Method for monitoring rare and endangered species in Yunnan province based on improved YOLOv11

Abstract

According to the improved YOLOv11-based Yunnan rare and endangered species monitoring method provided by the invention, the adaptive environment feature enhancement module is embedded, so that noise interference caused by special habitats such as high altitude, strong illumination and foggy conditions in the Yunnan region is effectively overcome, the robustness and stability of feature extraction in a complex environment are remarkably improved, and the accuracy of feature extraction is improved. The problem that in the prior art, the recognition accuracy is sharply reduced in the environment is solved. A traditional feature fusion structure is replaced by a cross-hierarchy feature interaction module, an efficient multi-hierarchy feature circulation interaction mechanism is established, the problem of missing detection of severely shielded targets and minimum population individuals is effectively solved, and the detection rate of the key targets is remarkably increased; by introducing the fine granularity identification module, fine difference characteristics among species are deeply mined, the problem of confusion of related species which are difficult to distinguish in the prior art is successfully solved, and the inter-class distinguishing precision is improved to a practical level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of protection of rare and endangered species in Yunnan, and particularly relates to a rare and endangered species monitoring method in Yunnan based on an improved YOLOv11. BACKGROUND

[0002] Yunnan is one of the regions with the most concentrated distribution of rare and endangered species in China, and has many rare species such as golden monkeys and green peacocks. However, due to the complex habitat environment (such as high altitude, dense forest, and fog), the small population size, and the high similarity in morphology of some closely related species, the traditional monitoring methods have obvious limitations in real-time and accuracy.

[0003] The existing endangered species monitoring methods have the following disadvantages: first, the adaptability to complex habitats is poor. In extreme environments such as high altitude, strong light, and rainforest fog, image feature extraction is easily disturbed, resulting in key identification features such as facial red spots of golden monkeys being covered by background noise; second, the detection ability for small populations and occluded targets is insufficient. For small-sized individuals such as infant mandrills, or targets such as western black-crowned gibbons occluded by vines, incomplete features often result in missed detection; third, the ability to distinguish closely related species is limited, and there is a semantic gap between protection practice, which cannot accurately distinguish species with similar morphology such as Qiaohua five-needle pine and Yunnan pine. SUMMARY

[0004] To solve the above technical problems, the present application provides a rare and endangered species monitoring method in Yunnan based on an improved YOLOv11, which can realize the monitoring and identification of rare and endangered species in Yunnan.

[0005] To achieve the above purpose, the technical scheme of the embodiments of the present application is as follows: In a first aspect, the present application provides a rare and endangered species monitoring method in Yunnan based on an improved YOLOv11, which comprises: obtaining an image dataset containing rare and endangered species in Yunnan in special habitats; inputting the image dataset into a trained species identification model for feature extraction and endangered species identification, and outputting the identification results; the identification results at least include the species category, location information, and target protection level information of the endangered species; wherein the species identification model is obtained by embedding a self-adaptive environment feature enhancement module at the output end of the first C3k2 module in the backbone network of the YOLOv11 model, replacing the FPN-PAN feature fusion network structure in the neck network with a cross-level feature interaction module, adding a fine-grained discrimination module before the classification convolution layer of the detection head, and performing stage-by-stage training; The adaptive environmental feature enhancement module is configured to process environmental noise in the image dataset through a dynamic filtering subnetwork and to strengthen species key identification features; the cross-level feature interaction module is configured to perform feature enhancement processing on the multi-level feature maps output by the backbone network based on a constructed three-level feature interaction channel of high-level semantics-mid-level structure-low-level details; and the fine-grained identification module is configured to perform feature fusion on the enhanced multi-scale feature maps to obtain a classification vector including common feature components and strengthened identification feature components.

[0006] In some embodiments, the adaptive environmental feature enhancement module includes an environmental noise suppression layer and a feature strengthening layer; and the adaptive environmental feature enhancement module is specifically configured to perform the following operations: The environmental noise suppression layer is based on dynamic convolution and is configured to perform regional analysis on the initial feature maps output by the first C3k2 module, to apply a convolution kernel matched with a current environmental interference type to different regions in the initial feature maps, and to obtain denoised feature maps; and the feature strengthening layer includes a channel attention module, a spatial attention module, and a coordinate attention mechanism, and is configured to sequentially perform channel attention weighting, spatial attention weighting, and coordinate information coding on the denoised feature maps to obtain enhanced feature maps.

[0007] In some embodiments, the environmental noise suppression layer is specifically configured to perform the following operations: The luminance pixel value of the initial feature maps is calculated and compared with a preset luminance pixel threshold to locate an overexposure region, and the Laplacian variance of the initial feature maps is calculated and compared with a preset variance threshold to locate a fog area scattering region; a dynamic learning mechanism is used to analyze the initial feature maps in real time to generate and adaptively adjust convolution kernel weight parameters matched with each region; for the overexposure region, a 7*7 low-pass filtering kernel is used for mean filtering; for the fog area scattering region, a 3*3 anisotropic filtering kernel is used for Gaussian filtering; and for a normal region, a 3*3 standard convolution kernel is maintained; the convolution kernel weight parameters are applied to the initial feature maps, and a region-by-region weighted convolution is performed to output the denoised feature maps.

[0008] In some embodiments, the feature strengthening layer is specifically further configured to perform the following operations: The channel attention module is used for performing channel importance evaluation on the de-noised feature map, calculating global average pooling of each channel, learning weights of each channel through a fully connected layer, and generating a channel weighted feature map; a first convolution kernel is used for extracting spatial correlation features, spatial weight distribution is performed on the channel weighted feature map to generate a spatial attention map, and weight focusing is performed on a target region to suppress irrelevant backgrounds, so that a spatial weighted feature map is output; coordinate information coding is performed on the spatial weighted feature map, the spatial weighted feature map is decomposed into horizontal and vertical direction position features, and target coordinate information is embedded to strengthen the perception of the target spatial position, so that the enhanced feature map is output.

[0009] In some embodiments, the multi-level feature map includes a shallow feature map, a middle feature map and a deep feature map; and the cross-level feature interaction module is specifically used for performing the following operations: A second convolution kernel is used for extracting a species category embedding vector and a habitat scene feature vector from the deep feature map; a habitat attention map is generated by calculating the matching degree between the species and the habitat; after 2 times upsampling of the deep feature map and the habitat attention map, cross-scale fusion is performed on the middle feature map, habitat semantic weight is assigned to the middle feature map through a dynamic weight coefficient, and a first fused middle feature map is output; after a third convolution kernel is used for extracting the structural features of the first fused middle feature map, a part attention mechanism is introduced, the structural features are processed based on a pre-stored landmark part template of a species specimen library, and a part attention map is generated; after 2 times upsampling of the part attention map and the first fused middle feature map, fusion is performed on the shallow feature map, and the structural constraint of the first fused middle feature map is transmitted to the shallow feature map through a residual connection, so that an optimized shallow feature map is output; after convolution operation is performed on the optimized shallow feature map by combining a Laplacian operator, gradient enhancement is performed on the extracted fuzzy edge region according to a gradient value dynamic adjustment activation function, and edge gradient anomalies are amplified through pixel-by-pixel weighting, so that an enhanced edge feature map is generated; after 0.5 times downsampling of the optimized shallow feature map and the enhanced edge feature map, secondary fusion is performed on the first fused middle feature map, so that a second fused middle feature map balancing structure and semantics is output; the second fused middle feature map is downsampled and fused with the deep feature map, so that a fused deep feature map with enhanced habitat correlation and boundary prediction ability is output.

[0010] In some embodiments, the fine-grained discrimination module includes a feature library association layer and an attention focusing layer; and the fine-grained discrimination module is specifically used for performing the following operations: The feature library association layer is configured to decouple the inputted optimized shallow feature map, the second fused middle feature map and the fused deep feature map, separate common features and discriminative features of species, match the discriminative features with reference features of close species in a preset endangered species specimen feature library, and generate a difference heat map; the attention focusing layer is configured to generate discriminative attention weights based on the difference heat map, and fuse the discriminative features after reinforcement according to the discriminative attention weights with the common features to form a classification vector for fine-grained species classification.

[0011] In some embodiments, the feature library association layer is further configured to perform the following operations: The optimized shallow feature map, the second fused middle feature map and the fused deep feature map are subjected to scale alignment and channel compression, and are spliced to form a target fused feature map; the target fused feature map is input into a feature decoupling unit in the feature library association layer, and features are separated through two parallel convolution branches; a first convolution branch adopts 3x3 convolution to extract cross-species common features, and outputs the common feature components; a second convolution branch combines a spatial attention mechanism, adopts 1x1 convolution and global average pooling to extract discriminative feature components specific to species.

[0012] In some embodiments, the attention focusing layer is further configured to perform the following operations: A preset endangered species specimen feature library is called to obtain reference feature templates of close species, and a reference feature matrix is formed; cosine similarity between the discriminative feature components and the reference feature matrix is calculated, and a difference heat map is generated; the difference heat map is normalized into discriminative attention weights, the discriminative attention weights are multiplied point by point with the discriminative feature components to obtain reinforced discriminative feature components; the reinforced discriminative feature components are fused with the common feature vector through residual connection to form a fused feature; the fused feature is subjected to global average pooling to compress and generate the classification vector.

[0013] In some embodiments, after the improved YOLOv11 algorithm is trained based on the image data set to obtain a trained species recognition model, the method further comprises: The trained species recognition model is subjected to lightweight processing and is deployed to an edge computing device; real-time monitoring images are analyzed locally by the edge computing device, and when an endangered species or an abnormal behavior of the endangered species is recognized, a warning information is automatically generated and sent.

[0014] In some embodiments, the species recognition model is trained through the following process: Initialize the algorithm with a sample data set in a general scenario, freeze the parameters of the adaptive environment feature enhancement module, the cross-level feature interaction module and the fine-grained discrimination module, and train the YOLOv11 basic architecture; obtain samples containing environmental interference, and label species categories, bounding boxes, species key discrimination features and species corresponding protection levels to construct a training data set; unfreeze all parameters, input the training data set into the trained YOLOv11 basic architecture for training, and dynamically adjust the weights of each module; and perform intensive training on samples of closely related species and critically endangered species to obtain the trained species recognition model.

[0015] The YOLOv11-based Yunnan rare and endangered species monitoring method provided by the application effectively overcomes the noise interference caused by the special habitats such as high altitude, strong light and fog in Yunnan, significantly improves the robustness and stability of feature extraction in complex environments, and solves the problem of sharp decline in recognition accuracy in such environments in the prior art; the cross-level feature interaction module replaces the traditional feature fusion structure, and an efficient multi-level feature cyclic interaction mechanism is established, effectively solving the problem of missing detection of severely occluded targets and extremely small population individuals, and significantly improving the detection rate of such key targets; by introducing the fine-grained discrimination module, the subtle difference features between species are deeply mined, and the confusion problem of closely related species that is difficult to distinguish in the prior art is successfully solved, and the inter-class discrimination accuracy is improved to a practical level. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the YOLOv11-based Yunnan rare and endangered species monitoring method provided by the embodiments of the application; Figure 2 is an architecture diagram of the improved YOLOv11 model provided by the embodiments of the application; Figure 3 is an architecture diagram of the backbone network in the improved YOLOv11 model provided by the embodiments of the application; Figure 4 is an architecture diagram of the neck network in the improved YOLOv11 model provided by the embodiments of the application; Figure 5 is an architecture diagram of the detection head in the improved YOLOv11 model provided by the embodiments of the application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the application. All other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0018] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" can be the same subset or a different subset of all possible embodiments, and can be combined with each other as long as there is no conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as understood by those skilled in the art to which the embodiments of the present application belong. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0019] The following describes an exemplary application of the improved YOLOv11-based Yunnan rare and endangered species monitoring device provided by the embodiments of the present application. The improved YOLOv11-based Yunnan rare and endangered species monitoring device provided by the embodiments of the present application can be implemented as a terminal or a server. In one implementation, the improved YOLOv11-based Yunnan rare and endangered species monitoring device provided by the embodiments of the present application can be implemented as various types of terminals such as a notebook computer, a tablet computer, a desktop computer, and a mobile device. In another implementation, the improved YOLOv11-based Yunnan rare and endangered species monitoring device provided by the embodiments of the present application can also be implemented as a server. The server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, etc. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiments of the present application. In the following, an exemplary application of the improved YOLOv11-based Yunnan rare and endangered species monitoring device as a server will be described.

[0020] The embodiments of the present application provide an improved YOLOv11-based Yunnan rare and endangered species monitoring method. Referring to Figure 1 , Figure 1 is a flowchart of an improved YOLOv11-based Yunnan rare and endangered species monitoring method provided by the embodiments of the present application. The steps shown in Figure 1 will be described.

[0021] In step S110, an image data set containing Yunnan rare and endangered species in special habitats is obtained.

[0022] In some embodiments, the rare and endangered species in Yunnan refers to a species distributed in Yunnan Province, which is threatened with extinction due to habitat destruction, small population size or severe threats, and is listed in the National Key Protected Wild Animal List, the National Key Protected Wild Plant List or the Yunnan Provincial Local Protection List. For example, Asian elephant, Yunnan golden monkey, and Fagraea faginea.

[0023] In some embodiments, the special habitat refers to a unique and non-universal living environment in Yunnan region, including high-altitude alpine meadow, tropical monsoon rainforest, karst cave, high mountain lake, etc. The light, humidity, vegetation background and other conditions of such environment are significantly different from the conventional environment, which can interfere with species image collection.

[0024] In some embodiments, the image data set is a collection of multiple images containing rare and endangered species in Yunnan and their special habitats. The data set not only contains images, but also contains corresponding standard information such as species category, target bounding box coordinate, protection level, etc., which are used for training, verification or testing of species identification model.

[0025] In step S120, the image data set is input into the trained species identification model for feature extraction and endangered species identification, and the identification result is output. The identification result at least includes the species category, location information and target protection level information of the endangered species.

[0026] The species identification model is obtained by embedding a self-adaptive environment feature enhancement module at the output end of the first C3k2 module in the backbone network of the YOLOv11 model, replacing the FPN-PAN feature fusion network structure in the neck network with a cross-level feature interaction module, adding a fine-grained discrimination module before the classification convolution layer of the detection head, and performing stage-by-stage training. The self-adaptive environment feature enhancement module is used to process the environmental noise in the image data set through a dynamic filtering sub-network, and to strengthen the key discriminant features of the species. The cross-level feature interaction module is used to perform feature enhancement processing on the multi-level feature maps output by the backbone network based on the constructed three-level feature interaction channel of high-level semantics-middle-level structure-low-level detail. The fine-grained discrimination module is used to perform feature fusion on the enhanced multi-scale feature maps to obtain a classification vector including common feature components and enhanced discriminant feature components.

[0027] In some embodiments, the species identification model is a target detection model based on YOLOv11, which integrates a self-adaptive environment feature enhancement module, a cross-level feature interaction module and a fine-grained discrimination module, and is obtained by stage-by-stage training. It is specially used for identifying rare and endangered species in Yunnan in special habitats.

[0028] In some embodiments, feature extraction refers to one of the core steps of the model processing the image, which refers to extracting information that can represent the characteristics of the species (such as texture, contour, structure) from the original image through convolution, pooling, fusion, etc. operations of the backbone network, neck network, etc. operation, and converting it into a feature map or feature vector that can be processed by a computer.

[0029] In some embodiments, the core structure of the YOLOv11 model is divided into three parts: backbone network, neck network, and head network.

[0030] Among them, the backbone network is the front part of the YOLOv11 model, mainly used to extract basic features from the input image, through multiple layers of convolution, pooling, etc. operation, the original image is converted into a feature map of different scales, providing basic features for subsequent feature fusion and recognition. The C3k2 module is the core residual structure module in the YOLOv11 backbone network, C3 represents the convolution block of residual connection, k2 represents the use of 2x2 size convolution kernel in the module, the main function is to reduce the gradient disappearance while extracting features, and enhance the model's reuse ability of features.

[0031] In some embodiments, the neck network is an intermediate structure connecting the backbone network and the head network, and its core function is to fuse the multi-level features output by the backbone network, and improve the model's detection ability for different scale targets, such as Asian elephants and rare insects. The original YOLOv11 commonly uses FPN-PAN structure.

[0032] In some embodiments, the head network is the output part of the YOLOv11 model, which processes the fused features output by the neck network through classification convolution layer, regression convolution layer, etc. and finally outputs the identification results such as species category and position coordinates of the target. The classification convolution layer is a key layer in the head network responsible for classifying, which converts the fused feature map into a probability vector of the corresponding species category through convolution operation, for example, outputting the probability value of Yunnan golden monkey and non-Yunnan golden monkey.

[0033] In some embodiments, the adaptive environmental feature enhancement module is an improved module embedded in the output end of the first C3k2 module of the backbone network in the present application, and the core is to contain a dynamic filtering sub-network. The dynamic filtering sub-network can dynamically adjust the filtering parameters according to the noise type of the special habitat in the input image, such as strong light spot and complex vegetation shadow, and filter environmental noise; it can strengthen the key identification features of the species such as the feather texture of birds and the leaf shape of plants while reducing noise.

[0034] In some embodiments, the cross-level feature interaction module is an improved module in the application for replacing the neck network FPN-PAN structure, the core of which is to build a three-level feature interaction channel of high-level semantic-middle-level structure-low-level detail. Here, the low-level detail feature: from the shallow layer of the backbone network, high resolution, contains rich texture, edge details, and uses precise positioning. The middle-level structure feature: from the middle layer of the backbone network, balances details and semantics, and contains object component and contour structure information. The high-level semantic feature: from the deep layer of the backbone network, low resolution, contains global and abstract semantic information of what species. The module establishes a direct and bidirectional interaction channel between the features at the three levels through a carefully designed connection path (such as dense connection, cross attention mechanism), so that the high-level semantic can guide the understanding of the middle-level and low-level features, and the low-level details can in turn optimize the positioning of the high-level, thereby realizing more robust feature enhancement.

[0035] In some embodiments, the fine-grained discrimination module is an improved module added in the application before the classification convolutional layer of the head network, the core of which is to fuse the enhanced multi-scale feature maps.

[0036] In some embodiments, the common feature component refers to features common to different species, such as the four limbs of mammals.

[0037] In some embodiments, the discriminative feature component refers to features unique to a species that can be distinguished from other species, such as the upturned nose of the Yunnan golden monkey.

[0038] In the embodiments of the application, the classification vector containing commonness + strengthened discrimination is obtained by fusion, which improves the accuracy of species classification, especially for distinguishing species with similar shapes.

[0039] The YOLOv11-based Yunnan rare and endangered species monitoring method provided by the application effectively overcomes the noise interference caused by special habitats such as high altitude, strong light, and fog in Yunnan, significantly improves the robustness and stability of feature extraction in complex environments, and solves the problem of sharp decline in recognition accuracy in such environments in the prior art; the cross-level feature interaction module replaces the traditional feature fusion structure, establishing an efficient multi-level feature cyclic interaction mechanism, effectively solving the problem of missing detection of severely occluded targets and extremely small population individuals, and significantly improving the detection rate of such key targets; by introducing the fine-grained discrimination module, the subtle difference features between species are deeply mined, successfully solving the confusion problem of closely related species that is difficult to distinguish in the prior art, and improving the inter-class discrimination accuracy to a practical level.

[0040] In some embodiments, the adaptive environment feature enhancement module comprises an environment noise suppression layer and a feature enhancement layer; and the adaptive environment feature enhancement module is specifically configured to perform the following operations: First, the environment noise suppression layer is implemented based on dynamic convolution, and is configured to perform region analysis on the initial feature map output by the first C3k2 module, and apply a convolution kernel matched with the current environment interference type to different regions in the initial feature map to obtain a denoised feature map.

[0041] In some embodiments, the first sub-layer of the adaptive environment feature enhancement module is configured to reduce the interference of environmental noise in a special habitat on the features of a species by using a dynamic convolution technique. The working logic is as follows: first, the initial feature map output by the first C3k2 module of the backbone network is divided into regions, and then a convolution kernel matched with the type of environmental interference in different regions, such as a high-reflectivity region, a shadow overlap region, a complex vegetation background region, etc., is dynamically generated or selected. Here, the size and weight parameters of the convolution kernel are adaptively adjusted according to the characteristics of the region, and finally the noise is filtered through convolution operation to output a denoised feature map.

[0042] In some embodiments, the initial feature map refers to the feature map output from the first C3k2 module of the backbone network. It represents the primary features extracted by the network and contains rich detailed information, but is also most susceptible to pollution by environmental noise in the original image.

[0043] In some embodiments, region analysis refers to the process of scanning or global analysis of the initial feature map by the dynamic convolution sub-network, aiming to identify the main interference type suffered by different regions, such as determining whether a region is a shadow, fog, or an occlusion, so as to assign appropriate dynamic convolution kernel parameters to each region.

[0044] In some embodiments, the denoised feature map refers to the output after processing by the environment noise suppression layer. Compared with the initial feature map, the part related to the background environmental noise in the feature response is weakened, while the contour and structural information of the target object is better preserved.

[0045] In this embodiment, unlike traditional fixed-parameter enhancement methods, the adaptive environment feature enhancement module can automatically adjust the denoising and enhancement strategy according to the specific environment of each image without human intervention. For foggy areas, the dynamically generated convolution kernel may tend to perform contrast enhancement; for rain and snow streak areas, the generated kernel may tend to directional filtering to remove the streaks; for motion blur areas, the generated kernel may perform deconvolution-based sharpening recovery, thereby achieving a filtering effect matched with the current environment interference type.

[0046] Secondly, the feature enhancement layer includes a channel attention module, a spatial attention module and a coordinate attention mechanism, which are used to sequentially perform channel attention weighting, spatial attention weighting and coordinate information coding on the de-noised feature map to obtain an enhanced feature map.

[0047] In some embodiments, the feature enhancement layer is a second sub-layer of the adaptive environmental feature enhancement module, located after the environmental noise suppression layer, and the core function is to highlight the key identification features of the species (such as the unique fur color of animals and the special leaf veins of plants) and suppress irrelevant background features based on de-noising through multi-layer attention mechanisms.

[0048] Here, the channel attention module first performs global average pooling on the feature map, compressing the spatial information of each channel into a scalar. Then, a small neural network is used to learn the importance weight of each channel. Finally, the weight is used to weight the entire feature channel.

[0049] Spatial attention module: After channel attention weighting, this module analyzes the information importance of different spatial positions (pixel points or local regions) in the feature map (for example, pixels in the species area are given high weight, and blank background or non-target area is given low weight), generates a spatial attention weight map, and then performs spatial weighting with the feature map to further highlight the spatial position features of the species in the image.

[0050] Coordinate attention mechanism, an attention mechanism that encodes channel information and spatial coordinate position at the same time, supplements the position coordinate information on the basis of channel attention and spatial attention. The core is to encode the coordinate information of the width and height directions of the feature map into two one-dimensional feature vectors, respectively, and generate attention weights after fusion with channel features, so that the model not only pays attention to which features are important, but also can clearly know the specific coordinate position of these features in the image, for example, accurately enhancing the coordinate information of key parts such as the beak and tail of birds, and improving the spatial positioning accuracy of features.

[0051] In some embodiments, the environmental noise suppression layer is specifically configured to perform the following operations: Firstly, the overexposure region is located by calculating the luminance pixel value of the initial feature map and the preset luminance pixel threshold, and the fog area scattering region is located by calculating the Laplacian variance of the initial feature map and the preset variance threshold. Secondly, the initial feature map is analyzed in real time through a dynamic learning mechanism to generate and adaptively adjust the convolution kernel weight parameters matched with each region; wherein, for the overexposure region, mean filtering is performed through a low-pass filtering kernel with a size of 7x7; for the fog area scattering region, Gaussian filtering is performed through an anisotropic filtering kernel with a size of 3x3; and for the normal region, a standard convolution kernel with a size of 3x3 is kept. Finally, the convolution kernel weight parameters are applied to the initial feature map, and the denoised feature map is output through region-by-region weighted convolution.

[0052] Here, the core of the dynamic learning mechanism is the MLP network, which dynamically generates the optimal convolution kernel weight according to the texture complexity, illumination intensity, target proportion and other indicators by analyzing the interference type of the current environment (such as the overexposure area, the fog area and the blur area). A large-scale smoothing kernel is used for the overexposure area, and an edge-preserving kernel is used for the fog area, so as to realize targeted denoising.

[0053] In some embodiments, the feature enhancement layer is specifically further configured to perform the following operations: Firstly, the channel importance of the denoised feature map is evaluated through the channel attention module, the global average pooling of each channel is calculated, and the weight of each channel is learned through a fully connected layer to generate a channel weighted feature map. Secondly, spatial correlation features are extracted through a first convolution kernel, spatial weight distribution is performed on the channel weighted feature map to generate a spatial attention map, and the weight is focused on the target region to suppress irrelevant background, and a spatial weighted feature map is output. Finally, the spatial weighted feature map is encoded with coordinate information, the spatial weighted feature map is decomposed into horizontal and vertical position features, and the target coordinate information is embedded to enhance the perception of the target spatial position, and the enhanced feature map is output.

[0054] Here, the feature enhancement layer cascades channel attention (SE) and spatial attention (PSA) to let the network focus on the key feature area of the species, such as the eye spot of the tail feather of the green peacock; and a coordinate attention mechanism is introduced to enhance the perception ability of the model to the target spatial position.

[0055] In some embodiments, the multi-level feature map includes a shallow feature map, a middle feature map and a deep feature map; and the cross-level feature interaction module is specifically configured to perform the following operations: The species category embedding vector and the habitat scene feature vector are extracted from the deep feature map by compressing channels through a second convolution kernel; a habitat attention map is generated by calculating the matching degree between the species and the habitat; the deep feature map and the habitat attention map are 2 times up-sampled, and then cross-scale fusion is performed with the middle layer feature map, the habitat semantic weight is assigned to the middle layer feature map through a dynamic weight coefficient, and a first fused middle layer feature map is output; after extracting the structural features of the first fused middle layer feature map through a third convolution kernel, a part attention mechanism is introduced, the structural features are processed based on the pre-stored landmark part template of the species specimen library, and a part attention map is generated; the part attention map and the first fused middle layer feature map are 2 times up-sampled, and then fusion is performed with the shallow layer feature map, and the structural constraint of the first fused middle layer feature map is transmitted to the shallow layer feature map through residual connection, and an optimized shallow layer feature map is output; after convolution operation is performed on the optimized shallow layer feature map by combining the Laplacian operator, the gradient-enhanced fuzzy edge region is dynamically adjusted according to the gradient value of the activation function, and the edge gradient anomaly is amplified through pixel-by-pixel weighting, and an enhanced edge feature map is generated; the optimized shallow layer feature map and the enhanced edge feature map are 0.5 times down-sampled, and then secondary fusion is performed with the first fused middle layer feature map, and a second fused middle layer feature map balancing structure and semantics is output; the second fused middle layer feature map is down-sampled and fused with the deep feature map, and a fused deep feature map with enhanced habitat association and boundary prediction ability is output.

[0056] The "species-habitat" semantic information in the high-level feature is used to generate an attention map, which guides the middle-level feature to focus on the potential activity area and reduces background interference; the species contour information in the middle-level feature is used in combination with the landmark part template to generate a part attention map, which constrains the low-level feature to focus on the key identification part and prevents detail diffusion; the edge information in the low-level feature (P3) is gradient-enhanced to amplify the subtle edge changes of the occluded target, and the enhanced details are fed back to the middle and high levels to improve the boundary prediction ability for incomplete targets.

[0057] Unlike the traditional feature fusion which is one-way or simply two-way, the cross-level feature interaction module realizes the circulation and mutual enhancement of information among the high, middle and low levels, and especially strengthens the transmission of weak features and occluded features.

[0058] In some embodiments, the fine-grained identification module includes a feature library association layer and an attention focusing layer; the fine-grained identification module is specifically configured to perform the following operations: Firstly, the feature library association layer is used to decouple the inputted optimized shallow feature map, the second fused middle layer feature map and the fused deep layer feature map, separate the common features and the discriminative features of species, and match the discriminative features with the reference features of related species in the preset endangered species specimen feature library to generate a difference heat map. Secondly, the attention focusing layer is used to generate discriminative attention weights based on the difference heat map, and the discriminative features after reinforcement are fused with the common features to form a classification vector for fine-grained species classification according to the discriminative attention weights.

[0059] In some embodiments, the feature library association layer is further used to perform the following operations: Firstly, the optimized shallow feature map, the second fused middle layer feature map and the fused deep layer feature map are subjected to scale alignment and channel compression to form a target fused feature map. Secondly, the target fused feature map is input into a feature decoupling unit in the feature library association layer, and feature separation is performed through two parallel convolution branches; the first convolution branch adopts 3x3 convolution to extract cross-species common features, and outputs the common feature component; the second convolution branch combines a spatial attention mechanism, adopts 1x1 convolution and global average pooling to extract species-specific discriminative feature components.

[0060] In some embodiments, the attention focusing layer is further used to perform the following operations: Firstly, a preset endangered species specimen feature library is called to obtain reference feature templates of related species to form a reference feature matrix; the cosine similarity between the discriminative feature components and the reference feature matrix is calculated, and a difference heat map is generated. Secondly, the difference heat map is normalized into discriminative attention weights, and the discriminative attention weights are multiplied point by point with the discriminative feature components to obtain reinforced discriminative feature components. Finally, the reinforced discriminative feature components are fused with the common feature vector through residual connection to form a fused feature; the fused feature is subjected to global average pooling to compress to generate the classification vector.

[0061] In some embodiments, after the above step S120 is performed, the method further comprises: performing lightweight processing on the trained species recognition model, and deploying it to an edge computing device; performing local analysis on real-time monitoring images through the edge computing device, and automatically generating and sending early warning information when an endangered species or an abnormal behavior of the endangered species is identified.

[0062] In some embodiments, the method further comprises: initializing the algorithm with a sample data set in a general scenario, freezing the parameters of the adaptive environment feature enhancement module, the cross-level feature interaction module and the fine-grained discrimination module, training the YOLOv11 basic architecture; obtaining a training data set labeled with species category, bounding box, species key discrimination feature and species corresponding protection level; unfreezing all parameters, inputting the training data set into the trained YOLOv11 basic architecture for training, and dynamically adjusting the weights of each module; and training the trained species recognition model by strengthening the training of samples of closely related species and critically endangered species.

[0063] The application realizes real-time detection and early warning of edge devices of protected areas through lightweight deployment, dynamically adapts to species changes by combining an incremental learning mechanism, and builds a complete closed loop from data collection, model training to field monitoring and protection response, solving the fundamental defects of existing monitoring technologies, such as lagging update and inability to adapt to long-term dynamic changes; not only can output species category and location information, but also can provide multi-dimensional information such as protection level, behavior state and environmental interference, providing precise and efficient technical support far beyond traditional methods for endangered species such as Yunnan golden monkeys and green peacocks in Gaoligongshan and Baimashan protected areas.

[0064] Next, an exemplary application of the embodiment of the application in an actual application scenario will be described.

[0065] The application improves the YOLOv11 algorithm, adds an adaptive environment feature enhancement module, a cross-level feature module interaction and a fine-grained discrimination module on the basis of the original architecture, solves the above problems, and realizes monitoring and identification of Yunnan rare and endangered species, and the specific process steps are as follows: S1, image data sets of Yunnan rare and endangered species (such as Yunnan golden monkeys and green peacocks) in special habitats (high altitude, dense forest, fog area) are collected and labeled with species category, bounding box, species key discrimination feature and species corresponding protection level.

[0066] S2, based on the processed image data set, the algorithm is constructed and trained, three improved modules are embedded on the basis of the traditional YOLOv11 algorithm, and the algorithm is trained in stages, the detection accuracy of small population individuals and closely related endangered species is improved, the core indicators such as monitoring critically endangered species detection rate are dynamically optimized.

[0067] S3, based on the trained algorithm, the identification effect of the algorithm on endangered species is verified in a typical protected area test set (such as Baimashan and Yuanjiang River Basin), and the adaptability to complex environment, the ability to detect occluded targets and the fine-grained discrimination ability are evaluated.

[0068] S4, using the trained algorithm, light weight deployment to the edge device of the protected area, realizing real-time detection of endangered species, abnormal behavior warning (such as injured individuals) and data linkage to the protection management platform.

[0069] S5, periodically access the patrol data during the monitoring process, update the species feature library through incremental learning, adapt to the dynamic changes of the number and behavior of endangered species, and form a monitoring and identification protection closed loop.

[0070] The architecture diagram of the improved YOLOv11 model is shown in Figure 2 The adaptive environment feature module, cross-level feature interaction module and fine-grained discriminative feature module are shown in the figure.

[0071] The above step S1 carries out data preparation work around the unique habitat of Yunnan rare and endangered species and the protection needs. When collecting data, focus on the special habitats of typical species distribution, such as the high-altitude rock slope where the Yunnan golden monkey in Baima Mountain is active, the seasonal rain forest where the green peacock in Xishuangbanna is inhabited, and the fog area where the forest musk deer in Gaoligong Mountain frequently appears. Targeted shooting is carried out through the deployment of infrared cameras, drones and other equipment. During the collection process, special attention should be paid to the different growth stages of small population species (such as the Wujia pine in Qiaohai), the behavior of individuals in the breeding period (such as the West Black Crown Gibbon's parenting behavior), and the scenes disturbed by the environment (such as the lazy monkey blocked by vines and the forest musk deer active in the fog), to ensure that the samples can cover the diversified forms of the species in complex habitats. At the same time, in view of the technical challenges of special habitats, such as image overexposure caused by strong ultraviolet rays at high altitudes, target occlusion caused by multi-layer vegetation in dense forests, and feature blurring caused by low visibility in fog areas, the image quality needs to be improved by adjusting the equipment parameters (such as infrared mode switching, exposure compensation) to provide real scene data support for subsequent algorithm training.

[0072] In addition to accurately marking the species category and the coordinates of the bounding box, key discriminative features of the species need to be highlighted, such as the facial red spot position of the Yunnan golden monkey, the eye-shaped pattern distribution of the green peacock tail feathers, and the petal vein sequence of the umbrella tree. These detailed features are the core basis for distinguishing closely related species (such as Yunnan Taxus and southern Taxus). At the same time, each sample needs to be associated with the corresponding protection level, such as the "critically endangered" and "endangered" levels in the IUCN (International Union for Conservation of Nature) Red List or the national first / second protection identification, so that the data not only contains visual features, but also carries semantic information of species protection. The annotation process needs to be reviewed by scientific researchers familiar with Yunnan rare species to ensure the accuracy of the discriminative feature and protection level annotation, laying a data foundation for subsequent algorithm learning of fine-grained features and association of protection decision needs.

[0073] The step S1 obtains images covering rare and endangered species in Yunnan in special habitats such as high altitude, dense forest and fog area, and the dataset is labeled with species category, key identification features and corresponding protection level.

[0074] Step S2: Based on the labeled dataset output in step S1, the improved YOLOv11 algorithm is constructed and trained. Three improved modules are embedded in the traditional YOLOv11 architecture. An adaptive environmental feature enhancement module is introduced to process high-altitude strong light, fog scattering and other interference through a dynamic filtering subnetwork (including an environmental noise suppression layer and a feature enhancement layer) to strengthen the key identification features of the Yunnan golden monkey face red spot and the green peacock eye-like pattern. Through the three-order circular interaction of the high-level semantic guidance flow (transferring habitat semantic features), the middle-level detail supplement flow (strengthening the edge of the occluded target) and the low-level difference amplification flow (highlighting the morphological differences of closely related species) of the cross-level feature interaction module, the feature integrity of small population individuals (such as Yunnan blue fruit tree seedlings) and occluded targets (such as lazy monkeys occluded by vines) is improved. Finally, a fine-grained identification module is introduced to load the Yunnan rare and endangered species specimen library features (such as the needle leaf angle template of Qiaohai five-needle pine and Yunnan pine), focus on the identification points of closely related species through the multi-head self-attention mechanism, and associate the IUCN protection level label.

[0075] The training process adopts a phased strategy. The first phase is basic adaptation. The algorithm is initialized with general samples (such as unoccluded adult individuals) in the image dataset of step S1. The parameters of the improved module are frozen, and only the YOLOv11 basic architecture is trained to enable it to quickly learn the basic morphological features of Yunnan rare species, with the overall average precision improved to more than 70%. The second phase is module enhancement. All parameters are unfrozen, and samples containing special habitat interference (such as fog area forest musk deer and high-altitude Yunnan golden monkey) are input. By dynamically adjusting the module weight (such as enhancing the filtering strength of the adaptive environment module), the algorithm adapts to extreme environmental features. In this phase, the detection accuracy of small population individuals (such as infant pygmy marmoset) needs to be monitored, and the average precision is required to be improved by more than 15% compared with the basic phase. The third phase is precise optimization. The samples of closely related species and critically endangered species are intensively trained. The difficult case mining strategy (such as repeatedly inputting confusing samples) is adopted, and the critically endangered species is given a weight of 3 times in the loss function to ensure that its detection rate is stable at more than 95%. Core indicators are monitored in real time throughout the training process. The classification error of closely related species is analyzed through the confusion matrix, the missing detection situation of critically endangered species is tracked through the recall rate curve, and the learning rate is dynamically adjusted combined with the feedback of the validation set to finally form the improved YOLOv11 algorithm adapted to the regional characteristics of Yunnan.

[0076] The specific improvements are as follows: I) Adaptive environmental feature enhancement module: this module is embedded into the output end of the C3k2 module of the Backbone of the YOLOv11 algorithm architecture, that is, the output end of the first CSPDarkne module, and the framework structure diagram of the Backbone backbone network is as shown in Figure 3

[0077] This module is divided into two layers in the YOLOv11 algorithm architecture. One is the environmental noise suppression layer, which is based on the realization of dynamic convolution (DynamicConv), introduces a dynamic learning mechanism, constructs an adjustment function to realize adaptive adjustment of the convolution kernel weight, and filters the overexposure area or fog area scattering noise caused by strong light at high altitude. The second is the feature enhancement layer, which introduces the cascade structure of channel attention (SE module) and spatial attention (PSA module), focuses on species-specific features such as the edge contrast of the eye spot of the green peacock tail feathers, and at the same time, through the coordinate attention mechanism (Coordinate Attention), the target position perception is strengthened.

[0078] 1) Environmental noise suppression layer: after receiving the initial feature map F0 (size 320x320x64) output by the first C3k2 module in the Backbone, the local area analysis is performed on F0, and through the brightness threshold, such as pixel value > 240, it is determined as an overexposure area and the fog degree calculation, such as Laplacian variance < 50, it is determined as a fog area blur, and the high-altitude strong light overexposure area (such as the bare rock reflection area of the Yunnan golden monkey activity in Baima Snow Mountain) and the fog area scattering area (such as the thick fog area of the forest musk deer activity in Gaoligong Mountain) are located.

[0079] Then, according to the dynamic learning mechanism, the adaptive convolution kernel weight parameters are generated according to the interference type. For the overexposure area, a large-scale smoothing kernel with a size of 7x7 (weight biased towards mean filtering) is used to suppress pixel value mutation; for the fog area blur area, a 3x3 edge preserving kernel (weight biased towards Gaussian filtering) is used to reduce edge diffusion; for the normal area, a 3x3 standard convolution kernel is used to extract details.

[0080] Finally, the dynamically generated convolution kernel weight parameters are applied to F0, and through the region-by-region weighted convolution (overexposure area weight 0.8, fog area weight 0.7, normal area weight 1.0), the denoised feature map F1 (320x320x64) is output, in which the gray scale of the overexposure area tends to be balanced, and the edge of the fog area (such as the outline of the four limbs of the forest musk deer) is clear.

[0081] ​2) Feature enhancement layer: This layer receives the de-noised feature map F1 (320x320x64) output by the ambient noise suppression layer. First, channel attention (SE module) is used to evaluate the importance of F1 in each channel, calculate the global average pooling value of each channel, and learn the weight through a fully connected layer (such as the green peacock tail feather eye spot channel weight is increased to 1.2, and the background vegetation channel weight is reduced to 0.3). The channel response of the species-specific feature is highlighted, and the output channel weighted feature map F2 (320x320x64) enters the spatial attention (PSA module), i.e., the spatial weight distribution of F2 is performed through a 7x7 convolution to extract spatial correlation features and generate a spatial attention map (320x320x1). The weight is focused on the target area (such as the needle leaf cluster of Qiaojia five-needle pine and the face of western black-crowned gibbon), and the irrelevant background (such as weeds and rocks) is suppressed. The output spatial weighted feature map F3 (320x320x64) is finally encoded with coordinate information, and the feature map is decomposed into horizontal and vertical position features. Through the embedding of target coordinate information (such as the x / y-axis position of the forest musk deer in the fog area), the model's perception of the target spatial position is strengthened, and the enhanced feature map F4 (320x320x64) is output.

[0082] Finally, the enhanced feature map F4 is processed by the Backbone subsequent module (subsequent C3k2, SPPF, C2PSA) to generate three-level feature maps: P3 (128x128x256): shallow features, retaining optimized details (such as the needle leaf texture of Taxus chinensis partially hidden by snow, and the edge contrast of the green peacock tail feather eye spot); P4 (64x64x512): medium-level features, balancing semantics and details (such as the body outline of the Yunnan snub-nosed monkey in the fir canopy and its habitat association); P5 (32x32x1024): deep-level features, strengthening high-level semantics (such as the matching relationship between species categories and habitats).

[0083] The adaptive environmental feature enhancement module can finally output high / medium / low environmental interference level labels, and the Backbone finally outputs P3 (shallow features, rich in details), P4 (medium-level features, balancing semantics and details), and P5 (deep-level features, rich in semantics) feature maps to the Neck part.

[0084] It should be noted that the dynamic learning mechanism in the embodiment is to judge the extreme environmental disturbance level where the rare and endangered species usually inhabit by analyzing the feature map in real time. When a high disturbance environment (such as vegetation background under direct sunlight at noon) is detected, the dynamic threshold suppression unit is automatically activated to filter background noise features (such as leaf reflection spots) exceeding the threshold; at the same time, the species features are strengthened, and the matching area in the feature map is gradient enhanced through the learned key feature template of the endangered species (such as the facial red spot of the Yunnan golden monkey and the eye-shaped spot of the green peacock), so as to highlight the morphological characteristics of the species. Unlike the traditional fixed enhancement method, the dynamic learning mechanism of this module enables the threshold and the strengthening weight to be dynamically updated with the input image, without the need for manual setting of environmental adaptation parameters.

[0085] The dynamic learning mechanism is based on dynamic convolution (DynamicConv) to extract multi-scale features of the input image, map them to an adaptive adjustment space, and construct a dynamic weight adjustment function in the space to adaptively adjust the convolution kernel weight. The calculation process is as follows: First, the texture complexity is calculated by extracting texture feature indexes such as energy, entropy, contrast, and correlation through a gray level co-occurrence matrix (GLCM); then the pixel value distribution is measured to determine the light intensity by using the image brightness channel; finally, the target proportion is calculated by counting the proportion of target area pixels based on the preliminary target detection results.

[0086] Then, the function is constructed and optimized based on the pre-constructed multi-layer perceptron (MLP) of the deep learning model. The multi-layer perceptron (MLP) includes an input layer, several hidden layers, and an output layer. The input layer receives three feature indexes: texture complexity, light intensity, and target proportion; the hidden layer uses a ReLU activation function to enhance the non-linear expression ability of the model; the output layer outputs the threshold θ and the strengthening weight α, which are used to adjust the judgment standard and weight distribution of the detection model. The specific content is as follows: Assume that the initial feature map is F0 Corresponding to texture complexity, light intensity, and target proportion, respectively, linear transformation is performed from the input layer to the hidden layer is the weight matrix of the i-th layer, which is responsible for weighted combination of the input feature F0; is the bias vector of the i-th layer, which is used to adjust the output after linear transformation to avoid the output value of the model always shifting towards zero, ensuring that the model can learn more complex feature mapping relationships), and then processed by the ReLU activation function to obtain the activation value Finally, the threshold θ and the strengthening weight α are obtained in the output layer, where , σ is the Sigmoid function, ​​is the weight matrix of the output layer, which is used to linearly transform the output of the previous layer , is the weight matrix of another group of output layers, is the output vector of the previous layer (usually the second-to-last layer), and the softplus function is used to ensure that the output is positive, and are the bias vectors in the corresponding operations, respectively, which are used to adjust the output after linear transformation, increasing the model fitting ability. The error between the actual detection result and the true label, such as cross-entropy loss, mean square error, etc., is taken as the optimization objective, and the back propagation algorithm is used to continuously adjust the weight parameters of the deep learning model, so that the threshold and reinforcement weight output by the constructed dynamic weight adjustment function can adapt to the image feature distribution in different environments.

[0087] Finally, the fluctuation range of the threshold θ and the reinforcement weight α in the historical data is recorded, and the update amplitude is limited according to the historical fluctuation range every time the weight is updated, to avoid overfitting and ensure the detection performance of the algorithm in different scenarios.

[0088] In this embodiment, the adaptive environmental feature enhancement module realizes dynamic filtering of environmental noise and directional reinforcement of key species identification features, effectively improving the feature robustness and expression stability of the model in complex imaging environments, aiming at the special habitat conditions of high altitude, strong light, and multiple fog in Yunnan region.

[0089] II) Cross-level feature interaction module: This module completely replaces the original FPN-PAN feature fusion network structure of the Neck, and the architecture of the Neck network is as shown in the accompanying Figure 4 .

[0090] This module constructs three layers of feature interaction channels in the YOLOv11 algorithm architecture, which are high-level semantic features, middle-level structure features, and low-level details. It solves the core problem that rare and endangered species are difficult to be effectively identified and detected by monitoring devices due to their sparse distribution in natural environment, hidden activities, and similar morphological features to the environment.

[0091] High-level semantic features (containing species category information) are guided by attention to focus on potential habitats, such as the activity area of Yunnan golden monkeys in the canopy of fir forests, and guide the middle-level structure features to focus on this area; middle-level structure features, such as the body outline of forest musk deer, are supplemented by details to focus on the low-level features, such as the musk gland area of forest musk deer, to avoid the dispersion of detailed features in complex backgrounds; low-level detailed features, such as the needle leaf texture of Taxus chinensis, are amplified by difference to strengthen the edge features that are partially blocked by snow or fallen leaves, while the subtle edge changes are fed back to the high-level to enhance the boundary prediction ability of the blocked area.

[0092] The three complement each other through cyclic interaction to ensure that the weak characteristics of small population species are not overwhelmed by complex backgrounds, and the complementarity of different scale features is more strengthened than traditional structures, solving the problem of insufficient transmission of weak features in the original Neck. The detailed process is as follows: 1) First, the attention-guided flow of high-level semantic features (including species category information) takes P5 output by Backbone as the core, focusing on potential habitats. P5 (deep semantic features) contains species category information (such as Yunnan golden monkeys) and habitat-related semantics (such as fir forest canopy). First, 1x1 convolution is used to compress the channel to 256 dimensions to extract core semantic features. Then, the habitat attention mechanism is introduced to calculate the matching degree between species and habitat in P5, such as the semantic similarity between Yunnan golden monkeys and fir forest canopy, to generate a habitat attention map (32x32) that focuses on high-matching areas (such as the dense area of branches and twigs of fir forest canopy).

[0093] Specifically, first, extract the species category embedding vector (such as the semantic encoding of Yunnan golden monkeys) and habitat scene feature vector (such as the texture and structure features of fir forest canopy) from P5 deep semantic features. Then, calculate the matching degree between the two by cosine similarity to get the correlation score between species and habitat at each spatial position. Finally, normalize the score through Sigmoid activation to generate a 32x32 habitat attention map, where high-score areas (such as the branch and twig area of fir forest canopy) have their weights strengthened as the focus area for subsequent feature fusion.

[0094] Next, P5 feature map and habitat attention map are 2x up-sampled (to 64x64) and fused with P4 (mid-level features) across scales. Through dynamic weight coefficients (Sigmoid activation learning), P5 habitat attention weights are assigned to P4 to guide P4 to focus on potential activity areas (such as the branch and twig gaps of Yunnan golden monkeys in fir forest canopy), and suppress irrelevant backgrounds (such as ground litter), outputting the fused P4' (64x64).

[0095] 2) The detail supplement flow of mid-level structural features takes the fused P4' (mid-level structural features) as the core, focusing on iconic parts. P4' contains species body contour information (such as the trunk shape of forest musk deer). After extracting structural features through 3x3 convolution, introduce part attention mechanism based on iconic part templates (such as musk gland area of forest musk deer, tail feather eye spot of green peacock) from species specimen library to generate part attention map (64x64) to highlight the identification areas that need to be focused on.

[0096] Subsequently, the P4' feature map and the part attention map are upsampled by 2 times (to 128x128) and fused with P3 (shallow detail features): the structural constraints of P4' are transmitted to P3 through residual connection, forcing the detail features of P3 (such as hair texture, leaf vein) to focus on the landmark parts (such as the edge of the musk gland of forest musk deer, the needle arrangement texture of Taxus chinensis), avoiding the details to be dispersed in the complex background (such as the entanglement of vines, the coverage of fallen leaves), and outputting the optimized P3' (128x128).

[0097] 3) The difference amplification flow of low-level detail features (such as the needle texture of Taxus chinensis) takes P3' (low-level detail features) as the core, strengthens the edge features of the partially occluded targets (such as the half-covered Taxus chinensis needle edge by snow, the finger outline of the western black-crowned long-armed gibbon occluded by vines), and realizes feedback. Starting from P3', it contains the edge details of the occluded targets (such as the half-covered Taxus chinensis needle edge by snow, the finger outline of the western black-crowned long-armed gibbon occluded by vines), through edge difference amplification (combined with Laplacian operator and dynamic activation function), the gradient of the fuzzy edge area (such as the gray transition zone between snow and needle) is enhanced, the subtle edge changes are amplified, and the enhanced edge feature map (128x128) is generated.

[0098] Specifically, first, the Laplacian operator is used to convolve the low-level detail feature map of P3' to extract edge gradient information (such as the gray transition zone gradient between snow and Taxus chinensis needle); then the gradient value is used to dynamically adjust the activation function (such as using LeakyReLU to enhance the response in the low-gradient fuzzy area, and keeping the ReLU output in the high-gradient clear area); finally, the edge gradient difference is amplified pixel by pixel to make the gray change of the fuzzy edge more significant, and the edge recognition of the occluded target (such as the needle covered by snow) is strengthened.

[0099] Next, P3' and the enhanced edge feature map are downsampled by 0.5 times (to 64x64), and fused with P4' twice to output P4'' (64x64). This step supplements the enhanced edge details to the middle-level features, improving the structural integrity of the occluded area; further downsample P4'' to 32x32, and fuse with P5 to output P5' (32x32), so that the low-level edge difference information is fed back to the high-level semantic features, enhancing the boundary prediction ability of the high-level for partially occluded targets (such as completing the branch boundary of Taxus chinensis occluded by snow according to the edge trend of the exposed needle).

[0100] Finally, the output enhanced third-order feature maps P3' (128x128, focusing on occlusion details), P4'' (64x64, balancing structure and semantics), and P5' (32x32, strengthening habitat association and boundary prediction) provide multi-scale feature support for the detection task of the Head. That is, the cross-level feature interaction module can output the target occlusion level (judge the degree of target occlusion by vegetation, rock, etc., such as "heavy occlusion" corresponds to <30% feature visibility), and the enhanced feature maps P3', P4'', and P5'.

[0101] In this embodiment, the cross-level feature interaction module replaces the original one-way or simple two-way FPN-PAN structure by constructing a three-order circular interaction path of high-level semantics-middle-level structure-low-level details, significantly enhancing the multi-level feature fusion and information transmission capability of severely occluded targets (such as long-armed apes in the state of vine shading) and extremely small-scale targets (such as dwarf monkey infants), and alleviating the feature attenuation problem caused by occlusion or small target size.

[0102] III) Fine-grained discriminant feature module This module is embedded in front of the input end of the final classification convolution layer of the classification submodule of the Head. The detection head Head architecture is shown in FIG. 8. Figure 5

[0103] This module is divided into two layers in the YOLOv11 algorithm architecture. One is the feature library association layer, which loads the fine-grained feature library of the species specimen library (such as the petal vein pattern template of the Fagara fortunei), and realizes feature matching through dynamic convolution. The other is the attention focusing layer, which focuses on the key discriminant points (such as the facial erythema position of the Yunnan snub-nosed monkey) of the closely related species through the multi-head self-attention mechanism (Multi-Head Self-Attention), and binds the classification result with the protection level label.

[0104] This module is designed for closely related species in the "List of Rare and Endangered Plants in Yunnan" and "List of National Key Protected Wild Animals".

[0105] Firstly, the feature library association layer decomposes the classification features into common features (such as the needle-shaped leaves of Pinaceae plants) and discriminant features (such as the five-needle-in-a-bundle feature of Pinus fenzeliana) through the feature decoupling unit.

[0106] ​Subsequently, comparative learning is performed through the attention focusing layer, access is made to the endangered species specimen feature library provided by the Forestry and Grassland Bureau of Yunnan Province, the difference degree of the current features and the close relative species (such as the petal vein order difference between Keteleeria evelyniana and Magnolia) is calculated, and the weight of the unique identification features is strengthened; finally, the classification vector fused with the commonness and the identification features after strengthening is output, and the classifier is input to complete the fine-grained identification. Compared with the traditional architecture, the design realizes accurate differentiation and correlation of the protection level information of close relative species, and finally can significantly improve the differentiation accuracy of the classification decision of close relative species, and the detailed process is as follows: 1) Feature library correlation layer: the enhanced feature maps P3', P4'' and P5' output by Neck are subjected to scale unification processing, P3' is down-sampled by 0.5 times to 64x64, P5' is up-sampled by 2 times to 64x64, and P4'' (64x64) forms a feature set of the same size, and then 1x1 convolution is performed to compress the channel number of the three to 128 dimensions, and then the fused feature map (64x64x384) is obtained.

[0107] Then, the features are separated through the feature decoupling unit (containing 2 parallel convolution branches), the 3x3 convolution is adopted to extract the cross-species common features (such as the needle-shaped leaf shape of Pinaceae plants and the stipule scar structure of Magnoliaceae plants), and the common feature component (64x64x64) is output; at the same time, the spatial attention mechanism is combined to focus on the detailed features (such as the five-needle bundle texture of Keteleeria evelyniana and the radial vein order of Keteleeria evelyniana petals) transmitted by P3', and the species-specific identification features are extracted through 1x1 convolution and global average pooling, and the identification feature component (64x64x64) is output.

[0108] 2) Attention focusing layer: the layer accesses the endangered species specimen feature library provided by the Forestry and Grassland Bureau of Yunnan Province, calls the benchmark features of close relative species (such as the needle leaf bundle number template of Keteleeria evelyniana vs. Pinus yunnanensis and the petal vein order template of Keteleeria evelyniana vs. Magnolia), forms a benchmark feature matrix (the dimension matches the identification feature component), and at the same time, the difference degree of the current features and the close relative species (such as the texture difference degree of five-needle bundle of Keteleeria evelyniana and three-needle bundle of Pinus yunnanensis and the structure difference degree of radial vein order of Keteleeria evelyniana and net vein order of Magnolia) is calculated through cosine similarity, the decoupled identification feature component is compared with the benchmark feature matrix element by element, and finally the difference heat map (64x64) is generated.

[0109] Based on the difference heat map, the identification attention weight (the weight of the high difference area tends to 1, and the weight of the low difference area tends to 0) is generated, which is multiplied point by point with the identification feature component, the response strength of the unique identification features (such as the needle leaf bundle node of Keteleeria evelyniana and the vein bifurcation point of Keteleeria evelyniana) is strengthened, and the common interference is suppressed.

[0110] Finally, the enhanced discriminant feature components and the common feature components are fused by residual connection (retaining the species basic category information while highlighting the discriminant details), and then the fused feature maps are compressed into 128-dimensional classification vectors through global average pooling, including common feature components (such as the basic category information of Pinaceae and Magnoliaceae) and enhanced discriminant feature components (such as the five-needle bundle of Pinus mairei and the radial vein order of Distimake rostratus), and finally the classification vector is input into the classifier of the Head to realize the accurate differentiation of closely related species (such as the differentiation of Pinus mairei and Pinus yunnanensis, and Distimake rostratus and Magnoliaceae plants), and to improve the fine-grained recognition accuracy.

[0111] The fine-grained discriminant feature mining module can finally output the species category (including fine-grained differentiation) and associate the preset protection level label.

[0112] In the embodiment, the fine-grained discriminant module introduces an identification mechanism based on the prior knowledge of the specimen library, strengthens the modeling of the subtle differences of morphologically similar closely related species (such as Pinus mairei and Pinus yunnanensis), and thus highlights the discriminant local features on the basis of common features, significantly improving the inter-class differentiation ability and species classification accuracy.

[0113] In summary, the embodiment constructs a full-chain technology closed loop from model optimization, field monitoring to continuous evolution through phased training, lightweight deployment design and incremental learning mechanism, effectively ensuring the practicality and adaptability of the system in the complex application scenarios of nature reserves.

[0114] The present application focuses on the Yunnan unique rare and endangered species and their typical living environment, covering a variety of critically endangered and endangered species listed in the "List of Rare and Endangered Plants in Yunnan" and "List of National Key Protected Wild Animals", such as the Yunnan golden monkey, the green peacock, Distimake rostratus, Pinus mairei, etc. By improving the detection rate and fine-grained recognition accuracy of these species in complex habitats, the dynamic monitoring of their population dynamics, activity trajectory, population size and behavior state can be realized, and individual abnormal state (such as injury, illness) and habitat disturbance events can be effectively identified and warned. The technical solution formed by the embodiment can provide accurate protection decision basis for nature reserves, thereby playing a key role in the survival and recovery of small population species and the overall protection of regional biodiversity, and maintaining the integrity of biodiversity in Yunnan.

[0115] Step S3: Based on the improved YOLOv11 algorithm trained in step S2, select the measured data of typical protected areas such as Baima Snow Mountain and Yuanjiang River Basin to construct a test set and verify the algorithm effect. The test set needs to cover the complex habitats unique to Yunnan, such as high-altitude bare rock (habitat of Yunnan snub-nosed monkey), rainforest fog (habitat of western black-crowned gibbons), and river valley shrubs (feeding area of green peafowl), and contains samples of different interference scenarios: such as snow mountain goats in strong light, clouded leopards in fog, and lazy monkeys blocked by vines, as well as areas of five-needle pine and Yunnan pine mixed vegetation in Qiaojia, to ensure that the verification scenarios are consistent with the actual monitoring needs.

[0116] The verification process focuses on three core capability evaluations. First, the adaptability to complex environments is evaluated by comparing the detection accuracy of the model in high-altitude strong light and fog scenarios (such as the detection rate of Yunnan snub-nosed monkeys in Baima Snow Mountain) with that in regular environments, to assess the noise reduction effect of the adaptive environment feature enhancement module. Second, the detection of obscured targets is evaluated by counting the number of individuals obscured by branches, leaves, and rocks (such as black-necked cranes obscured by shrubs in Yuanjiang River Basin), to verify the ability of the cross-level feature interaction module to capture incomplete features. Third, the fine-grained differentiation is evaluated by calculating the classification accuracy of closely related species (such as Huagai wood and Magnolia plants, and clouded leopards and musk deer), to test the differentiation effect of the fine-grained identification module.

[0117] During the evaluation process, key indicators are recorded simultaneously. The overall detection accuracy needs to be stable at 85% or above, the detection rate of critically endangered species (such as green peafowl) needs to be no less than 90%, and the classification confusion rate of closely related species needs to be controlled within 5%. If the indicators of a certain scenario do not meet the threshold (such as the detection rate of clouded leopards in fog areas being lower than the threshold), the parameter settings of the corresponding module (such as the filtering strength of the dynamic convolution kernel) are analyzed, and typical difficult cases (such as feature blurring samples caused by heavy fog) in the test set are fine-tuned to form an optimized model that adapts to the actual monitoring needs of Yunnan protected areas.

[0118] Step S4: Based on the algorithm verified in step S3, perform lightweight processing to adapt to the edge devices of protected areas. Remove redundant convolution layers and quantize weight parameters (such as converting 32-bit floating-point numbers to 16-bit) to reduce the model size and computational complexity, while ensuring that the loss of detection accuracy is controlled within 5%, so that it can run stably on devices such as Jetson edge terminals and infrared camera embedded modules, meeting the real-time processing needs in offline environments.

[0119] After deployment, the edge device can perform localized analysis on the collected real-time images, quickly identify endangered species such as Yunnan snub-nosed monkey and green peacock, and automatically trigger abnormal behavior warning when it detects slow movement and abnormal body of forest musk deer, and generate real-time data containing species information, location coordinates and warning type. These data are transmitted wirelessly (such as Beidou short message) to the reserve management platform, which integrates multiple device monitoring results to form visual information such as species distribution heat map and activity trajectory tracking, providing accurate action guidance for patrol personnel and realizing closed-loop management from real-time detection to protection response.

[0120] Step S5: Based on the real-time monitoring in step S4, a dynamic updating mechanism is established to adapt to the natural changes of species, i.e. periodically collecting field data recorded by reserve patrol personnel, including newly discovered small population individuals (such as Yunnan blue fruit tree seedlings), seasonal changes of species behavior (such as courtship behavior of green peacock during breeding season), and new identification characteristics of closely related species (such as needle leaf shape difference of Taxus chinensis at different growth stages), and organizing these data into incremental learning samples.

[0121] With the help of incremental learning algorithm, the model is fine-tuned with new samples to update the feature templates of species in the fine-grained feature library, so that the model can identify new individuals appearing after the change of population number, or adapt to the dynamic change of behavior mode (such as new activity track of Yunnan snub-nosed monkey due to change of food distribution), and through continuous iteration, the monitoring system is always synchronized with the actual state of the protected object, finally ensuring that the protection measures for endangered species can be adjusted in time with the population dynamics.

[0122] It should be noted that the description of the device of the embodiments of the present application is similar to the description of the above-mentioned method embodiments, and has similar beneficial effects as the method embodiments, therefore, no further description is given. For technical details not disclosed in the present device embodiment, please refer to the description of the method embodiments of the present application for understanding.

[0123] It should be noted that, in the embodiments of the present application, if the improved YOLOv11-based Yunnan rare and endangered species monitoring method described above is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. According to such an understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing a terminal to perform all or part of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read Only Memory), a magnetic disk or an optical disk, and various media that can store program codes. Thus, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0124] As an example, the executable instructions can but need not correspond to files in a file system, can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, subprograms, or portions of code. The executable instructions can be deployed to be executed on one electronic device or alternatively on multiple electronic devices that are located at one site or distributed across multiple sites and interconnected by a communication network.

[0125] The above merely describes the embodiments of the present application, but is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present application shall be included in the protection scope of the present application.

[0126] It should be understood that the "one embodiment" or "an embodiment" mentioned throughout the specification means that the specific features, structures, or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily mean the same embodiment. In addition, these specific features, structures, or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that the size of the sequence number of each process in various embodiments of the present application does not mean the execution order, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application. The sequence number of the above-mentioned embodiments of the present application is only for description, not representing the advantages and disadvantages of the embodiments.

[0127] It should be noted that, in the present document, the terms "comprising", "containing" or any other similar term are intended to encompass non-exclusive inclusions, such that a process, a method, or an apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element. In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The above-described device embodiments are merely illustrative, for example, the division of the units is only a logical functional division, and in actual implementation, there can be another division, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.

[0128] The above description is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A Yunnan rare and endangered species monitoring method based on improved YOLOv11, characterized by, The method comprises: acquiring an image data set containing rare and endangered species in Yunnan in special habitats; inputting the image data set into a trained species identification model for feature extraction and endangered species identification, and outputting an identification result; the identification result at least includes species category, location information and target protection level information of the endangered species; wherein the species identification model is obtained by embedding a self-adaptive environment feature enhancement module at the output end of the first C3k2 module in the backbone network of the YOLOv11 model, replacing the FPN-PAN feature fusion network structure in the neck network with a cross-level feature interaction module, adding a fine-grained identification module before the classification convolution layer of the detection head, and performing stage-by-stage training; the self-adaptive environment feature enhancement module is used to process environmental noise in the image data set through a dynamic filtering subnetwork, and to strengthen key species identification features; the cross-level feature interaction module is used to perform feature enhancement processing on the multi-level feature maps output by the backbone network based on the constructed three-layer feature interaction channel of high-level semantics-middle-level structure-low-level details; the fine-grained identification module is used to perform feature fusion on the enhanced multi-scale feature maps to obtain a classification vector including common feature components and enhanced identification feature components.

2. The method of claim 1, wherein, The self-adaptive environment feature enhancement module includes an environment noise suppression layer and a feature enhancement layer; the self-adaptive environment feature enhancement module is specifically used to perform the following operations: The environment noise suppression layer is realized based on dynamic convolution, and is used to perform region analysis on the initial feature map output by the first C3k2 module, apply a convolution kernel matched with the current environmental interference type to different regions in the initial feature map, and obtain a denoised feature map; The feature enhancement layer includes a channel attention module, a spatial attention module and a coordinate attention mechanism, and is used to sequentially perform channel attention weighting, spatial attention weighting and coordinate information coding on the denoised feature map to obtain an enhanced feature map.

3. The method of claim 2, wherein, The environment noise suppression layer is specifically used to perform the following operations: By calculating the brightness pixel value of the initial feature map and the preset brightness pixel threshold, the overexposed region is located, and by calculating the Laplacian variance of the initial feature map and the preset variance threshold, the fog area scattering region is located; Through a dynamic learning mechanism, the initial feature map is analyzed in real time, and a convolution kernel weight parameter matched with each region is generated and adaptively adjusted; wherein, for the overexposed region, a low-pass filtering kernel with a size of 7x7 is used for mean filtering; for the fog area scattering region, a anisotropic filtering kernel with a size of 3x3 is used for Gaussian filtering; for the normal region, a standard convolution kernel with a size of 3x3 is maintained; The convolution kernel weight parameter is applied to the initial feature map, and the denoised feature map is output through region-by-region weighted convolution.

4. The method of claim 2, wherein, The feature enhancement layer is specifically used to perform the following operations: Through the channel attention module, the channel importance of the denoised feature map is evaluated, the global average pooling of each channel is calculated, and the weight of each channel is learned through a fully connected layer to generate a channel weighted feature map; The spatial correlation feature is extracted by a first convolution kernel, spatial weight distribution is performed on the channel weighted feature map, a spatial attention map is generated, weight focus is generated on the target region, irrelevant background is suppressed, and a spatial weighted feature map is output; Coordinate information coding is performed on the spatial weighted feature map, the spatial weighted feature map is decomposed into horizontal and vertical direction position features, target coordinate information is embedded, perception of the target spatial position is strengthened, and the enhanced feature map is output.

5. The method of claim 1, wherein, The multi-level feature map includes a shallow feature map, a middle feature map and a deep feature map; the cross-level feature interaction module is specifically configured to perform the following operations: Species category embedding vectors and habitat scene feature vectors are extracted from the deep feature map by compressing channels through a second convolution kernel; a habitat attention map is generated by calculating the matching degree between species and habitats; After 2 times upsampling of the deep feature map and the habitat attention map, cross-scale fusion is performed on the middle feature map, habitat semantic weight is given to the middle feature map through a dynamic weight coefficient, and a first fused middle feature map is output; After extracting the structural features of the first fused middle feature map through a third convolution kernel, a part attention mechanism is introduced, the structural features are processed based on the pre-stored landmark part template of the species specimen library, and a part attention map is generated; After 2 times upsampling of the part attention map and the first fused middle feature map, fusion is performed on the shallow feature map, and the structural constraint of the first fused middle feature map is transmitted to the shallow feature map through residual connection, and an optimized shallow feature map is output; After convolution operation is performed on the optimized shallow feature map by combining Laplacian, gradient enhancement is performed on the extracted fuzzy edge region according to the gradient value of the dynamic activation function, and the edge gradient anomaly is amplified through pixel-by-pixel weighting, and an enhanced edge feature map is generated; After 0.5 times downsampling of the optimized shallow feature map and the enhanced edge feature map, secondary fusion is performed on the first fused middle feature map, and a second fused middle feature map balancing structure and semantics is output; The second fused middle feature map is down-sampled and fused with the deep feature map, and a fused deep feature map with enhanced habitat correlation and boundary prediction ability is output.

6. The method of claim 1, wherein, The fine-grained identification module includes a feature library association layer and an attention focusing layer; the fine-grained identification module is specifically configured to perform the following operations: The common features and identification features of species are separated from the input optimized shallow feature map, the second fused middle feature map and the fused deep feature map through the feature library association layer, and the identification features are matched with the reference features of the close species in the pre-set endangered species specimen feature library, and a difference heat map is generated; The identification attention weight is generated based on the difference heat map through the attention focusing layer, the identification attention weight is strengthened, the strengthened identification features are fused with the common features, and a classification vector for fine-grained species classification is formed.

7. The method of claim 6, wherein, The feature library association layer is specifically further configured to perform the following operations: Scale alignment and channel compression are performed on the optimized shallow feature map, the second fused middle-level feature map and the fused deep-level feature map to splice a target fused feature map; The target fused feature map is input into a feature decoupling unit in the feature library association layer, and feature separation is performed through two parallel convolution branches; the first convolution branch adopts 3×3 convolution to extract cross-species common characteristics, and outputs the common characteristic component; The second convolution branch combines a spatial attention mechanism, adopts 1×1 convolution and global average pooling to extract species-specific discriminative feature components.

8. The method according to any of claims 6 or 7, characterized in that, The attention focusing layer is specifically further used to perform the following operations: A preset endangered species specimen feature library is called to obtain a reference feature template of a close relative species, and a reference feature matrix is formed; Cosine similarity between the discriminative feature components and the reference feature matrix is calculated, and a difference heat map is generated; The difference heat map is normalized into discriminative attention weights, and the discriminative attention weights are multiplied point by point with the discriminative feature components to obtain reinforced discriminative feature components; The reinforced discriminative feature components and the common characteristic vector are fused through residual connection to form a fused feature; Global average pooling is performed on the fused feature to compress and generate the classification vector.

9. The method of claim 1, wherein, After the image data set is input into the trained species recognition model for feature extraction and endangered species recognition, and an identification result is output, the method further comprises: The trained species recognition model is subjected to lightweight processing and is deployed to an edge computing device; Real-time monitoring images are analyzed locally by the edge computing device, and when an endangered species or an abnormal behavior of the endangered species is identified, a warning information is automatically generated and sent.

10. The method of claim 6, wherein, The species recognition model is trained through the following process: Initialize the algorithm with a sample data set in a general scenario, freeze the parameters of the adaptive environmental feature enhancement module, the cross-level feature interaction module and the fine-grained discriminative module, and train a YOLOv11 basic architecture; Obtain samples containing environmental interference, and label species categories, bounding boxes, species key discriminative features and species corresponding protection levels to construct a training data set; Unfreeze all parameters, input the training data set into the trained YOLOv11 basic architecture for training, and dynamically adjust the weights of each module; Intensive training is performed on samples of close relative species and critically endangered species to obtain the trained species recognition model.

Images