Cloud-based intelligent biodiversity monitoring methods and systems

Abstract

This application discloses a cloud-based intelligent biodiversity monitoring method and system, relating to the field of intelligent monitoring. Specifically, species images collected by front-end devices are first uploaded to the cloud platform, where a general recognition model performs a preliminary assessment. If the general model's recognition result has high confidence or indicates a non-confusing species, the result is directly output. Crucially, when the general model's recognition result is uncertain or involves predefined easily confused species, the system extracts the species' Region of Interest (ROI) image and feeds it into a specially trained expert attention classifier. This classifier extracts multi-scale spatial salient features and performs deep fusion to accurately capture subtle differences, ultimately providing expert-level prediction results. This recognition strategy effectively overcomes the challenges of species identification in complex environments, significantly improving the accuracy and reliability of biodiversity monitoring.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring, and more specifically, to a cloud-based intelligent biodiversity monitoring method and system. Background Technology

[0002] With the intensification of global climate change and human activities, biodiversity faces unprecedented threats. To effectively protect and manage biological resources, long-term, continuous, and intelligent monitoring of biodiversity is crucial. Traditional biodiversity monitoring methods often rely on manual field surveys, which are not only time-consuming and labor-intensive but also limited by human and material resources, making it difficult to achieve large-scale, high-frequency monitoring. Therefore, constructing a cloud-based intelligent biodiversity monitoring solution, utilizing advanced information technology and artificial intelligence to achieve automated species identification, efficient data management, and analysis, is a key issue that urgently needs to be addressed in the field of biodiversity conservation.

[0003] However, while some research has attempted to apply artificial intelligence to species identification, there are few mature cloud-based intelligent biodiversity monitoring solutions that can effectively address the challenges of species identification in complex scenarios. This is mainly because biodiversity monitoring faces numerous challenges. For example, in the wild, collected species images often suffer from poor image quality due to low light, complex backgrounds, species occlusion, and varying shooting angles. More importantly, many species exhibit small inter-class differences but large intra-class differences. Different species may be extremely similar in morphology, while the same species may show significant differences under different conditions, such as posture, age, individual differences, lighting, occlusion, and distance. For instance, hog badgers and Asian badgers are extremely similar in size, coat color distribution, and nocturnal habits. Their distinguishing features are very subtle and localized; for example, the front of a hog badger's snout is hairless, while an Asian badger's snout is hairy. Under low light, low resolution, or specific shooting angles, these key features are easily lost or become blurred, making it difficult for general identification models to accurately distinguish them.

[0004] Therefore, a new solution is needed to effectively address the aforementioned technical problems and improve the accuracy and efficiency of intelligent biodiversity monitoring. Summary of the Invention

[0005] To overcome the problems in the prior art, according to one aspect of this application, a cloud platform-based intelligent biodiversity monitoring method is provided, comprising: the cloud platform receiving original images of species objects collected by front-end devices; inputting the original images of species objects into a general recognition model to obtain a general prediction result, the general prediction result including a bounding box, a predicted species, and a confidence score; in response to the general prediction result satisfying a preset condition, extracting a species object ROI image from the original images of species objects based on the bounding box; in response to the general prediction result not satisfying the preset condition, displaying the original images of species objects and the general prediction result; inputting the species object ROI image into an expert attention classifier to obtain an expert prediction result; and displaying the expert prediction result and the original images of species objects.

[0006] According to another aspect of this application, a cloud-based intelligent biodiversity monitoring system is provided, comprising: a species object original image acquisition module, used for receiving original images of species objects collected by a front-end device via a cloud platform; a general prediction module, used for inputting the original images of species objects into a general recognition model to obtain a general prediction result, the general prediction result including a bounding box, a predicted species, and a confidence score; a species object ROI image cropping module, used for cropping a species object ROI image from the original images of species objects based on the bounding box in response to the general prediction result meeting a preset condition; a general prediction result comparison module, used for displaying the original images of species objects and the general prediction result in response to the general prediction result not meeting the preset condition; an expert prediction module, used for inputting the species object ROI image into an expert attention classifier to obtain an expert prediction result; and a display module, used for displaying the expert prediction result and the original images of species objects.

[0007] Compared with existing technologies, this application provides a cloud-based intelligent biodiversity monitoring method and system. Specifically, species images collected by front-end devices are first uploaded to the cloud platform, where a general recognition model makes a preliminary judgment. If the general model's recognition result has high confidence or does not involve easily confused species, the result is directly output. Crucially, when the general model's recognition result is uncertain or involves predefined easily confused species, the system extracts the species' ROI image and feeds it into a specially trained expert attention classifier. This classifier extracts multi-scale spatial salient features of the species, such as edges, textures, and shapes, and performs deep fusion to accurately capture subtle differences, ultimately providing expert-level prediction results. This recognition strategy fully leverages the efficiency of the general model and the accuracy of the expert model, effectively overcoming the challenges of species identification in complex environments and significantly improving the accuracy and reliability of biodiversity monitoring. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of a cloud-based intelligent biodiversity monitoring method according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of data flow in a cloud-based intelligent biodiversity monitoring method according to an embodiment of this application.

[0011] Figure 3 This is a schematic diagram of the data flow in step S3 of the cloud-based intelligent biodiversity monitoring method according to an embodiment of this application.

[0012] Figure 4 This is a flowchart of step S5 in the cloud-based intelligent biodiversity monitoring method according to an embodiment of this application.

[0013] Figure 5 This is a block diagram of a cloud-based intelligent biodiversity monitoring system according to an embodiment of this application. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] To address the limitations of existing technologies, this application proposes a cloud-based intelligent biodiversity monitoring method. Figure 1 This is a flowchart of a cloud-based intelligent biodiversity monitoring method according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating the data flow of a cloud-based intelligent biodiversity monitoring method according to an embodiment of this application. Figure 1 and Figure 2As shown, the cloud-based intelligent biodiversity monitoring method according to an embodiment of this application includes: S1, the cloud platform receives original images of species objects collected by a front-end device; S2, the original images of species objects are input into a general recognition model to obtain a general prediction result, the general prediction result including a bounding box, a predicted species, and a confidence score; S3, in response to the general prediction result meeting a preset condition, a species object ROI image is extracted from the original images of species objects based on the bounding box; S4, in response to the general prediction result not meeting the preset condition, the original images of species objects and the general prediction result are displayed; S5, the species object ROI image is input into an expert attention classifier to obtain an expert prediction result; S6, the expert prediction result and the original images of species objects are displayed.

[0016] In step S1, the cloud platform receives the original images of species objects collected by the front-end devices. It should be understood that in the field of intelligent biodiversity monitoring, front-end devices, such as infrared cameras and sensors, continuously collect a large amount of species image data in the field environment. This data is the foundation for species identification and biodiversity analysis. However, due to the complexity and diversity of the field environment and the vast monitoring range, front-end devices are usually distributed in different geographical locations and may face problems such as unstable network connections and limited storage space. To achieve centralized management, efficient processing, and remote access to this dispersed and massive amount of data, and to provide a unified entry point for subsequent intelligent identification and analysis, this application transmits the original images of species objects collected by the front-end devices to the cloud platform for centralized processing. This not only solves the problem of insufficient local processing capabilities of the front-end devices but also provides powerful computing power support for subsequent complex calculations and model inference, and ensures the secure storage and sharing of data.

[0017] For example, one feasible approach for step S1 is as follows: First, a front-end device, such as an infrared camera deployed in a protected area, captures an image of an animal at night. This infrared camera has a built-in image acquisition module and a wireless communication module. When the infrared camera senses animal activity and triggers a capture, the image acquisition module generates raw image data containing the species. Subsequently, this raw image data is encrypted and packaged into data packets via the infrared camera's built-in wireless communication module, for example, through a 4G / 5G cellular network or satellite communication network. These data packets are transmitted in real-time or periodically to a pre-set cloud platform server. The cloud platform server has a data receiving service deployed on it, which continuously listens for connection requests and data transmissions from the front-end device. Once a data packet is received from the infrared camera, the data receiving service unpacks and decrypts it, and stores the raw image data in the cloud platform's data storage module, such as an object storage service. During storage, a unique identifier can be generated for each image, and metadata information such as its acquisition time and geographical location can be recorded for subsequent retrieval and management. For example, an image of a hog badger taken at night in a protected area, along with its raw image data, shooting timestamp, and GPS coordinates, will be packaged and uploaded to a cloud platform via a 4G network. After receiving the data, the cloud platform's data receiving service will store it in a designated data bucket, generate a unique filename, and record its metadata.

[0018] In step S2, the original image of the species is input into a general recognition model to obtain a general prediction result, which includes a bounding box, predicted species, and confidence score. Correspondingly, in intelligent biodiversity monitoring, the amount of original image data collected by front-end devices is enormous and diverse, containing a large amount of non-target information, such as aerial shots, background interference, and images with insufficient clarity. Directly refining all original images would result in a huge waste of computational resources and low processing efficiency. Therefore, in order to quickly filter out images containing potential species targets and with relatively usable quality, and to initially identify the species and their approximate locations in the images, thereby providing a focused area for subsequent refined identification, this application requires inputting the original image of the species into a general recognition model to obtain a general prediction result.

[0019] Specifically, in an exemplary embodiment of this application, the general recognition model is the standard YOLOv5 model. It is worth noting that standard YOLOv5, as a classic single-stage object detection model, is characterized by its small size, fast inference speed, and high accuracy. It can quickly and accurately achieve multi-target detection, and is particularly suitable for edge devices with limited computing and storage resources, which highly aligns with the heavy data transmission load and high timeliness requirements in biodiversity monitoring in the field. Furthermore, YOLOv5 performs excellently in general object recognition, effectively handling complex backgrounds and lighting variations in images, laying the foundation for subsequent fine-grained recognition.

[0020] One feasible approach for step S2 is as follows: First, obtain the original image of the species from the data storage module of the cloud platform. For example, an original image of a badger captured by a front-end infrared camera and uploaded to the cloud may have an original resolution of 1920x1080 pixels. This original image will first undergo preprocessing, including data augmentation and adaptive image scaling. In the data augmentation stage, techniques such as random cropping, flipping, and color dithering can be applied to increase the diversity of training data and improve the model's generalization ability. Adaptive image scaling will adjust the image size to the standard input size set during YOLOv5 model training, such as 640x640 pixels, and perform normalization processing (e.g., scaling the pixel values ​​from the 0-255 range to the 0-1 range) to adapt to the model's input requirements.

[0021] Next, the preprocessed original image is input into the general recognition model, namely the standard YOLOv5 model. The YOLOv5 model is generally divided into an input end, a backbone network, a neck end, and an output end. The input end is responsible for data augmentation, adaptive image scaling, and adaptive anchor box calculation. Adaptive anchor box calculation refers to the fact that during the training phase, YOLOv5 uses methods such as K-means clustering to automatically calculate a set of preset anchor box sizes and aspect ratios that are most suitable for the current dataset based on the true size distribution of target objects in the training dataset, in order to improve the model's detection efficiency and accuracy for targets of different sizes. The backbone network adopts the CSPNet (C3) structure, and its role is to perform feature extraction. Through multiple convolutional operations, such as convolutional layers, batch normalization layers, activation functions, and pooling layers, it extracts feature maps of different scales from the image. For example, after a series of convolution and pooling operations, the image is gradually downsampled to obtain feature maps of different sizes such as 128x128, 64x64, and 32x32. These feature maps contain abstract features such as the image's edges, textures, and shapes. The weights and bias parameters in these convolutional layers are obtained through backpropagation and optimization on a large-scale dataset.

[0022] Subsequently, these feature maps are fed into the Neck layer, where feature fusion is performed using an FPN+PAN structure. FPN propagates strong semantic features from top to bottom, fusing high-level (semantically rich but low-resolution) feature maps with low-level (less semantically rich but high-resolution) feature maps through upsampling to enhance the semantic information of the low-level features. PAN, on the other hand, propagates strong localization features from bottom to top, fusing low-level feature maps with high-level feature maps through downsampling to enhance the localization accuracy of the high-level features. Through this combined operation, features extracted from different backbone layers are parametrically fused to generate fused multi-scale feature maps. These feature maps simultaneously contain rich semantic information and precise localization information to better capture targets of different sizes.

[0023] Finally, the fused feature maps are fed into the output. The output task is to map the three feature maps output by Neck to predefined anchor boxes, calculate the loss function, and then optimize the model parameters using gradient descent. Based on these feature maps, it generates a large number of candidate bounding boxes on the image and predicts for each bounding box whether it contains an object, the object's category, and a confidence score for that prediction. For example, for an input image of a hog badger, the YOLOv5 model might detect one or more bounding boxes in the image, each containing a predicted species, such as hog badger or Asian badger, and a confidence score, such as 0.65. After obtaining the initial prediction results, post-processing, namely non-maximum suppression (NMS), is performed to eliminate overlapping redundant bounding boxes and retain the best prediction results. NMS sorts bounding boxes of the same category according to their confidence scores, then iteratively selects the bounding box with the highest score and removes other bounding boxes that overlap too much with the highest-scoring bounding box, thus ensuring that each object is identified by only one best bounding box. Finally, the general recognition model outputs a general prediction result, including one or more bounding boxes (e.g., coordinates of [x_min, y_min, x_max, y_max]), the corresponding predicted species, such as hog badger, and the confidence score of the prediction, for example, 0.65.

[0024] In step S3, in response to the general prediction result meeting preset conditions, a species object ROI image is extracted from the original image of the species object based on the bounding box. It should be understood that while general recognition models can quickly perform preliminary screening and recognition of a large number of images, they often have limitations in terms of recognition accuracy when dealing with fine-grained classification problems, especially for easily confused species with highly similar morphologies. For example, the problem mentioned in the background art of hog badgers being widely identified as Asian badgers, and the situation where sparrows are unidentified due to their small size, both illustrate the challenges that general models face when dealing with complex or similar species. To avoid performing time-consuming and resource-intensive expert-level fine-grained recognition on all images, while ensuring deeper analysis of species that are difficult for general models to accurately distinguish, an intelligent screening mechanism is needed. Therefore, when the general prediction result meets the preset conditions, the species object ROI image is extracted from the original image based on the bounding box in order to accurately focus the recognition on the species subject, remove irrelevant background interference, and thus provide a cleaner and more discriminative input for the subsequent expert attention classifier, thereby improving the recognition accuracy of easily confused species and the overall system processing efficiency.

[0025] Specifically, in an exemplary embodiment of this application, the preset condition is that the predicted species in the general prediction results belong to a predefined set of easily confused species or have a confidence score less than a preset threshold. It is worth noting that in intelligent biodiversity monitoring, while general identification models, such as YOLOv5, can efficiently perform preliminary species detection, they have limitations in handling certain specific situations. On the one hand, for easily confused species with highly similar morphologies, such as hog badgers and Asian badgers, the general model may struggle to capture the subtle features that distinguish them, resulting in low identification accuracy. On the other hand, even for non-easily confused species, if the confidence score given by the general model is low, it indicates significant uncertainty in its identification results, potentially leading to misjudgment. Directly adopting these low-confidence or easily confused species general prediction results reduces the reliability and accuracy of the entire monitoring system, thereby affecting subsequent biodiversity data analysis and conservation decisions. Therefore, to compensate for the shortcomings of general models and ensure that only highly confident results that have undergone more refined verification are ultimately adopted, setting the preset condition as the predicted species in the general prediction results belonging to a predefined set of easily confused species or having a confidence score less than a preset threshold is crucial. This enables the system to intelligently identify images that require further expert-level refinement, thereby improving overall recognition accuracy and reliability.

[0026] In particular, in one exemplary embodiment of this application, Figure 3 This is a schematic diagram of the data flow in step S3 of the cloud-based intelligent biodiversity monitoring method according to an embodiment of this application. Figure 3 As shown, step S3, in response to the general prediction result satisfying the preset conditions, involves extracting the ROI image of the species object from the original image of the species object based on the bounding box, including: S31, determining whether the predicted species in the general prediction result belongs to a predefined set of easily confused species; S32, if yes, determining that the general prediction result satisfies the preset conditions; S33, if no, determining whether the confidence score in the general prediction result is less than a preset threshold; S34, if the confidence score is less than the preset threshold, determining that the general prediction result satisfies the preset conditions.

[0027] One feasible approach to step S3 is as follows: First, extract the predicted species and confidence score, along with the corresponding bounding box information, from the general prediction results. For example, the general prediction results may contain a bounding box [x1,y1,x2,y2], the predicted species hog badger, and a confidence score of 0.65.

[0028] Next, proceed to S31. This set of easily confused species is pre-defined based on the knowledge of biological experts and historical identification data. It includes morphologically similar species pairs that are easily confused by the general model, such as hog badgers and Asian badgers, sparrows and mountain sparrows, etc. If the predefined set of easily confused species includes hog badgers, and the predicted species in the general prediction result is hog badgers, then it is judged as yes.

[0029] If the judgment result is yes, meaning the predicted species belongs to the easily confused species set, then proceed to S32 to determine that the general prediction result meets the preset conditions. This means that even if the confidence score is high, expert-level fine identification is still required because the species belongs to the easily confused category.

[0030] If the result of S31 is negative, meaning the predicted species does not belong to the easily confused species set, then proceed to S33 to determine whether the confidence score in the general prediction result is less than a preset threshold. This preset threshold is set based on experience and actual needs and is used to measure the reliability of the general model's recognition result. For example, the preset threshold can be set to 0.7. If the confidence score in the general prediction result is 0.55, then it is determined to be positive, meaning 0.55 is less than 0.7.

[0031] If the result of S33 is yes, meaning the confidence score is less than a preset threshold, then proceed to S34 to determine that the general prediction result meets the preset conditions. This means that although it is not a easily confused species, the general model is not confident enough in identifying this species and further expert confirmation is required. In other words, as long as either the predicted species belongs to a predefined set of easily confused species or the confidence score is less than a preset threshold, the general prediction result is determined to meet the preset conditions.

[0032] Once the general prediction result is determined to meet the preset conditions, the system will accurately extract the ROI image of the species from the original image based on the bounding box information provided in the general prediction result. For example, if the bounding box given by the general prediction result is [100,200,400,500], the system will extract a rectangular area from the original image with (100,200) as the upper left corner and (400,500) as the lower right corner to form an ROI image containing only the main species.

[0033] In step S4, in response to the general prediction result not meeting the preset conditions, the original image of the species and the general prediction result are displayed. Accordingly, when processing a large number of images, the general recognition model can provide high-confidence, accurate recognition results for most non-confusing species or images of good quality. However, feeding these results into a computationally intensive expert attention classifier for secondary recognition would create unnecessary computational burden and time delays, reducing overall monitoring efficiency. Therefore, to optimize resource allocation, improve processing efficiency, and ensure users can obtain reliable recognition information promptly, it is reasonable to directly display the prediction result when the general recognition model can accurately identify non-confusing species with high confidence. This not only avoids redundant computational steps but also provides rapid feedback on monitoring information, meeting users' needs for real-time performance and efficiency.

[0034] For example, one feasible way to do step S4 is as follows: First, obtain the general prediction result output by the general recognition model, which includes bounding boxes, predicted species, and confidence scores. For example, the prediction result of the general recognition model for an original image is: bounding box [200, 250, 500, 600], predicted species is magpie, and confidence score is 0.92.

[0035] Next, it is determined whether the general prediction result does not meet the preset conditions. The preset conditions are set as follows: the predicted species in the general prediction result belongs to a predefined set of easily confused species, or the confidence score is less than a preset threshold, for example, the preset threshold is 0.7.

[0036] The specific judgment process is as follows: First, determine whether the predicted species, the magpie, in the general prediction result belongs to a predefined set of easily confused species. Based on the judgment that it is not in this set of easily confused species, next, determine whether the confidence score of 0.92 in the general prediction result is less than a preset threshold of 0.7. Clearly, 0.92 is not less than 0.7. Since the predicted species, the magpie, does not belong to the set of easily confused species, and its confidence score of 0.92 is not less than the preset threshold of 0.7, the general prediction result does not meet the preset conditions.

[0037] When the general prediction result does not meet the preset conditions, the cloud platform will directly display the original image of the species and the general prediction result. Specifically, the cloud platform's user interface module will retrieve the corresponding original image of the species from the data storage module and load and display it. Simultaneously, the bounding box, predicted species, and confidence score from the general prediction result will be visually overlaid or annotated on the original image. For example, on the display screen, the user can see the original magpie image, with a rectangular box precisely marking the magpie's location, and the predicted species (magpie) and confidence score (92%) clearly displayed next to the box or below the image. This direct display method allows monitoring personnel to quickly and intuitively obtain reliable information identified by the general model without further manual intervention or complex fine-grained identification processes, thereby improving monitoring efficiency and the timeliness of information acquisition.

[0038] In step S5, the ROI image of the species is input into an expert attention classifier to obtain expert prediction results. It should be understood that although general recognition models can perform preliminary screening, their accuracy drops significantly when faced with easily confused species that are morphologically similar (such as hog badgers and Asian badgers), or when the confidence level of the general model is low. These subtle inter-class differences, such as the hog badger's hairless snout while the Asian badger's hairy snout, often require more refined feature extraction and discrimination capabilities. General models focus on global features and rapid recognition, making it difficult to capture these local and subtle distinguishing features. To overcome this challenge and significantly improve the recognition accuracy for these complex situations, this application inputs the ROI image of the species into an expert attention classifier to obtain expert prediction results. This expert classifier focuses on extracting multi-dimensional, highly discriminative features from local regions and uses an attention mechanism to focus on the most critical distinguishing points, thereby achieving accurate identification of easily confused species, compensating for the shortcomings of general models, and ensuring the reliability of monitoring results.

[0039] In particular, in one exemplary embodiment of this application, Figure 4 This is a flowchart of step S5 in the cloud-based intelligent biodiversity monitoring method according to an embodiment of this application. Figure 4As shown, step S5, inputting the species object ROI image into the expert attention classifier to obtain expert prediction results, includes: S51, passing the species object ROI image through the multi-layer convolutional structure of the expert attention classifier to obtain the species object ROI edge feature map, species object ROI texture feature map, and species object ROI shape feature map; S52, inputting the species object ROI edge feature map, species object ROI texture feature map, and species object ROI shape feature map into the spatial attention module respectively to obtain the species object ROI edge spatial salient feature map, the species object ROI spatial salient texture feature map, and the species object ROI shape feature map. S53. A salient shape feature map of the species object ROI spatial space is generated; S54. The salient edge feature map, the salient texture feature map, and the salient shape feature map of the species object ROI spatial space are fused to obtain a multi-scale spatial salient fused feature map of the species object ROI; S55. The multi-scale spatial salient fused feature map of the species object ROI is input into the fully connected layer of the expert attention classifier for feature flattening to obtain a classification feature vector of the species object ROI; S56. The classification feature vector of the species object ROI is input into the classification head of the expert attention classifier to obtain the expert prediction result.

[0040] One possible approach to step S5 is as follows: First, obtain the resulting ROI image of the species object. For example, a badger ROI image cropped from the original image might have a size of 200x200 pixels.

[0041] Next, proceed to S51. Understandably, the expert attention classifier employs a deep convolutional neural network architecture. Its multi-layered convolutional structure is specifically designed and trained to extract features of different levels and types from an image in parallel or sequentially. For example, the classifier might contain multiple parallel convolutional branches, each focusing on extracting a specific type of feature: one branch, through specific convolutional kernels, such as Sobel or Canny operator-inspired kernels and activation functions, more sensitively captures edge information of the image. This branch might consist of a series of convolutional layers, batch normalization layers, and ReLU activation functions. For instance, the first convolutional kernel might have a size of 3x3, a stride of 1, and output 64 feature channels, followed by batch normalization and ReLU activation, then downsampling through pooling layers. This process is repeated multiple times to ultimately generate a species. One branch extracts the edge feature map of the ROI (Region of Interest). Another branch extracts texture information from the image, such as hair texture and skin texture, through different combinations of convolutional kernels. This branch may use multi-scale convolutional kernels, such as 3x3 or 5x5, in parallel processing to capture texture details at different scales. It also includes convolutional layers, batch normalization layers, and activation functions to generate texture feature maps of the ROI. A third branch focuses on extracting the overall outline and local morphological features of the species. This branch may contain deeper convolutional structures to learn more abstract shape representations and generate shape feature maps of the ROI. The weights and bias parameters in these convolutional layers are trained on a dataset of easily confused species with a large number of finely labeled features through backpropagation and optimization algorithms, aiming to maximize the expressive power of different feature maps for subtle differences between species.

[0042] Then, S52 is performed. That is, each spatial attention module adaptively learns and generates a spatial weight map based on the content of the input feature map. Each spatial attention module contains one or more convolutional layers and a sigmoid activation function. For example, for a species object ROI edge feature map, its corresponding spatial attention module might first compress its channel count through a 1x1 convolutional layer, then generate a single-channel weight map with the same spatial size as the feature map through another convolutional layer, and finally normalize the weight values ​​to between 0 and 1 using the sigmoid function. This weight map indicates which spatial locations in the feature map are more important for the current task, i.e., species identification. For example, for a badger ROI image, the edge attention module might assign higher weights to areas such as the snout edge and ear contour; the texture attention module might focus on the density and direction of hair distribution; and the shape attention module might focus on the overall body shape and head shape. By performing element-wise multiplication between the original feature map and the corresponding spatial weight map, the feature representation of important regions can be enhanced, while the features of unimportant or interfering regions can be suppressed, thereby obtaining the spatial salient feature map of the species object ROI edge, the spatial salient texture feature map of the species object ROI, and the spatial salient shape feature map of the species object ROI.

[0043] Next, S53. The fusion method can employ weighted summation, feature connection, or more complex fusion networks. For example, a weighted summation method can be used to superimpose the three spatial salient feature maps according to preset weights, which can also be learned through training, to obtain a comprehensive multi-scale spatial salient fusion feature map, i.e., a multi-scale spatial salient fusion feature map of the species object ROI.

[0044] Specifically, in the expert attention classifier, the edge, texture, and shape feature maps of the species object's ROI have been enhanced for saliency in their respective dimensions through multi-layer convolutional structures and spatial attention modules. This means that each feature map independently highlights its key information in species identification. However, while these features are saliency in themselves, they are complementary, and a single-dimensional feature is often insufficient to fully distinguish between species with highly similar morphologies. For example, edges alone may not be enough to distinguish the subtle differences in fur between a hog badger and an Asiatic badger. Therefore, to fully utilize this multi-dimensional, saliency-enhancing information and overcome the limitations of single features, they need to be effectively fused. This fusion aims to maintain the saliency of each feature while integrating information from different dimensions to form a more comprehensive and discriminative multi-scale spatially saliency-enhancing feature map, thereby significantly improving the accuracy of fine-grained species identification.

[0045] Based on this, in a preferred exemplary embodiment of this application, step S53, fusing the spatial salient feature map of the species object ROI edge, the spatial salient texture feature map of the species object ROI, and the spatial salient shape feature map of the species object ROI to obtain a multi-scale spatial salient fused feature map of the species object ROI, includes: first, calculating the mean feature value of each position in the spatial salient feature map of the species object ROI edge, the spatial salient texture feature map of the species object ROI, and the spatial salient shape feature map of the species object ROI to obtain a mean feature map of the species object ROI. It should be understood that in the expert attention classifier, the edge, texture, and shape feature maps have already been saliency-enhanced through the spatial attention module, but they are still independent representations in their respective dimensions. In order to subsequently establish spatial gradient-aware response calibration for different feature distributions, a unified reference that can simultaneously contain the spatial description of edge, texture, and shape primitives at the local scale is needed. By calculating the mean feature value of each position (i.e., each pixel or feature grid point) in these three spatial salient feature maps, a comprehensive and smooth feature map can be obtained. This mean feature map, to a certain extent, represents the average spatial distribution and intensity of species objects across these core visual dimensions, providing a stable benchmark for subsequent relative gradient response calculations. It can initially integrate local responses under different convolutional dimensions, forming a preliminary feature space representation, laying the foundation for more refined fusion later.

[0046] Next, based on the mean feature map of the species object ROI, the spatial relative gradient responses of the species object ROI edge spatial salient feature map, the species object ROI spatial salient texture feature map, and the species object ROI spatial salient shape feature map are calculated respectively to obtain the species object ROI edge spatial salient response feature map, the species object ROI spatial salient texture response feature map, and the species object ROI spatial salient shape response feature map, that is: ;in, This is a feature map of the mean ROI of a species. It is a salient feature map of the marginal space of the ROI of a species object. Species object ROI spatial salient texture feature map Significant shape feature map of ROI space for species objects The Middle Each feature map It is a dot product by position. It is subtracted based on position. To calculate the reciprocal of each eigenvalue in the feature map, It is the first of the three salient response feature maps of species object ROI edge space, species object ROI space salient texture response feature map, and species object ROI space salient shape response feature map. Each feature map. It is understandable that the mean feature map of the ROI of a species object... Given that the spatial description of edges, textures, and shape primitives at the local scale is already included, by... Comparison, or differencing, can highlight each individual feature map. Local variations and differences relative to the overall average level. This relative gradient response captures more precisely the uniqueness and salience of each dimension (edge, texture, shape) at a specific spatial location, rather than simply focusing on absolute values. For example, if the edge features of a region are... It is very prominent in the middle, but... If the gradient is relatively flat, then the relative gradient response will be high, indicating that the edge has high discriminative power in that region. This allows for the calibration of spatial gradient perception for different feature distributions, enabling subsequent fusion to more effectively utilize this calibrated, highly discriminative local information.

[0047] Then, the spatially significant response feature maps of species object ROI edge space, spatially significant texture response feature map, and spatially significant shape response feature map are processed by local hot zone convolution kernel function to obtain spatially significant fused perceptual feature maps of species object ROI edge space, spatially significant fused perceptual feature map of species object ROI, and spatially significant fused perceptual feature map of species object ROI, namely: ;in, yes The convolution kernel function, It is added based on the position point. yes eigenvalue standard deviation The three salient fusion perceptual feature maps of species object ROIs are: the salient fusion perceptual feature map of species object ROIs' edge space, the salient fusion perceptual feature map of species object ROIs' texture space, and the salient fusion perceptual feature map of species object ROIs' shape space. Each feature map. This should be understandable. This represents a convolution operation, where the convolution kernel function has the same characteristics as the feature map. Same standard deviation This means that the shape and strength of the convolution kernel adaptively adjust according to the statistical properties of the input feature map, enabling it to better capture the saliency of local regions. This local hotspot convolution kernel function processing can focus on local regions with corresponding saliency cues, further enhancing the feature representation of these key local regions through the statistical correlation presented by the convolutional neighborhood, which is the same as the saliency of the saliency dimension. For example, the subtle feature of the hairless snout of a badger, even if not obvious in the edge feature map, may have a unique texture pattern in the texture feature map. Through this local hotspot processing, it can be ensured that these local information with high saliency in their respective dimensions are fully perceived and enhanced, providing high-quality local discriminative features for the final fusion.

[0048] Finally, the spatial salient fusion perceptual feature map of the species object ROI edge, the spatial salient fusion perceptual feature map of the species object ROI, and the spatial salient fusion perceptual feature map of the species object ROI are fused point by point to obtain the multi-scale spatial salient fusion feature map of the species object ROI, namely: ;in, This is a multi-scale spatially salient fusion feature map of the species object ROI. In other words, this final fusion step aims to obtain the overall spatial distribution perceptual response by simulating visual perception, based on the highly saliency (i.e., salient local region discrimination) of the labeled ROIs with corresponding convolutional dimensions, thereby achieving effective feature map fusion. The point-plus-fusion method integrates local semantic structures of different dimensions based on spatial salient distribution. This fusion approach ensures that local regions exhibiting high discriminative power in edge, texture, and shape dimensions are emphasized to the greatest extent, while suppressing irrelevant or redundant information. The resulting multi-scale spatially salient fusion feature map of the species object ROI is a highly condensed and highly discriminative feature representation. It integrates key information of the species in different visual dimensions, providing optimized and comprehensive input for subsequent fully connected layer classification, thus significantly improving the accuracy of fine-grained species identification.

[0049] Next, S54 is performed. The fused feature map is multi-dimensional (e.g., height x width x number of channels). To be input into the fully connected layer for classification, it needs to be flattened into a one-dimensional feature vector. This flattening operation arranges all pixel values ​​of the multi-dimensional feature map sequentially into a long vector. Subsequently, this one-dimensional feature vector is input into the fully connected layer of the expert attention classifier. Each neuron in the fully connected layer is connected to all neurons in the previous layer, i.e., the flattened feature vector, and has independent weights and bias parameters. The specific processing is as follows: each neuron multiplies all its inputs, i.e., each dimension value of the feature vector, with the corresponding connection weights, then sums all the products and adds a bias term. This weighted sum is then transformed through a non-linear activation function, such as ReLU, to introduce non-linearity, enabling the model to learn and express more complex feature relationships. Through the stacking of multiple fully connected layers, neurons learn the complex non-linear relationships between these flattened features, further extracting high-level semantic information to obtain the species object ROI classification feature vector. This information is crucial for the final species classification. The weights and bias parameters in these fully connected layers are also continuously adjusted and optimized during training through backpropagation and optimization algorithms.

[0050] Finally, S55 is performed. The classification head consists of one or more fully connected layers and a Softmax activation function. It maps the classification feature vector to the probability distribution of each species category. Specifically, the classification head receives the flattened feature vector as input, performs a linear transformation through the first fully connected layer, then possibly passes through a ReLU activation function, and then passes through a second fully connected layer to map the features to a dimension equal to the number of species categories. Finally, the Softmax activation function converts these output values ​​into a vector representing the probability of each category, with all probabilities summing to 1. For example, for an input image of a hog badger ROI, the classification head will output a probability vector where the probability of the hog badger category might be 0.98, and the probability of the Asiatic badger category might be 0.02. Ultimately, the category with the highest probability is selected as the expert prediction result, e.g., hog badger, along with a corresponding confidence score.

[0051] In step S6, the expert prediction results and the original image of the species are displayed. That is, for easily confused species that are difficult for general identification models to distinguish accurately, or for identification results with low confidence, the expert attention classifier refines the process, resulting in more accurate and reliable expert predictions. These results are derived after multi-dimensional feature extraction and attention mechanisms focusing on key details, representing the highest accuracy of the current identification process. To ensure that monitoring personnel can obtain this meticulously verified final identification information in a timely and intuitive manner, and accurately judge and record the species in the original images, thus providing a solid data foundation for subsequent biodiversity research and conservation strategy formulation, displaying the expert prediction results and the original image of the species is crucial. This is not only the final confirmation of the complex identification task, but also a key step in the closed loop of the entire intelligent monitoring process, ensuring the integrity and usability of the information.

[0052] For example, one feasible approach to step S6 is as follows: First, obtain the expert prediction result output by the expert attention classifier, which includes the predicted species and confidence score. For instance, for an image previously identified as a hog badger by the general model but with a low confidence score (0.65) or judged as an easily confused species, after processing by the expert attention classifier, the final expert prediction result might be a hog badger with a confidence score of 0.98. Simultaneously, it is also necessary to retrieve the original image of the species object corresponding to this prediction result from the data storage module of the cloud platform.

[0053] Next, the cloud platform's user interface module is responsible for visualizing this information. Specifically, the original image is loaded and displayed on the user interface, such as on a web-based biodiversity monitoring and management platform. To visually represent the identification results, the species category and confidence score from the expert prediction results are clearly labeled next to the original image or displayed as text below the image. For example, on the display, the user can see the original image of a hog badger, with the expert-predicted species: hog badger, confidence score: 98%, etc. displayed below the image. If the general identification model has already provided bounding box information in step S2, and the image has entered steps S3 and S5 for refined identification, then in step S6, the bounding box can also be overlaid on the original image to more intuitively indicate the species' location in the image.

[0054] Furthermore, to facilitate data management and analysis for users, these expert predictions and corresponding raw image information are stored in the cloud platform's database and linked to metadata such as front-end devices, collection time, and geographical location. Users can quickly retrieve and view specific identification records based on criteria such as species name, identification time, and confidence level using the platform's query and filtering functions. For example, monitoring personnel can log in to the platform to view all images identified as hog badgers by the expert classifier and further verify their accuracy, or use them for subsequent population statistics and behavioral analysis. This display and storage mechanism ensures the transparency and traceability of identification results and facilitates the long-term accumulation and in-depth analysis of biodiversity data.

[0055] In summary, the cloud-based intelligent biodiversity monitoring method based on the embodiments of this application is explained. Specifically, species images collected by the front-end device are first uploaded to the cloud platform, where a general recognition model makes a preliminary judgment. If the general model's recognition result has high confidence or is not a easily confused species, the result is directly output. Crucially, when the general model's recognition result is uncertain or involves predefined easily confused species, the system extracts the species' ROI image and feeds it into a specially trained expert attention classifier. This classifier extracts multi-scale spatial salient features of the species, such as edges, textures, and shapes, and performs deep fusion to accurately capture subtle differences, ultimately providing expert-level prediction results. This recognition strategy fully leverages the efficiency of the general model and the accuracy of the expert model, effectively overcoming the challenges of species identification in complex environments and significantly improving the accuracy and reliability of biodiversity monitoring.

[0056] Figure 5 This is a block diagram of a cloud-based intelligent biodiversity monitoring system according to an embodiment of this application. Figure 5 As shown, the cloud-based intelligent biodiversity monitoring system 100 according to an embodiment of this application includes: a species object original image acquisition module 110, used to receive original images of species objects collected by front-end devices on the cloud platform; a general prediction module 120, used to input the original images of species objects into a general recognition model to obtain a general prediction result, the general prediction result including a bounding box, a predicted species, and a confidence score; a species object ROI image cropping module 130, used to crop a species object ROI image from the original images of species objects based on the bounding box in response to the general prediction result meeting a preset condition; a general prediction result comparison module 140, used to display the original images of species objects and the general prediction result in response to the general prediction result not meeting the preset condition; an expert prediction module 150, used to input the species object ROI image into an expert attention classifier to obtain an expert prediction result; and a display module 160, used to display the expert prediction result and the original images of species objects.

[0057] Those skilled in the art will understand that the specific operations of each step in the cloud-based intelligent biodiversity monitoring system described above have been referenced. Figures 1 to 4 The cloud-based intelligent biodiversity monitoring method is described in detail here, and therefore, its repeated description will be omitted.

Claims

1. A cloud-based intelligent biodiversity monitoring method, characterized in that, include: The cloud platform receives raw images of species objects collected by front-end devices; The original image of the species object is input into a general recognition model to obtain a general prediction result, which includes a bounding box, a predicted species, and a confidence score. In response to the general prediction result meeting the preset conditions, the ROI image of the species object is extracted from the original image of the species object based on the bounding box; In response to the general prediction result not meeting the preset conditions, the original image of the species object and the general prediction result are displayed; The ROI image of the species object is input into an expert attention classifier to obtain expert prediction results; Display the expert prediction results and the original image of the species object; The process of inputting the ROI image of the species object into an expert attention classifier to obtain expert prediction results includes: The species object ROI image is passed through a multi-layer convolutional structure of an expert attention classifier to obtain the species object ROI edge feature map, species object ROI texture feature map, and species object ROI shape feature map. The species object ROI edge feature map, species object ROI texture feature map, and species object ROI shape feature map are respectively input into the spatial attention module to obtain the species object ROI spatial salient feature map, species object ROI spatial salient texture feature map, and species object ROI spatial salient shape feature map. The method involves fusing the spatial salient feature map of the species object's ROI edge, spatial salient texture, and spatial salient shape to obtain a multi-scale spatial salient fused feature map of the species object's ROI. This includes: calculating the mean eigenvalues ​​at each location in the spatial salient feature map of the species object's ROI edge, spatial salient texture, and spatial salient shape to obtain a mean feature map of the species object's ROI; and based on the mean feature map of the species object's ROI, calculating the spatial relative gradient responses of the spatial salient feature map of the species object's ROI edge, spatial salient texture, and spatial salient shape to obtain a spatial salient response feature map of the species object's ROI edge and species object ROI. The OI spatial salient texture response feature map and the species object ROI spatial salient shape response feature map are obtained. Local hot zone convolution kernel functions are applied to the species object ROI edge spatial salient fusion perception feature map, the species object ROI spatial salient texture fusion perception feature map, and the species object ROI spatial salient shape fusion perception feature map to obtain the species object ROI edge spatial salient fusion perception feature map, the species object ROI spatial salient texture fusion perception feature map, and the species object ROI spatial salient shape fusion perception feature map. Point-by-point fusion is then performed on the species object ROI edge spatial salient fusion perception feature map, the species object ROI spatial salient texture fusion perception feature map, and the species object ROI spatial salient shape fusion perception feature map to obtain the species object ROI multi-scale spatial salient fusion feature map. The multi-scale spatial saliency fusion feature map of the species object ROI is input into the fully connected layer of the expert attention classifier for feature flattening to obtain the species object ROI classification feature vector; The ROI classification feature vector of the species object is input into the classification head of the expert attention classifier to obtain the expert prediction result.

2. The cloud-based intelligent biodiversity monitoring method according to claim 1, characterized in that, The general recognition model is the standard YOLOv5 model.

3. The cloud-based intelligent biodiversity monitoring method according to claim 1, characterized in that, The preset condition is that the predicted species in the general prediction result belongs to a predefined set of easily confused species or the confidence score is less than a preset threshold.

4. The cloud-based intelligent biodiversity monitoring method according to claim 3, characterized in that, In response to the general prediction result satisfying a preset condition, the ROI image of the species object is extracted from the original image of the species object based on the bounding box, including: Determine whether the predicted species in the general prediction results belong to a predefined set of easily confused species; If so, the general prediction result is determined to meet the preset conditions; If not, determine whether the confidence score in the general prediction result is less than a preset threshold; If the confidence score is less than a preset threshold, the general prediction result is determined to meet the preset conditions.

5. A cloud-based intelligent biodiversity monitoring system according to claim 1, characterized in that, include: The original image acquisition module for species objects is used by the cloud platform to receive original images of species objects collected by the front-end devices. A general prediction module is used to input the original image of the species object into a general recognition model to obtain a general prediction result, which includes a bounding box, a predicted species, and a confidence score. The species object ROI image cropping module is used to crop the species object ROI image from the original image of the species object based on the bounding box in response to the general prediction result meeting the preset conditions. A general prediction result comparison module is used to display the original image of the species object and the general prediction result in response to the general prediction result not meeting the preset conditions. The expert prediction module is used to input the ROI image of the species object into the expert attention classifier to obtain the expert prediction result; The display module is used to display the expert prediction results and the original image of the species object.

Images