Construction and verification method of large-scale wild animal data set

Abstract

The invention provides a large-scale wild animal data set construction and verification method, and the method comprises the steps: constructing a target data set which comprises a plurality of types of wild animal sample images; inputting a test set in the target data set into a hybrid expert model to obtain a target classification result, the hybrid expert model being obtained by training a training set in the target data set; and based on the analysis of the target classification result, determining a verification result of the test set in the target data set.

Description

Technical Field

[0001] This invention relates to the field of dataset construction technology, specifically to a method for constructing and validating a large-scale wildlife dataset. Background Technology

[0002] Several datasets have already driven the development of related research. Caltech-UCSD Birds-200-2011, as the most representative fine-grained recognition dataset, contains 11,788 images of 200 bird species, providing detailed bounding boxes and part annotations (including 15 key parts such as beaks, wings, and tails) as well as 312 binary attribute annotations, providing a dataset foundation for researchers to develop and validate bird fine-grained recognition algorithms based on partial features. NABirds, as a large-scale dataset of North American birds, covers 48,562 images of 1,011 species. Its annotations adopt a hierarchical classification system in the form of a phylogenetic tree, and compared with CUB-200, NABirds has more diverse data sources.

[0003] While existing wildlife datasets have made significant strides in advancing related research, several limitations restrict their application effectiveness and generalization capabilities. The primary issue is the insufficient realism of the image acquisition environment. Early datasets such as CUB-200 often used specimen-style photography or images acquired under controlled environments, lacking the realism of complex field environments. This limitation makes it difficult for trained models to adapt to environmental changes in real-world application scenarios, reducing recognition accuracy in real-world situations. Furthermore, images acquired in the wild face variations in lighting, occlusion, and background interference, increasing the difficulty and uncertainty of annotation. Professional annotation relies on the participation of wildlife biologists, resulting in high annotation costs and hindering the quality, expansion, and update cycle of the datasets.

[0004] In summary, the current dataset construction process lacks sufficient automation. Image acquisition, annotation verification, and quality control rely heavily on manual operations and specialized knowledge, resulting in low dataset construction efficiency. This low level of automation not only limits the expansion and iteration speed of datasets but also hinders the implementation of diverse annotations. Therefore, a method for constructing and intelligently verifying large-scale wildlife datasets is urgently needed to significantly improve the accuracy and efficiency of wildlife dataset construction. Summary of the Invention

[0005] This invention provides a method for constructing and validating a large-scale wildlife dataset, comprising: constructing a target dataset, which includes sample images of multiple types of wildlife; inputting a test set from the target dataset into a hybrid expert model to obtain a target classification result, wherein the hybrid expert model is trained on a training set from the target dataset; and determining the validation result of the test set from the target dataset based on the analysis of the target classification result.

[0006] In the above scheme, constructing the target dataset includes: collecting multi-source heterogeneous data through a preset monitoring system, the multi-source heterogeneous data including original image data and video data of multiple types of wild animal samples; preprocessing the multi-source heterogeneous data and extracting key frames to determine the target frame image data; and using a preset automatic data annotation tool to annotate the target frame image data and the original image data respectively to obtain the target dataset containing images of multiple types of wild animal samples.

[0007] In the above scheme, the hybrid expert model is trained on the training set of the target dataset. The training process includes: acquiring the training set and performing feature preprocessing; dividing the preprocessed training set into multiple training subsets, and training each expert model using the cross-entropy loss function based on each training subset; after extracting multi-scale features from the images in the training set and performing multi-level feature fusion, using a multilayer perceptron to map the fused features to the routing decision space; calculating the activation weight of each expert model based on the fused features; determining the target expert model among multiple expert models based on a preset threshold filtering strategy; and fusing the prediction probabilities of multiple target expert models to obtain the target classification result.

[0008] In the above scheme, the preprocessed training set is divided into multiple training subsets, and each expert model is trained using the cross-entropy loss function based on each training subset. This includes: dividing the preprocessed training set into multiple training subsets according to the expertise of each expert model, where each training subset corresponds to one expert model; and training each expert model based on each training subset with cross-entropy loss as the main objective and by introducing three types of regularization terms.

[0009] In the above scheme, after extracting multi-scale features from the images in the training set and performing multi-level feature fusion, a multilayer perceptron is used to map the fused features to the routing decision space. This includes: performing multi-scale processing on the images in the training set to extract global and local features respectively; achieving multi-level fusion through attention mechanisms or feature concatenation to obtain fused features; and using a multilayer perceptron to map the fused features to the routing decision space to obtain routing feature vectors. The routing decision space includes multiple clustered image regions, and each clustered image region corresponds to one or more expert models.

[0010] In the above scheme, the activation weight of each expert model is calculated based on the fused features, including: calculating the activation weight of each expert model by using a four-layer cascaded structure based on the routing feature vector.

[0011] In the above scheme, the target expert model among multiple expert models is determined based on a preset threshold filtering strategy, including: filtering out the target expert model among multiple expert models by setting a fixed threshold, a dynamic threshold, and a relative threshold.

[0012] In the above scheme, based on the analysis of the target classification results, the verification results of the test set in the target dataset are determined, including: based on the prediction probability of each target expert model, classification similarity calculation is performed to obtain similarity results to complete the quantitative analysis of prediction uncertainty.

[0013] In the above scheme, based on the analysis of the target classification results, the verification results of the test set in the target dataset are determined, including: based on the prediction probability of each target expert model, comparing the prediction results of each target expert with the true label; and by setting confidence threshold and stage entropy threshold, identifying high-confidence error sample images and high-disagreement error sample images.

[0014] In the above scheme, based on the analysis of the target classification results, the verification results of the test set in the target dataset are determined, including: fine-tuning each expert model and updating the expert pool based on the identified erroneous sample images.

[0015] The technical solutions of the embodiments of the present invention have at least the following beneficial effects:

[0016] (1) This method includes data collection, data cleaning, data annotation, and efficient intelligent data verification, constructing a complete toolchain from raw data to high-quality labeled datasets. By integrating data sources such as actual collected data and data acquired from the network, the richness of the dataset is improved; by using automated annotation tools, the efficiency of data annotation is improved while reducing manual costs; and by combining an efficient intelligent verification method for large-scale wildlife datasets based on hybrid expert models, the quality of the dataset is effectively improved.

[0017] (2) This method effectively solves the problems of sample representativeness and model generalization in fine-grained recognition by intelligently sampling and preprocessing the data. The model, which combines expert models, considers both the cluster structure preservation and the dynamics of the inherent difficulty of the samples, and accurately quantifies the combined impact of different visual modes on model performance. Compared with traditional random sampling, this mechanism maintains high baseline accuracy even when the amount of training data is reduced, and significantly reduces the cost of data annotation.

[0018] (3) This method improves the accuracy of the model in fine-grained identification of wild animals by adjusting the threshold of the number of activated experts and the confidence weight parameters of the hybrid expert model, so as to make the target classification results more accurate and thus improve the accuracy and efficiency of the overall dataset construction. Attached Figure Description

[0019] Figure 1A flowchart illustrating a method for constructing and validating a large-scale wildlife dataset according to an embodiment of the present invention is shown schematically.

[0020] Figure 2 A flowchart illustrating the construction of a target dataset according to an embodiment of the present invention is shown schematically;

[0021] Figure 3 The flowchart illustrating the process of obtaining a hybrid expert model by training a training set according to an embodiment of the present invention is shown. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0023] Figure 1 A flowchart illustrating a method for constructing and validating a large-scale wildlife dataset according to an embodiment of the present invention is shown.

[0024] Please refer to the following for details. Figure 1 The specific process of the method for constructing and verifying a large-scale wildlife dataset according to an embodiment of the present invention includes operations S110 to S130.

[0025] In operation S110, a target dataset is constructed, which includes images of various types of wild animal samples.

[0026] Figure 2 A flowchart illustrating the construction of a target dataset according to an embodiment of the present invention is shown.

[0027] Please refer to the following for details. Figure 1 The specific process for constructing the target dataset in this embodiment of the invention includes operations S210~S230.

[0028] When operating S210, multi-source heterogeneous data is collected through a preset monitoring system. The multi-source heterogeneous data includes raw image data and video data of various wild animal samples.

[0029] For example, a multimodal intelligent monitoring system for birds or other wild animals in open scenarios is constructed, which includes hardware devices and system architecture.

[0030] For example, the hardware configuration includes a high-definition smart camera equipped with a 1 / 1.2-inch CMOS sensor and supporting 40x optical zoom for acquiring high-resolution images and video data of wildlife. The hardware configuration also includes self-developed intelligent acoustic monitoring equipment, such as a built-in microphone, power supply, network module, and edge computing module. The edge computing module has noise reduction and sound recognition algorithms to improve audio data quality and analysis efficiency. The hardware configuration also includes an infrared camera for automatically detecting and recording images and video data of animal species within a 10-meter radius.

[0031] For example, the system architecture can adopt a cloud-based collaborative processing mode, where edge devices use lightweight deep learning algorithms to initially process the raw audio and video stream data, and then upload the processed data to the cloud server for further in-depth analysis.

[0032] In operation S220, multi-source heterogeneous data is preprocessed and key frames are extracted to determine the target frame image data.

[0033] For example, the original image and video data of the various wild animal samples were standardized in format. For instance, the data format was unified, the image data was converted to JPEG format to achieve an optimal balance between file size and image quality; the audio data was standardized to WAV format with a uniform sampling rate of 48kHz to ensure complete preservation of acoustic information; and the video data was encoded in MP4 format using H.264.

[0034] Furthermore, keyframe extraction is performed on the video data. For example, the second frame of the video is used as the initial reference frame, and then the feature similarity between each subsequent frame and the initial frame is calculated to select the most representative keyframes, thereby achieving effective data compression and image quality optimization.

[0035] During operation S230, based on the preset automatic data annotation tool, the target frame image data and the original image data are annotated respectively to obtain a target dataset containing multiple types of wild animal sample images.

[0036] According to an embodiment of the present invention, after obtaining the target frame image data and the preprocessed original image based on the above operations, the data is automatically labeled based on deep learning.

[0037] Understandably, this involves building an automated data labeling pipeline that includes four preset automated labeling tools.

[0038] For example, a spatial positioning and annotation tool uses object detection algorithms to accurately identify individual animals based on image data preprocessing and feature extraction, and generates bounding box annotation information with location coordinates and confidence scores.

[0039] For example, a fine-grained classification and labeling tool receives preprocessed image data, and through feature encoding and classifier prediction, outputs hierarchical classification and labeling information containing four levels: order, family, genus, and species. It can also provide classification confidence assessment, providing a basis for subsequent quality control.

[0040] For example, a pixel-level segmentation annotation tool first takes the bounding box information generated by spatial positioning annotation as input, guides the segmentation algorithm to accurately segment the target image, and generates pixel-level binary mask annotation results.

[0041] For example, a multimodal semantic annotation tool converts image data into multimodal input and generates structured text annotations using biologically specific prompt templates, covering animal behavior, environmental features, and ecological status. This achieves standardized association between image and text information, supporting the needs of subsequent multimodal learning tasks.

[0042] In operation S120, the test set in the target dataset is input into the hybrid expert model to obtain the target classification result, where the hybrid expert model is trained on the training set in the target dataset.

[0043] According to an embodiment of the present invention, the hybrid expert model is obtained by training on the training set of the target dataset. After the hybrid expert model is trained on the training set, the test set is then input into the hybrid expert model to obtain the target classification result.

[0044] The process of training this hybrid expert model using the training set is explained in detail below.

[0045] Figure 3 The flowchart illustrating the process of obtaining a hybrid expert model by training a training set according to an embodiment of the present invention is shown.

[0046] Please refer to the following for details. Figure 3 The specific process of obtaining a hybrid expert model by training a training set in this embodiment of the invention includes operations S310 to S360.

[0047] In operation S310, the training set is acquired and feature preprocessing is performed.

[0048] For example, the training set includes 1000 images of various wild animals. Deep features are extracted using pre-trained models (such as ResNet and ViT), and the images are divided into multiple clusters based on feature similarity using clustering algorithms (such as K-means). The number of training samples is allocated proportionally according to the cluster size to ensure a balanced data distribution; at the same time, a sample difficulty score (such as the clarity of classification boundaries) is calculated to select representative samples.

[0049] For example, for each cluster Intelligent selection of the number of training samples according to proportion:

[0050]

[0051] Where, n j G represents the number of samples to be selected for the j-th cluster; j This represents the total number of original samples in the j-th cluster; This represents the floor function; This means ensuring that at least two samples are selected for each cluster.

[0052] Furthermore, calculate the difficulty score for each sample.

[0053]

[0054] in, d represents the predicted probability of the pre-trained model for the class c to which sample i belongs. i This represents the difficulty score of the i-th sample (the larger the score, the more difficult the sample). This indicates that the maximum probability value is taken among all categories C.

[0055] Finally, a general dataset is constructed, containing basic features and difficulty labels, which is used as initial training data for all expert models to carry out subsequent specialized training.

[0056] In operation S320, the preprocessed training set is divided into multiple training subsets, and each expert model is trained based on each training subset using the cross-entropy loss function.

[0057] According to an embodiment of the present invention, the preprocessed training set is divided into multiple training subsets, and each expert model is trained using the cross-entropy loss function based on each training subset, including: dividing the preprocessed training set into multiple training subsets according to the expertise of each expert model, wherein each training subset corresponds to an expert model; and training each expert model based on each training subset with cross-entropy loss as the main objective and by introducing three types of regularization terms.

[0058] For example, specialized training subsets are allocated from a general dataset based on the expert models' areas of expertise (e.g., felines, birds). During training, cross-entropy loss is the primary objective, and three types of regularization terms are introduced: sparsity (L1 regularization) to reduce redundant parameters, diversity (difference loss) to encourage models to focus on different features, and confidence (label smoothing) to improve prediction certainty. Through joint optimization, each expert model achieves high accuracy on its specific animal subset while maintaining complementarity between models.

[0059] When operating the S330, after extracting multi-scale features from the images in the training set and performing multi-level feature fusion, a multilayer perceptron is used to map the fused features to the routing decision space.

[0060] According to an embodiment of the present invention, after extracting multi-scale features from images in the training set and performing multi-level feature fusion, a multilayer perceptron is used to map the fused features to the routing decision space. This includes: performing multi-scale processing on the images in the training set to extract global and local features respectively; achieving multi-level fusion through an attention mechanism or feature concatenation to obtain fused features; and using a multilayer perceptron to map the fused features to the routing decision space to obtain a routing feature vector. The routing decision space includes multiple clustered image regions, and each clustered image region corresponds to one or more expert models.

[0061] Understandably, the image to be verified undergoes multi-scale processing, extracting global features (overall contour, pose) and local features (eye details, texture, etc.) separately, and then achieving multi-level fusion through attention mechanisms or feature concatenation. The fused features are input into a multilayer perceptron (MLP), projected onto the routing decision space, and generate a low-dimensional vector. This vector encodes the correlation between the image and the areas of expertise of each expert model, providing a basis for subsequent weight calculations.

[0062] In operation S340, the activation weights of each expert model are calculated based on the fused features.

[0063] According to an embodiment of the present invention, the activation weight of each expert model is calculated based on the fused features, including: calculating the activation weight of each expert model by adopting a four-layer cascaded structure based on the routing feature vector.

[0064] Specifically, a four-layer cascaded structure is adopted, where the input processing layer reduces the dimensionality of the routing feature vector; the weight prediction network learns the weight mapping relationship through two fully connected layers; the ReLU activation function enhances the nonlinear modeling capability; and Softmax normalization ensures that the sum of the weights is 1. For example, a "cheetah" image may generate expert weights [0.6 (feline expert), 0.3 (running posture expert), 0.1 (other)]. This adaptive calculation ensures that the sum of the activation weights is 1, realizing the probabilistic output of the expert model's selection decision.

[0065] The specific calculation formula is as follows:

[0066]

[0067] Where g is the expert weight vector (normalized by softmax), W r It is the router's weight matrix. b represents the feature extraction result of the router on the input x.r This represents the router's bias term. Softmax is the normalization function that ensures the weights sum to 1.

[0068] Its loss function expression is:

[0069]

[0070] Among them, L total It is the total loss, L CE It is the cross-entropy loss (primarily the classification loss), L sparse It is a sparsity loss (experts are encouraged to choose sparsity), L diversity Indicating a loss of diversity (encouraging expert diversity), L confidence This means confidence loss (improving prediction confidence). This represents the weighting coefficient of each loss term.

[0071] In operation S350, based on a preset threshold filtering strategy, a target expert model is determined from multiple expert models, where the target expert model includes multiple models.

[0072] When operating the S360, the predicted probabilities of multiple target expert models are fused to obtain the target classification result.

[0073] Specifically, a three-pronged strategy is used to select the target expert model from multiple expert models. For example, this involves setting a fixed threshold (e.g., experts with a weight <0.1 are directly eliminated), a dynamic threshold (adjusting the threshold based on the mean and variance of the current batch weights), and a relative threshold (retaining only the top 3 experts). For instance, if the dynamically calculated mean is 0.2, experts with a weight >0.15 might be retained.

[0074] The expert selection formula is:

[0075]

[0076] Among them, E selected It is the set of selected experts, TOPK is the selection of the K highest-scoring experts, i is the expert index, and g is the set of selected experts. i c is the routing weight of the i-th expert. i N is the confidence score of the i-th expert. experts is the total number of experts, and T represents the threshold for the number of experts selected.

[0077] The final target classification result is obtained by weighted fusion of multiple expert prediction probabilities:

[0078]

[0079] in the formula This represents the final predicted probability that input x belongs to category c. g is the probability predicted by the i-th expert that input x belongs to category c. i c is the routing weight of the i-th expert. i The numerator is the confidence score of the i-th expert, the numerator is the sum of the weighted prediction probabilities of all selected experts, and the denominator represents the weight normalization factor.

[0080] It should be noted that the number of experts can also be dynamically controlled. By setting a maximum number of active experts (e.g., 3), if the number of remaining experts exceeds the limit after threshold filtering, only the top few with the highest weights will be retained. This mechanism avoids wasting computational resources. For example, only three experts—"Felidae," "Texture," and "Birds"—are called instead of all models, which can significantly improve the computational efficiency of the hybrid expert model.

[0081] In operation S130, based on the analysis of the target classification results, the verification results of the test set in the target dataset are determined.

[0082] According to an embodiment of the present invention, after obtaining the target classification result, a verification analysis is performed.

[0083] Understandably, after determining the validation results of the test set, the hybrid expert model is adjusted to improve its accuracy in fine-grained identification of wild animals, thereby increasing the accuracy of target classification results and improving the accuracy and efficiency of the overall dataset construction.

[0084] Specifically, in embodiments of the present invention, classification similarity calculation is performed based on the prediction probability of each target expert model to obtain similarity results and complete the quantitative analysis of prediction uncertainty.

[0085] Understandably, uncertainty is assessed by calculating the similarity of the distribution of expert predictions. For example, if all experts output "leopard" with similar probabilities, the consistency is high and the result is reliable; if some experts output "cheetah" and some output "tiger," the entropy value increases, triggering a manual review process.

[0086] Furthermore, embodiments of the present invention also perform multi-dimensional detection and analysis of erroneous samples, including: comparing the prediction results of each target expert with the true labels based on the prediction probability of each target expert model; and determining high-confidence erroneous sample images and high-disagreement erroneous sample images by setting confidence thresholds and stage entropy thresholds.

[0087] For example, incorrectly predicted samples are compared with the true labels, and a confidence threshold (e.g., 0.8) is set to distinguish between high and low confidence errors. Further analysis of expert disagreement is conducted, and the disagreement among experts is assessed based on entropy values. Entropy values ​​higher than a threshold (e.g., 1.5) are identified as high-disagreement errors.

[0088] Furthermore, embodiments of the present invention also perform adaptive dynamic optimization of expert models, including: fine-tuning each expert model and updating the expert pool based on the identified erroneous sample images.

[0089] For example, based on the identified error sample images, the expert model can be optimized in two ways: fine-tuning (using high-confidence error samples to strengthen specific category learning) and updating the expert pool (eliminating experts with consistently low weight / high error rates and adding new experts for weak categories). For instance, if "snake recognition" frequently makes mistakes, an expert model focused on crawling can be introduced to continuously improve the system's adaptability.

[0090] Based on the aforementioned method for constructing and validating large-scale wildlife datasets, embodiments of this invention validate the method. For example, firstly, data from 100 wildlife classes are extracted from all data sets, and each class is divided into a training set and a test set at a ratio of 1:9: a small amount of data is used for training, and the majority is used for inference. Then, the MOE system integrating six model architectures—CrossX, DCL, PMG, PIM, MPSA, and TransFG—is trained using the training set data. Finally, the MOE (Hybrid Expert Model) integrating the six architectures is tested on the test set. The validation results of this embodiment show that the prediction accuracy after integration is 94.29%, a significant improvement compared to the prediction accuracy of each individual model.

[0091] Through the embodiments of the present invention, fine-grained identification of wild animals in a large-scale wildlife dataset is achieved based on a hybrid expert model. By performing expert selection and prediction fusion on the test image set of the fine-grained identification task in the large-scale wildlife dataset, and adjusting the threshold for the number of activated experts and the confidence weight parameters, the accuracy of the target classification results is improved, thereby enhancing the accuracy and efficiency of the overall dataset construction.

[0092] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing and validating a large-scale wildlife dataset, characterized in that, The method includes: Construct a target dataset, which includes images of multiple types of wild animal samples; The test set in the target dataset is input into the hybrid expert model to obtain the target classification result, wherein the hybrid expert model is trained on the training set of the target dataset; Based on the analysis of the target classification results, the verification results of the test set in the target dataset are determined.

2. The method for constructing and validating a large-scale wildlife dataset according to claim 1, characterized in that, The target dataset to be constructed includes: Multi-source heterogeneous data is collected through a pre-set monitoring system, including raw image data and video data of various wild animal samples; The multi-source heterogeneous data is preprocessed and key frames are extracted to determine the target frame image data; Based on a preset automatic data annotation tool, the target frame image data and the original image data are annotated respectively to obtain a target dataset containing sample images of multiple types of wild animals.

3. The method for constructing and validating a large-scale wildlife dataset according to claim 1, characterized in that, The hybrid expert model is obtained by training on the training set in the target dataset, and the training process includes: Obtain the training set and perform feature preprocessing; The preprocessed training set is divided into multiple training subsets, and each expert model is trained using the cross-entropy loss function based on each training subset. After extracting multi-scale features from the images in the training set and performing multi-level feature fusion, a multilayer perceptron is used to map the fused features to the routing decision space. Based on the fused features, the activation weight of each expert model is calculated; Based on a preset threshold filtering strategy, a target expert model is determined among the multiple expert models, wherein the target expert model includes multiple models; The predicted probabilities from multiple target expert models are fused to obtain the target classification result.

4. The method for constructing and validating a large-scale wildlife dataset according to claim 3, characterized in that, The step of dividing the preprocessed training set into multiple training subsets and training each expert model using the cross-entropy loss function based on each training subset includes: Based on the area of ​​expertise of each expert model, the preprocessed training set is divided into multiple training subsets, wherein each training subset corresponds to one expert model. Based on each of the training subsets, cross-entropy loss is used as the primary objective, and each of the expert models is trained by introducing three types of regularization terms.

5. The method for constructing and validating a large-scale wildlife dataset according to claim 3, characterized in that, After extracting multi-scale features from the images in the training set and performing multi-level feature fusion, a multilayer perceptron is used to map the fused features to the routing decision space, including: The images in the training set are processed at multiple scales to extract global and local features respectively; Multi-level fusion is achieved through attention mechanisms or feature concatenation to obtain fused features; The fused features are mapped to the routing decision space using a multilayer perceptron to obtain a routing feature vector. The routing decision space includes multiple clustered image regions, and each clustered image region corresponds to one or more expert models.

6. The method for constructing and validating a large-scale wildlife dataset according to claim 3 or 5, characterized in that, The step of calculating the activation weight of each expert model based on the fused features includes: Based on the routing feature vector, the activation weights of each expert model are calculated using a four-layer cascaded structure.

7. The method for constructing and validating a large-scale wildlife dataset according to claim 3 or 5, characterized in that, The method of determining the target expert model among the multiple expert models based on a preset threshold filtering strategy includes: By setting fixed thresholds, dynamic thresholds, and relative thresholds, the target expert model is selected from the multiple expert models.

8. The method for constructing and validating a large-scale wildlife dataset according to claim 1 or 3, characterized in that, The analysis based on the target classification results, determining the validation result of the test set in the target dataset, includes: Based on the prediction probability of each target expert model, classification similarity is calculated to obtain similarity results and complete the quantitative analysis of prediction uncertainty.

9. The method for constructing and validating a large-scale wildlife dataset according to claim 8, characterized in that, The analysis based on the target classification results, determining the validation result of the test set in the target dataset, includes: Based on the predicted probability of each target expert model, the prediction results of each target expert are compared with the true labels; By setting confidence thresholds and stage entropy thresholds, high-confidence erroneous sample images and high-disagreement erroneous sample images are identified.

10. The method for constructing and validating a large-scale wildlife dataset according to claim 9, characterized in that, The analysis based on the target classification results, determining the validation result of the test set in the target dataset, includes: Based on the identified erroneous sample images, each expert model is fine-tuned and the expert pool is updated.

Images