooting angle, and crop growth status. Traditional feature pyramids or PAN / FPN structures often treat features of different scales "equally" during fusion, using direct addition or splicing to complete the fusion, which can easily lead to noise at one scale dominating the fusion result. Simultaneously, relying solely on stepwise upsampling / downsampling across layers results in information attenuation, making it difficult to stably compensate high-level semantics to the fine-resolution level, thus causing false positives and false negatives in natural scenes. Therefore, it is crucial to construct a cross-stage fusion framework that can connect the outputs of different stages of the backbone, achieving cross-scale information transmission in a bottom-up and top-down bidirectional link, and introducing a learnable normalized weight mechanism at the fusion nodes to dynamically recalibrate the contribution of multi-scale features, enabling the model to automatically select more effective information sources under different background textures and scale conditions.

[0056] Meanwhile, the deployment of agricultural drones on the edge is constrained by power consumption, computing power, and real-time performance. Model structure modifications must strike a feasible balance between improving accuracy and controlling complexity. This requires enhancing detection accuracy and robustness in complex field scenarios while keeping the overall parameter scale and inference latency within the limits of airborne capabilities. Furthermore, it must support exporting weights in an edge-deployable format for stable operation in actual inspection tasks.

[0057] This invention proposes an attention-enhanced YOLO method for agricultural pest identification in low-altitude inspection scenarios. This method improves the feature representation ability of small-scale pest targets, the effectiveness of multi-scale feature fusion, and the detection stability in edge-side application scenarios. Targeting typical application scenarios such as low-altitude UAV plant protection inspections in cold-region agriculture, this method uses an integrated modeling approach—"image input—feature extraction enhancement—cross-stage weighted feature fusion—detection output"—as its main thread. Without changing the UAV aerial data acquisition method or existing detection output format, it achieves robust pest target identification in complex natural scenes through targeted replacement of the core feature extraction module and the Neck fusion structure.

[0058] Specific Implementation Plan 1: Combining Figures 1 to 3 As shown, this invention provides an EIPE-CSWFPN enhanced YOLO identification method for low-altitude inspection in cold-region agriculture, comprising the following steps:

[0059] S100: Using aerial images captured by drones as input, the traditional feature extraction module is replaced by the EIPE module in the backbone network feature extraction stage. The EIPE module is embedded in multiple key stages of the backbone network while maintaining consistent input and output feature sizes and interfaces. It is used to enhance the response to changes in local structure, edge contours, and fine-grained texture, and alleviate the problem of weakening small target detail information during downsampling and channel stacking in traditional convolutional feature extraction. By continuously strengthening the distinguishable cues of pest targets in the shallow to mid-layer feature space, the subsequent fusion and detection stages obtain more stable and discriminative basic feature representations, thereby reducing the probability of misclassifying texture as a target or target as background in complex backgrounds.

[0060] The Edge Information Perception and Enhancement Module (EIPE) addresses the challenges of low-altitude drone pest detection, where targets are small-scale, have fragmented boundaries, and are easily obscured by leaf texture and lighting variations. It uses edge structure information as a core element to improve positioning reliability, directly replacing the traditional C2F unit in YOLOv10's network structure. This enables edge sensitivity and multi-scale representation capabilities from the backbone feature extraction stage. The replacement concentrates modifications on the feature enhancement stage within the backbone, without altering the input / output interface of the detection head, ensuring controllable engineering implementation. Simultaneously, it enhances pest outlines and local mutation information before the network enters the neck fusion stage, reducing edge attenuation and missed detection risks caused by downsampling and background interference from the source.

[0061] EIPE uses input feature maps Starting from this point, multiple parallel, multi-scale branches are constructed to cover different receptive fields. Each branch reshapes the features into a fixed spatial grid (e.g., using adaptive average pooling) through adaptive average pooling. , , , Thus, coarse-to-fine contextual representations are obtained simultaneously at the same level; the branch features after adaptive average pooling are fed into the edge information extractor to generate edge enhancement terms, and residuals are superimposed with the branch features to obtain the enhanced branch output; the enhancement results of all scale branches are concatenated in the channel dimension, while retaining one path from the original input. The direct bypass path to the concatenation end not only ensures that the original semantic information is not lost due to pooling reshaping, but also effectively prevents gradient vanishing and feature degradation through the residual propagation mechanism, thereby improving the stability and convergence speed of the network training process. The concatenated features are then processed through... Convolution is used for fusion and channel compression to obtain output features. ;

[0062] The internal computation of the edge information extractor follows a closed-loop logic of "local transformation—smoothing contrast—differential edge extraction—gated compression"; the input features of any scale branch... First through convolution Complete channel alignment and redundancy compression, then... convolution The local structural response is extracted and then upsampled to a spatial scale consistent with the branch features to obtain the intermediate features. :

[0063]

[0064] To explicitly separate high-frequency components at the edges, intermediate features are introduced. Scale-consistent local average pooling smoothing term (kernel size) (stride=1, padding=1), we get :

[0065]

[0066] in, The scale index represents the multi-scale branch; Indicates the first The intermediate feature maps obtained after convolution transformation and upsampling of each scale branch, and,

[0067]

[0068] in, , and These represent intermediate feature maps. The number of channels, height, and width; Indicates the core size is A local average pooling operation with a step size of 1 and a fill width of 1 is performed on each channel without changing the number of channels or the spatial size of the feature map. Indicates the intermediate feature map The smooth feature map obtained after local average pooling has a size similar to... Maintaining consistency, that is:

[0069]

[0070] High-frequency residuals are generated through differential methods. :

[0071]

[0072] This difference is equivalent to extracting the locally abrupt changes that would otherwise be smoothed out, making the contours, boundary transitions, and fine-grained textures more prominent in the feature domain; subsequently, the high-frequency residuals are... Send in Convolution followed by sigmoid mapping Obtain edge response terms :

[0073]

[0074] At the branch level of EIPE, edge response terms As an edge enhancement term, it is injected back into the scale branch features in a residual manner to form an enhanced output. :

[0075]

[0076] At the multi-scale fusion level, the enhanced output of each scale branch After channel concatenation and fusion with the original input features from the bypass, a joint representation is obtained that combines fine-grained boundary sensitivity with global contextual consistency. Because edge enhancement is performed in parallel at multiple scales, small target pests can obtain stable contour responses under different resolutions and texture complexities. At the same time, bypass semantic information and multi-scale context are uniformly integrated into the output features. In this way, the differences between pests and easily confused areas such as leaf texture and branch shadows are more easily captured by the network, thus achieving more robust detection and localization performance under the constraint of edge computing power.

[0077] S200. After completing the backbone feature extraction in step S100, the traditional Neck structure is replaced by the Cross-Stage Weighted Feature Pyramid (CSWFPN) to unify and integrate features output from different scales and stages. The CSWFPN incorporates multi-scale features from each stage of the backbone network into the same fusion framework, using a bidirectional fusion link from bottom to top to complete cross-stage information transfer: the bottom-up path converges features from different levels step by step, enabling semantic and detail information to form a unified expression at a higher level; the top-down path performs backflow fusion on this basis, enabling high-level semantics to be back-compensated to a finer resolution level, avoiding information attenuation caused by relying solely on step-by-step upsampling / downsampling; during the bidirectional fusion process, the CSWFPN introduces a weighted fusion mechanism to adjust the contribution of features from different stages and scales, enabling the fusion result to adaptively highlight effective information in the case of complex background textures or strong noise, suppressing the dominant influence of redundant or interfering features on the fusion result, thereby improving the stability and consistency of multi-scale features in natural inspection scenarios.

[0078] The Cross-Stage Weighted Feature Pyramid Network (CSWFPN) addresses the challenges of low-altitude drone pest identification, such as large target scales, complex background textures, and the difficulty of capturing both detail and semantics with single-level features. It replaces the original neck feature fusion structure in the network, unifying and transmitting multi-scale features output from the backbone network within a single fusion framework. This structure takes four stages of features (P2-4, P3-8, P4-16, and P5-32) as input. Each stage first undergoes convolution to shape the channels, and then upsampling and downsampling align the features of different resolutions to the target. The standard scale ensures that subsequent fusion is strictly aligned and computable in both spatial and channel dimensions. The bottom-up path gradually converges low-level details and structural information to deeper feature representations, while the top-down path flows high-level semantic information back to a higher resolution level, achieving semantic compensation for details and detail constraint on semantics. At the same time, cross-stage bypass input is introduced at key nodes, enabling fusion nodes to simultaneously receive multi-source feature streams from different stages and directions, reducing information attenuation caused by step-by-step transmission, and more accurately distinguishing pests from the background in complex natural scenes while simultaneously reducing false detections and false negatives.

[0079] The fusion process of CSWFPN is accomplished by two types of structural units. The first type is the Concat fusion unit: the aligned features are concatenated at the same scale, and feature reshaping and nonlinear refinement are performed through C2f. This is used to fully couple the semantic information of the upsampled backflow with the detailed information at the same scale, forming a structured multi-scale aggregated representation. The second type is the Fusion1, Fusion2, and Fusion3 weighted fusion units: Conv1×1 is introduced before these nodes to align the channels of each input, so that features from different sources enter the same fusion space. Then, learnable weights are introduced in the fusion nodes to perform normalized weighted fusion of two or three inputs. The fusion result is then further refined by C2f and passed to a deeper Fusion layer through downsampling, so that the weighting calibration is continuously effective at multiple scales. The final output forms a three-layer pyramid feature, which directly connects to the detection head. The interface remains stable while the fusion capability is significantly enhanced. The key gain of CSWFPN comes from the learnable weighting mechanism introduced in the fusion nodes.

[0080] For the fusion of two inputs, let the feature maps of the two inputs be respectively... and The corresponding learnable weights are respectively and Two-way fused output feature map Defined as:

[0081]

[0082] For the fusion of three inputs, let the feature maps of the three inputs be respectively... , and The corresponding learnable weights are respectively , and Three-way fusion output feature map Defined as:

[0083]

[0084] Among them, the upper right corner is marked " "Indicates feature fusion of two paths, marked in the upper right corner" "Indicates the fusion of three features; and This represents the feature maps from different sources of the two input fusion nodes. , and This represents the feature maps from different sources of the input three-way fusion nodes; , and Indicates the source or scale index of each input feature; , , , and These are learnable weights that correspond one-to-one with the corresponding input features, used to dynamically adjust the contribution of features from different sources to the fusion result; These are preset positive constants used to avoid zero denominators and improve the stability of numerical calculations; the input features and weight parameters corresponding to different fusion nodes are set separately and updated during model training.

[0085] The normalization term keeps the output within a consistent numerical range, preventing unbounded growth of amplitude caused by increased weights, and ensuring numerical stability and smooth gradient propagation during the fusion process. The weight factors are learned through training and can automatically increase the contribution of more reliable scales and suppress the interference of noisy scales under different aerial photography altitudes, viewpoint changes, lighting conditions and background texture intensity. This allows multi-scale information to participate in the classification and bounding box regression of the detection head in a more reasonable proportion, ultimately resulting in improved global context preservation capabilities and a significant reduction in information loss in complex natural scenes.

[0086] S300. After completing the EIPE enhanced feature extraction in step S100 and the CSWFPN cross-stage weighted fusion in step S200, the fused multi-scale features are input into the detection head for target classification and position regression output. Since the front-end feature extraction stage enhances the fine-grained structural expression of small targets, and the fusion stage establishes cross-stage information connectivity and improves the multi-scale collaborative effect through the weighting mechanism, the detection head can maintain a more stable pest target response under different flight altitudes, different lighting conditions, and complex leaf backgrounds, thus achieving reliable positioning and identification of small-scale pest targets.

[0087] Specific Implementation Scheme 2: The present invention provides an EIPE-CSWFPN enhanced YOLO recognition system for low-altitude inspection of cold-region agriculture. This system has program modules corresponding to the above steps, and executes the steps in the above-mentioned EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection of cold-region agriculture when running.

[0088] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0089] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to implement, when called by a processor, the steps of an EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture.

[0090] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0091] Simulation Experiment

[0092] Model training and validation experiments were conducted using the IP102 pest detection dataset and a self-made aerial photography dataset of cold-region agricultural pests. The IP102 dataset is a publicly available benchmark dataset for pest detection, covering 102 common agricultural pests, with a total of over 75,000 images, of which approximately 19,000 images have object detection bounding box annotations.

[0093] Meanwhile, to further verify the model's generalization ability and engineering applicability in real-world UAV application scenarios, a self-made aerial photography dataset of cold-region agricultural pests was constructed and introduced. This dataset was collected by UAVs during the growing season of typical cold-region agricultural crops, covering major cold-region food crops such as rice, corn, and soybeans. The aerial photography focused on real field environments under conditions of low temperature, strong winds, short light periods, and complex backgrounds. The collected images have resolutions mainly ranging from 1920×1080 to 3840×2160, covering various natural lighting scenarios including sunny days, cloudy days, and partial shadows. The dataset contains 4,200 UAV aerial images, of which 3,600 images have been manually labeled with bounding boxes, accumulating approximately 18,000 labeled pest instances, covering 15 categories of common or prevalent cold-region agricultural pests, specifically including pests with distinct cold-region characteristics such as rice planthoppers, corn borers, and soybean pod borers. All labels use rectangular bounding boxes, and a two-person cross-validation process was used to ensure the consistency and accuracy of the labeling results.

[0094] In the aforementioned datasets, class imbalance exists to varying degrees in the sample distribution. Particularly in the IP102 dataset, some classes have nearly 3,000 samples, while others have fewer than 100. This can easily lead to insufficient learning of long-tail classes and unstable detection recall during model training. To address this issue, targeted data augmentation was performed on the classes with fewer samples. Methods such as random horizontal flipping, random contrast enhancement, random saturation transformation, adding Gaussian noise, and adding salt-and-pepper noise were used to increase sample diversity, thereby increasing the number of effective training samples and alleviating class bias without changing the semantic boundaries of the classes. After data augmentation, the entire dataset was randomly divided into training, validation, and test sets in a 7:2:1 ratio for model training, parameter tuning, and performance evaluation.

[0095] The experimental environment was configured as follows: Python 3.10 was used as the programming language, and Ubuntu 22.04 LTS was the operating system. The deep learning framework was built on PyTorch 1.9 and ran in an accelerated environment with CUDA 12.1 and cuDNN 8. The hardware platform was equipped with an NVIDIA GeForce RTX 3090 graphics card, providing sufficient computing power for model training and validation.

[0096] Regarding hyperparameter settings, the training batch size is fixed at 64, and the total number of training epochs is fixed at 300. The optimizer is configured in adaptive (Auto) mode, which automatically selects the optimization method that best matches the current data and convergence state from the three strategies of SGD, Adam, and AdamW during training. This ensures efficient parameter updates and convergence stability, and reduces the uncertainty caused by manually selecting the optimizer.

[0097] To ensure the effectiveness of the training process is comparable and quantifiable, detection precision and recall were simultaneously calculated in the experiment. Detection precision was expressed as mAP50 with an IoU threshold of 0.5, and recall was... With accuracy Each by , , The result of counting is:

[0098]

[0099]

[0100] Experimental results

[0101] To balance airborne power consumption constraints with real-time inference requirements, this solution optimizes the architecture based on a lightweight YOLOv10s backbone and exports the trained weights in an edge-deployable format. The overall model parameter size is controlled at 8.9M, which can serve as a baseline for evaluating the complexity of storage usage, computing power overhead, and inference latency when deployed on UAV edge devices.

[0102] To evaluate the superiority of the model in pest detection, it was compared with commonly used deep learning object detection models. The experimental results are shown in the table below. Compared with the original YOLOv10s, this method improved mAP50 by 2.6% and recall by 2.1%. The recognition performance using the YOLOv10s method is shown in the figure below. Figure 4 As shown in the figure, the recognition effect using this method is as follows: Figure 5 As shown in the figure, these improvements demonstrate that EC-YOLO not only outperforms six popular object detection models, but also exhibits higher accuracy and robustness in detection.

[0103] Table 1 Algorithm Comparison

[0104]

[0105] Furthermore, to enhance the model's applicability in cold-region agriculture, comparative experiments were conducted on a self-made aerial photography dataset of cold-region agricultural pests. The results show that on this cold-region dataset, EC-YOLO maintains a stable improvement in mAP50 and recall compared to the original YOLOv10s, with mAP50 increasing by approximately 3.1% and recall by approximately 2.8%, indicating that the proposed method has good robustness and practicality under the complex environmental conditions of cold-region agriculture.

[0106] Table 2 Algorithm Comparison

[0107]

[0108] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. An EIPE-CSWFPN-enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture, characterized in that, Includes the following steps: S100: Using aerial images collected by drones as input, the C2F unit in YOLOv10 is replaced with the EIPE edge information perception enhancement module in the backbone network feature extraction stage. S200. After completing the backbone feature extraction in step S100, the Neck structure in YOLOv10 is replaced with the Cross-Stage Weighted Feature Pyramid (CSWFPN). Multi-scale features from each stage of the backbone network are incorporated into the same fusion framework. Cross-stage information transmission is completed using a bidirectional fusion link that is both bottom-up and top-down. A weighted fusion mechanism is introduced during the bidirectional fusion process to adjust the contribution of features from different stages and scales. S300. After completing the EIPE enhanced feature extraction in step S100 and the CSWFPN cross-stage weighted fusion in step S200, the multi-scale feature input obtained by the fusion is used to perform target classification and position regression output, so as to realize the reliable positioning and identification of small-scale pest targets.

2. The EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture according to claim 1, characterized in that: In step S100, the EIPE edge information perception enhancement module uses the input feature map Starting from this point, multiple parallel multi-scale branches are constructed to cover different receptive fields; each branch reshapes the features into a fixed spatial grid through adaptive average pooling, thereby obtaining coarse-to-fine contextual representations at the same level simultaneously; The branch features after adaptive average pooling are fed into the edge information extractor to generate edge enhancement terms, and then the enhanced branch output is obtained by superimposing the residuals of the branch features. The augmentation results of all scale branches are concatenated in the channel dimension, while retaining one path from the original input. The direct bypass path to the splicing end; the spliced ​​features are accessed via... Convolution is used for fusion and channel compression to obtain output features. .

3. The EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture according to claim 2, characterized in that: Step S100 specifically includes, Input features for any scale branch First through convolution Complete channel alignment and redundancy compression, then... convolution The local structural response is extracted and then upsampled to a spatial scale consistent with the branch features to obtain the intermediate features. : To explicitly separate high-frequency components at the edges, intermediate features are introduced. The scale-consistent local average pooling smoothing term is obtained. : in, The scale index represents the multi-scale branch; Indicates the first The intermediate feature maps obtained after convolution transformation and upsampling of each scale branch, and, in, , and These represent intermediate feature maps. The number of channels, height, and width; Indicates the core size is Local average pooling operation with a step size of 1 and a fill width of 1; High-frequency residuals are generated through differential methods. : high frequency residual Send in Convolution followed by sigmoid mapping Obtain edge response terms : At the branch level of the EIPE edge information perception enhancement module, the edge response item As an edge enhancement term, it is injected back into the scale branch features in a residual manner to form an enhanced output. : At the multi-scale fusion level, the enhanced output of each scale branch After channel concatenation and fusion with the original input features from the bypass, a joint representation, i.e., the output feature, is obtained, which combines fine-grained boundary sensitivity with global context consistency. .

4. The EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture according to claim 3, characterized in that: In step S200, in the cross-stage weighted feature pyramid (CSWFPN), features from four stages—P2-4, P3-8, P4-16, and P5-32—are used as input. Each path first undergoes channel shaping through convolution, and then upsampling and downsampling are used to align features of different resolutions to the target scale. The bottom-up path gradually converges low-level details and structural information to deeper feature representations, while the top-down path feeds back high-level semantic information to higher resolution levels, achieving semantic compensation for details and constraint of semantics by details. At the same time, cross-stage bypass inputs are introduced at key nodes, enabling the fusion node to simultaneously receive multi-source feature streams transmitted from different stages and directions.

5. The EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection in cold-region agriculture according to claim 4, characterized in that: In step S200, the fusion process of the cross-stage weighted feature pyramid (CSWFPN) is completed by two types of structural units. The first type is the Concat fusion unit: the aligned features are concatenated at the same scale, and the features are reshaped and nonlinearly refined by the C2f unit. This is used to fully couple the semantic information of the upsampled backflow with the detailed information at the same scale to form a structured multi-scale aggregated representation. The second type is the Fusion1, Fusion2, and Fusion3 weighted fusion units: Conv1×1 is introduced before these nodes to align the channels of each input, so that features from different sources enter the same fusion space; then, learnable weights are introduced within the fusion nodes to perform normalized weighted fusion of two or three inputs, and the fusion result is then connected to the C2f unit and passed to a deeper Fusion layer through downsampling, so that the weighting calibration is continuously effective at multiple scales; finally, the output forms a three-layer pyramid feature, which is directly connected to the detection head; For the fusion of two inputs, let the feature maps of the two inputs be respectively... and The corresponding learnable weights are respectively and Two-way fused output feature map Defined as: For the fusion of three inputs, let the feature maps of the three inputs be respectively... , and The corresponding learnable weights are respectively , and Three-way fusion output feature map Defined as: Among them, the upper right corner mark Indicates feature fusion from two paths, indicated by the upper right corner label. This indicates the fusion of three features; and This represents the feature maps from different sources of the two input fusion nodes. , and This represents the feature maps from different sources of the input three-way fusion nodes; , and Indicates the source or scale index of each input feature; , , , and These are learnable weights that correspond one-to-one with the corresponding input features; It is a preset positive constant.

6. An EIPE-CSWFPN enhanced YOLO identification system for low-altitude inspection in cold-region agriculture, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-5 above, and executes the steps in the above-described EIPE-CSWFPN enhanced YOLO recognition method for low-altitude inspection of cold-region agriculture when it is run.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the EIPE-CSWFPN enhanced YOLO identification method for low-altitude inspection of cold-region agriculture as described in any one of claims 1-5.

Images