Animal feature recognition method and system based on computer vision

Abstract

The invention discloses an animal feature recognition method and system based on computer vision, and relates to the technical field of animal image recognition, and the method comprises the steps: calculating and storing a region association degree matrix based on the historical monitoring data of a target region; obtaining a to-be-recognized image of the target sub-region at the current moment, and obtaining at least one preliminary recognition result and an initial confidence coefficient thereof; acquiring identification information of N associated regions having the highest association degree with the target sub-region; obtaining a confidence coefficient after interactive verification; the occurrence prior probability of the species under the current ecological condition is calculated; and calculating a final confidence coefficient, and outputting an enhanced recognition result containing the species tag and the confidence coefficient. The method has the advantages that by creatively fusing multi-dimensional information, the accuracy and robustness of wild animal recognition in complex open environments such as forests are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of animal image recognition technology, specifically to a method and system for animal feature recognition based on computer vision. Background Technology

[0002] In the field of forest ecological monitoring, accurate and efficient identification and monitoring of wildlife species are fundamental to assessing biodiversity, understanding population dynamics, and developing conservation strategies. Traditional methods mainly rely on manual field surveys, infrared camera traps, and a small number of fixed monitoring points. These methods not only consume significant manpower and resources but also suffer from discontinuous data collection and limited coverage, making it difficult to achieve large-scale, real-time dynamic perception. In recent years, with the popularization of computer vision technology, deep learning-based automatic identification models have been attempted to be applied to automatically analyze monitoring images or video streams, aiming to replace some manual labor and improve monitoring efficiency. However, purely visual recognition models are unstable in the complex real-world forest environment. Their performance is easily affected by factors such as changes in lighting, occlusion, shooting angle, and rare animal postures, leading to frequent misjudgments, missed judgments, or ambiguous confidence levels at single monitoring points, thus limiting their reliability and automation level.

[0003] Existing technologies typically process data from each monitoring point in isolation, lacking a systematic utilization of animal behavior patterns and the ecological environment. Specifically, existing methods have two main limitations: First, the identification process relies solely on image information from a single point and a single moment, failing to effectively utilize animal migration patterns in space and activity patterns in time, i.e., ignoring the inherent spatiotemporal correlations between observations from different geographical locations; second, they do not incorporate rich prior ecological knowledge, such as species habits, activity rhythms, and habitat preferences, as stable constraints into the identification decision, resulting in insufficient discrimination ability when faced with species with similar visual features or poor image quality. Therefore, there is an urgent need for an intelligent identification method that can integrate multi-source information and possess context-aware capabilities to improve the accuracy, robustness, and reliability of animal feature identification in complex real-world scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, a method and system for animal feature recognition based on computer vision are provided. This technical solution solves at least one of the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A computer vision-based method for animal feature recognition includes:

[0007] Based on historical monitoring data of the target area, calculate and store the regional correlation matrix W, which represents the spatiotemporal correlation of animal identification results between any two geographic sub-regions. The matrix element W(Ri, Rj, s) represents the correlation degree of species s between region Ri and region Rj, where Ri and Rj are the identifiers of geographic sub-regions, and s is the species identifier.

[0008] Obtain the image of the target sub-region R0 to be identified at the current time, and use the benchmark animal identification model to obtain at least one preliminary identification result and its initial confidence P0(s);

[0009] Query the region correlation matrix W to obtain the identification information of the N associated regions with the highest correlation to R0;

[0010] Based on the identification information of the associated region, the preliminary identification result of R0 is interactively verified to obtain the confidence level Pfused(s) after interactive verification.

[0011] Obtain the ecological factor vector corresponding to the current identified event. Based on a pre-defined ecological factor knowledge base, the prior probability Pprior(s|) of species s under current ecological conditions is calculated. );

[0012] Combining the post-interactive verification confidence Pfused(s) with the prior probability Pprior(s| ), calculate the final confidence level Pfinal(s), and output an enhanced identification result containing species label and confidence level based on the final confidence level Pfinal(s).

[0013] Preferably, the calculation and storage of the regional correlation matrix W, which characterizes the spatiotemporal correlation of animal identification results between any two geographical sub-regions based on historical monitoring data of the target area, specifically includes:

[0014] Acquire historical monitoring data, including images of multiple geographic sub-regions over a period of time and verified animal identification tags;

[0015] For species s, calculate the basic co-occurrence association C(Ri, Rj, s) between regions Ri and Rj:

[0016]

[0017] Where I(R, t, s) is an indicator function, which has a value of 1 if species s is identified in region R at time t, and a value of 0 otherwise. To prevent the smoothing factor from being divided by zero, the summation is performed across all times t in the historical time period to obtain the basic co-occurrence correlation C(Ri, Rj, s) between regions Ri and Rj.

[0018] For species s, calculate the temporal transition correlation Atemp(Ri, Rj, s) between regions Ri and Rj:

[0019]

[0020]

[0021] Indicates time lag The probability that species s will migrate from region Ri to region Rj. The time decay weighting function is... The preset time decay constant;

[0022] Calculate the comprehensive correlation between regions Ri and Rj based on the basic co-occurrence correlation and temporal transition correlation between Ri and Rj:

[0023] ,

[0024] in, , The weighting coefficients are preset, and α+β=1.

[0025] Preferably, the specific steps for calculating the confidence level after interactive verification are as follows:

[0026] From the regional correlation matrix W, select N regions Rk that make W(R0, Rk, s) greater than a preset threshold as the correlation regions, k=1, 2, ..., N;

[0027] Obtain the identification confidence Pk(s) of species s in each associated region Rk within a similar time period;

[0028] The weighted log-likelihood ratio fusion method was used to calculate the confidence Pfused(s) after cross-validation.

[0029]

[0030]

[0031]

[0032] For normalized correlation weights, The log-likelihood ratio represents the strength of evidence that the target is species s relative to the logarithmic form of the target not being species s.

[0033] Preferably, the step of obtaining the ecological factor array corresponding to the current identification event... Based on a pre-defined ecological factor knowledge base, the prior probability Pprior(s|) of species s under current ecological conditions is calculated. Specifically, it includes:

[0034] Constructing ecological factor vectors In this context, each dimension e represents an ecological factor, and m is the total number of ecological factors, including at least one of the following: season, day and night duration, temperature, rainfall, light intensity, altitude, slope, and distance from water source.

[0035] Based on the ecological factor vector, for the target sub-region R0 at the current time, generate its corresponding ecological factor array. , Let R0 be the target sub-region at the current moment in terms of ecological factors. The above value;

[0036] Based on the ecological factor vector, for a biological species s, generate its corresponding optimal ecological factor array. , For biological species s in the ecological factor dimension The upper range of required numerical values;

[0037] The ecological factor requirements for species s are established based on historical data, with an emphasis on... For biological species s in the dimension of ecological factors The focus of demand;

[0038] Calculate species s in a given ecological factor The prior probability of the following occurrence, Pprior(s| );

[0039]

[0040]

[0041] Let be the deviation of the j-th ecological factor, if Exceed The upper limit, then Depend on and The difference between the upper and lower limits is determined if Below The lower limit, then for and The difference between the lower limits is determined. Let be the degree of conformity of the j-th ecological factor.

[0042] Preferably, the combination of the post-interactive verification confidence Pfused(s) and the prior probability Pprior(s| The final confidence level Pfinal(s) is calculated, and the enhanced identification result containing species label and confidence level is output based on the final confidence level Pfinal(s). Specifically, this includes:

[0043] Based on the post-verification confidence Pfused(s) and the prior probability Pprior(s| The product of ) is used as the final confidence level Pfinal(s);

[0044] The species characteristics corresponding to the maximum final confidence score Pfinal(s) are selected as the final species labels;

[0045] The species label and final confidence score Pfinal(s) are output as the enhanced identification result.

[0046] Preferably, it also includes a similarity verification mechanism, which specifically includes:

[0047] If the difference between the maximum final confidence score Pfinal(s) and the remaining final confidence scores is less than a preset similarity threshold, a similarity verification mechanism is triggered. The similarity verification mechanism outputs the maximum final confidence score Pfinal(s) and species labels whose difference with the maximum final confidence score Pfinal(s) is less than the preset similarity threshold for manual verification.

[0048] Preferably, the method further includes adaptive updating of the regional correlation matrix, wherein the adaptive updating of the regional correlation matrix includes:

[0049] Based on the new verified identification results, the regional correlation matrix is ​​updated again according to the preset update cycle;

[0050] During the update, a time decay mechanism is adopted to apply time decay weights to historical correlation data and add them together with the correlation calculated from historical monitoring data in the most recent update cycle.

[0051] Furthermore, a computer vision-based animal feature recognition system is proposed, comprising:

[0052] The data storage module is used to store historical monitoring data of the target area and the regional correlation matrix W calculated based on the data. The matrix W represents the spatiotemporal correlation of animal identification results between any two geographical sub-regions. The matrix element W(Ri,Rj,s) represents the correlation degree of species s between region Ri and region Rj.

[0053] The image recognition module is used to acquire the image to be recognized in the target sub-region R0 at the current time, and call the benchmark animal recognition model to obtain at least one preliminary recognition result and its initial confidence P0(s);

[0054] The association information query module is used to query the regional association degree matrix W in the data storage module to obtain the identification information of the N associated regions with the highest association degree with the target sub-region R0.

[0055] The interactive verification module is used to perform interactive verification on the preliminary recognition result obtained by the image recognition module based on the recognition information of the associated region obtained by the associated information query module, and obtain the confidence level Pfused(s) after interactive verification.

[0056] The ecological prior calculation module is used to obtain the ecological factor vector F0 corresponding to the current identified event, and calculate the prior probability Pprior(s|F0) of the occurrence of species s under the current ecological conditions based on the preset ecological factor knowledge base.

[0057] The result fusion and output module is used to combine the post-interactive verification confidence Pfused(s) and the prior probability Pprior(s|F0) to calculate the final confidence Pfinal(s), and output an enhanced identification result containing species label and confidence based on the final confidence Pfinal(s).

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention significantly improves the accuracy and robustness of wildlife identification in complex open environments such as forests by creatively integrating multi-dimensional information. First, by constructing and utilizing a regional correlation matrix representing the spatiotemporal patterns of animal activity, collaborative interactive verification across monitoring points is achieved. This effectively utilizes the migration and co-occurrence patterns of animal populations, overcoming the limitations of single-point identification being susceptible to accidental interference, such as brief occlusion and fluctuations in image quality, thus significantly enhancing the spatiotemporal consistency of identification results. Second, by introducing prior probabilities of species occurrence calculated based on current ecological factors, domain ecological knowledge, such as species habitat preferences and activity rhythms, is used as prior constraints within a Bayesian framework. This is then fused with visual recognition results for decision-making, greatly improving the ability to distinguish species with similar visual features or those appearing in unsuitable habitats, effectively reducing false alarms. Finally, the entire solution constitutes a dynamically enhanced intelligent sensing system. It not only integrates visual, spatiotemporal, and ecological evidence through a weighted fusion strategy to output more credible enhanced results, but its built-in adaptive update mechanism also enables the system knowledge to be continuously optimized over time, thus possessing stronger environmental adaptability and long-term application value. This provides core technical support for achieving highly reliable and automated wide-area ecological intelligent monitoring. Attached Figure Description

[0060] Figure 1 This is a flowchart of the computer vision-based animal feature recognition method proposed in this scheme;

[0061] Figure 2 This is a flowchart of the method for calculating the regional correlation matrix W proposed in this scheme;

[0062] Figure 3 This is a flowchart of the method for calculating confidence after interactive verification proposed in this scheme;

[0063] Figure 4 The flowchart of the proposed method for outputting enhanced identification results that include species labels and confidence scores is shown below.

[0064] Figure 5 This is an architecture diagram of the electronic devices in this solution;

[0065] Figure 6 This is a schematic diagram of the computer-readable storage medium structure in this scheme.

[0066] The numbers on the map are:

[0067] 500 - Electronic device; 501 - Bus; 502 - CPU; 503 - ROM; 504 - RAM; 505 - Communication port; 506 - Input / output component; 507 - Hard disk; 508 - User interface; 600 - Computer-readable storage medium. Detailed Implementation

[0068] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0069] Reference Figure 1 As shown, a computer vision-based animal feature recognition method includes:

[0070] Based on historical monitoring data of the target area, a regional correlation matrix W is calculated and stored, representing the spatiotemporal correlation of animal identification results between any two geographic sub-regions. The matrix element W(Ri, Rj, s) represents the correlation degree of species s between regions Ri and Rj, where Ri and Rj are the identifiers of the geographic sub-regions, and s is the species identifier. By mining the animal activity patterns implicit in historical data, the correlation of different geographic sub-regions in the appearance of specific species is quantified from a spatiotemporal perspective. Matrix W essentially encodes the migration paths, activity ranges, and ecological connections between regions of species, providing data-driven, quantifiable "spatial memory" and "empirical knowledge" for subsequent collaborative identification. This forms the basis for achieving networked intelligent perception rather than isolated, single-point perception.

[0071] The image of the target sub-region R0 at the current time is acquired. A baseline animal recognition model is used to obtain at least one preliminary recognition result and its initial confidence level P0(s). The baseline animal recognition model is responsible for extracting visual features from the original image and making a preliminary judgment, outputting possible species labels and their confidence levels. The image recognition and classification model is a mature existing technology in this field, and will not be elaborated upon here. P0(s) reflects the initial judgment strength of the model based solely on visual evidence from the current single image. This is the starting point of the entire recognition process, but its results may have uncertainties or errors due to factors such as image quality, occlusion, and rare poses, providing processing objects for subsequent enhancement and correction.

[0072] By querying the regional correlation matrix W, the identification information of the N most closely related regions to R0 is obtained, achieving a crucial leap from "single-point observation" to "contextual awareness." Using the pre-constructed correlation matrix W, the system intelligently locates the N other regions most closely associated with animal activity at the current monitoring point R0 historically. The identification information of these related regions, such as the species and confidence levels identified within similar time periods, constitutes strong spatiotemporal contextual evidence for the current identification task. This is equivalent to not only looking at the current photo of R0 when judging the animal at R0, but also referring to what its "neighbors" or "ecologically connected areas" have recently seen, greatly enriching the decision-making basis.

[0073] Based on the identification information of the associated region, the preliminary identification result of R0 is interactively verified to obtain the confidence Pfused(s) after interactive verification. The spatiotemporal context evidence from the associated region is integrated with the preliminary visual evidence of R0 itself. The output Pfused(s) of this step is a confidence adjusted by spatial co-verification, which weakens the random error in single-point observation, strengthens the identification result consistent with the spatiotemporal behavior pattern, and thus improves the robustness and consistency of the result.

[0074] The process involves obtaining the ecological factor vector F0 corresponding to the current identified event, and calculating the prior probability Pprior(s|F0) of species s under the current ecological conditions based on a pre-defined ecological factor knowledge base. Domain ecology knowledge is introduced as a prior constraint. The ecological factor vector F0 (such as time, climate, and topography) characterizes the specific environmental context of the identified event. By querying the ecological factor knowledge base (which stores habitat preferences, activity rhythms, etc. for each species), the theoretical probability Pprior(s|F0) of species s' occurrence under a given environment F0 can be calculated. This is equivalent to judging "what might be here" based on the environment before judging "what is this" based on image evidence, providing a prior probability distribution that conforms to ecological laws for the identification process. This helps correct results that are extremely unlikely to occur under specific environments but might be misjudged by the visual model.

[0075] Combining the post-validation confidence Pfused(s) with the prior probability Pprior(s| The system calculates a final confidence level, Pfinal(s), and outputs an enhanced recognition result containing species labels and the confidence level based on Pfinal(s). This combines spatially co-validated visual evidence (Pfused(s)) with ecological prior knowledge (Pprior(s|F0)), typically using Bayesian methods such as multiplication, to calculate the final confidence level Pfinal(s) that integrates all available information. This final confidence level considers visual features, spatiotemporal behavioral patterns, and ecological suitability, making it far more reliable than a single initial visual recognition result. The system ultimately outputs species labels based on this confidence level, forming an enhanced recognition result that significantly improves accuracy, robustness, and interpretability.

[0076] Reference Figure 2 As shown, based on historical monitoring data of the target area, the regional correlation matrix W, which represents the spatiotemporal correlation of animal identification results between any two geographical sub-regions, is calculated and stored. Specifically, it includes:

[0077] Acquire historical monitoring data, including images of multiple geographic sub-regions over a period of time and verified animal identification tags;

[0078] For species s, calculate the basic co-occurrence association C(Ri, Rj, s) between regions Ri and Rj:

[0079]

[0080] Where I(R, t, s) is an indicator function, which has a value of 1 if species s is identified in region R at time t, and a value of 0 otherwise. To prevent the smoothing factor from being divided by zero, the summation is performed across all times t in the historical time period to obtain the basic co-occurrence correlation C(Ri, Rj, s) between regions Ri and Rj.

[0081] For species s, calculate the temporal transition correlation Atemp(Ri, Rj, s) between regions Ri and Rj:

[0082]

[0083]

[0084] Indicates time lag The probability that species s will migrate from region Ri to region Rj. The time decay weighting function is... The preset time decay constant;

[0085] Calculate the comprehensive correlation between regions Ri and Rj based on the basic co-occurrence correlation and temporal transition correlation between Ri and Rj:

[0086] ,

[0087] in, , The weighting coefficients are preset, and α+β=1.

[0088] In this scheme, the determination of the regional correlation matrix W includes two core, complementary dimensions, aiming to refine the spatiotemporal patterns of animal activity. The basic co-occurrence correlation C(Ri,Rj,s) quantifies the static, synchronous ecological connections between animals by statistically analyzing the frequency of species s co-occurring in two regions at the same time and normalizing the data. This includes co-occurrence behaviors caused by shared territory or water sources. The temporal transfer correlation Atemp(Ri,Rj,s) is more dynamic. It captures and quantifies the dynamic patterns of animal migration and roaming between regions by finding the maximum transfer probability of species s from region Ri to region Rj. This involves considering the transfer probability under different time lags τ, taking the maximum value, and applying a time decay weight w(τ). This reflects temporal causal relationships such as migration paths and activity routes. Finally, by using preset weight coefficients α and β, these two dimensions are linearly fused into a comprehensive correlation degree W, so that the matrix can not only indicate whether there is a correlation between regions, but also distinguish and characterize whether the correlation is mainly due to the habit of "appearing at the same time" or the migration behavior of "arriving in sequence", thus constructing a comprehensive and detailed spatiotemporal correlation model of animal activities.

[0089] This system achieves a refined and computable representation of animal activity patterns, upgrading from simple "proximity" to a quantitative correlation degree that includes "co-occurrence strength" and "migration tendency," enabling subsequent steps to utilize richer and more accurate contextual information. Secondly, it distinguishes and integrates co-occurrence and migration patterns, allowing the system to utilize both population distribution information and individual or group movement patterns, resulting in a more comprehensive and ecologically accurate assessment of correlations. Finally, the calculation of temporal migration correlation degree incorporates a time-decay weight function and the selection of the maximum migration probability, effectively highlighting regular and periodic migration patterns while reducing the interference of accidental and random migration events. This enhances the stability and indicative power of the correlation matrix, providing a highly reliable and discriminative weighting basis for subsequent interactive verification steps.

[0090] Reference Figure 3 As shown, the specific steps for calculating confidence after interactive verification are as follows:

[0091] From the regional correlation matrix W, select N regions Rk that make W(R0, Rk, s) greater than a preset threshold as the correlation regions, k=1, 2, ..., N;

[0092] Obtain the identification confidence Pk(s) of species s in each associated region Rk within a similar time period;

[0093] The weighted log-likelihood ratio fusion method was used to calculate the confidence Pfused(s) after cross-validation.

[0094]

[0095]

[0096]

[0097] For normalized correlation weights, The log-likelihood ratio represents the strength of evidence that the target is species s relative to the logarithmic form of the target not being species s.

[0098] This scheme integrates preliminary visual evidence from the target region R0 itself with spatiotemporal collaborative evidence from N related regions using a scientific probabilistic fusion method to form an enhanced judgment. Specifically, it involves several key steps: First, threshold screening ensures that only regions with sufficiently high correlation are adopted as valid correlation evidence, preventing interference from noise information from weakly correlated or even irrelevant regions. Second, weight assignment normalizes the elements W(R0,Rk,s) of the region correlation matrix to obtain... This ensures that the contribution of evidence from each associated region to the final decision is proportional to its association strength with the current target region, giving opinions from highly associated regions greater "voice." Finally, the weighted log-likelihood ratio fusion is the essence of the entire process. The confidence level P(s) of each region regarding species s is transformed into the log-likelihood ratio (LLR) of evidence supporting "the target is species s" relative to "the target is not species s". Then, the evidence from the current region is linearly superimposed with the weighted evidence from all associated regions, and finally, a logistic function is used to transform the combined total evidence strength back into a standardized probability value Pfused(s). This method has a solid theoretical foundation in probability and information theory, effectively fusing uncertain information from multiple sources to obtain consistent and robust posterior probability estimates. It achieves the scientific fusion of heterogeneous evidence from multiple sources, combining the raw output of a single image classification model with spatiotemporally correlated identification results from other monitoring points through a rigorous probabilistic graphical model framework (i.e., the sum of log-likelihood ratios corresponds to the accumulation of independent or related evidence). This ensures that the final confidence decision is based on a broader and more reliable evidence network, significantly improving the reliability of the decision. Furthermore, the allocation of evidence weights is scientific and well-founded, directly using quantifiable spatiotemporal correlation as the fusion weight, directly linking the "credibility of evidence" to the "relevance of evidence sources." This is more in line with actual physical laws than simple averaging or artificially assigning weights, significantly improving the rationality and accuracy of the fusion results. Third, the final output confidence level has better stability and discriminative power. After multi-regional collaborative verification, those misidentifications that contradict spatiotemporal behavior patterns will be weakened because their evidence is not supported in the associated regions, while those species that appear in multiple associated regions will be strengthened by evidence. This makes Pfused(s) more resistant to single-point random errors than the initial P0(s), thus laying a high-quality and high-confidence input foundation for further integration with ecological prior knowledge.

[0099] Specifically, obtain the array of ecological factors corresponding to the current identified event. Based on a pre-defined ecological factor knowledge base, the prior probability Pprior(s|) of species s under current ecological conditions is calculated. Specifically, it includes:

[0100] Constructing ecological factor vectors In this context, each dimension e represents an ecological factor, and m is the total number of ecological factors, including at least one of the following: season, day and night duration, temperature, rainfall, light intensity, altitude, slope, and distance from water source.

[0101] Based on the ecological factor vector, for the target sub-region R0 at the current time, generate its corresponding ecological factor array. , Let R0 be the target sub-region at the current moment in terms of ecological factors. The above value;

[0102] Based on the ecological factor vector, for a biological species s, generate its corresponding optimal ecological factor array. , For biological species s in the ecological factor dimension The upper range of required numerical values;

[0103] The ecological factor requirements for species s are established based on historical data, with an emphasis on... For biological species s in the dimension of ecological factors The focus of demand;

[0104] Calculate species s in a given ecological factor The prior probability of the following occurrence, Pprior(s| );

[0105]

[0106]

[0107] Let be the deviation of the j-th ecological factor, if Exceed The upper limit, then Depend on and The difference between the upper and lower limits is determined if Below The lower limit, then for and The difference between the lower limits is determined. Let be the degree of conformity of the j-th ecological factor.

[0108] This scheme constructs a quantitative framework incorporating multi-dimensional ecological factors. The ecological factor array F0 precisely characterizes the specific environmental state at the time of the identified event. The optimal ecological factor array and the demand focus together constitute the "ecological profile" of species s. The former defines the ideal range of each environmental factor preferred by the species, while the latter quantifies the restrictive strength of each factor on the species' distribution. For example, for some species, temperature may be more important than slope. The core formula for probability calculation... This approach embodies a weighted conformity assessment: for each ecological factor, if the measured value falls within the preferred range of species *s*, then it is considered fully conforming to that factor; if it deviates, the conformity is reduced according to the degree of deviation. Finally, the weighted conformity of all factors is normalized to obtain a comprehensive prior probability between 0 and 1. This value intuitively reflects "how likely it is to see species *s* in the current environment F0," completely independent of image visual information and purely based on ecological principles. This prior probability calculation method brings crucial domain knowledge enhancement and error correction capabilities to the system. First, it achieves the quantification and procedural embedding of ecological domain knowledge, transforming expert experience, such as species habitat preferences and activity rhythms, into structured data and algorithms, enabling the computer to "understand" and apply ecological principles for reasoning, significantly improving the system's intelligence level. Second, it provides a powerful and interpretable ecological rationality check for visual recognition results. When the visual model makes a suspected misjudgment due to image blurring, occlusion, or other reasons—for example, identifying a warm-loving reptile on a cold winter night—the extremely low prior probability calculated by this method effectively challenges and suppresses this result, thus systematically reducing the occurrence of such "unreasonable" false alarms. Furthermore, its weighted evaluation and exponential decay model design are both reasonable and flexible. The weighting mechanism ensures that different ecological factors participate in decision-making with the correct influence; the exponential decay model can significantly distinguish between slight and severe deviations, making the prior probability respond smoothly and sensitively to continuous changes in environmental conditions. Finally, the output Pprior(s|F0) of this step, as a robust ecological prior, is fused with visual and spatiotemporal evidence, fundamentally enhancing the reliability and interpretability of the entire system's decision-making.

[0109] Reference Figure 4 As shown, combining the post-validation confidence Pfused(s) and the prior probability Pprior(s| The final confidence level Pfinal(s) is calculated, and the enhanced identification result based on the final confidence level Pfinal(s) includes species labels and confidence levels. Specifically, it includes:

[0110] Based on the post-validation confidence Pfused(s) and prior probability Pprior(s| The product of ) is used as the final confidence level Pfinal(s);

[0111] The species characteristics corresponding to the maximum final confidence score Pfinal(s) are selected as the final species labels;

[0112] The species label and final confidence score Pfinal(s) are output as the enhanced identification result.

[0113] The final confidence level, Pfinal(s), is obtained by directly multiplying the cross-validated confidence level Pfused(s) with the ecological prior probability Pprior(s|F0). This operation has a clear explanation in probability theory: it can be understood as obtaining "likelihood" evidence (Pfused(s)) reflecting spatiotemporal consistency, and then incorporating "prior" knowledge (Pprior(s|F0)) reflecting environmental suitability, thus obtaining a comprehensive "posterior" confidence level. This product-based fusion rule requires that the final accepted identification result must be supported by both visual / spatiotemporal evidence and ecological evidence; neither can be lacking. Subsequently, the species label that maximizes the final confidence level is selected as the output, which is an optimal decision rule following the maximum posterior probability principle. The entire process, from probability fusion to decision output, forms a logical closed loop, ensuring that the final conclusion given by the system is an optimal and interpretable judgment that integrates all available information, bringing decisive performance improvements and application value to the system. First, a hierarchical and complementary strong fusion is achieved. The product rule requires both visual-spatiotemporal evidence and ecological priors to be "high" in order to generate a high final confidence score. This enables the system to effectively correct identification results that seem reasonable on a single evidence channel but are unreasonable when considered comprehensively, greatly enhancing the rigor and reliability of decision-making. Second, the mechanism outputs a unified, quantitative final confidence score Pfinal(s). This value integrates all information from image features, spatiotemporal patterns, and the ecological environment. Compared to a single initial identification confidence score, its numerical meaning is richer, its discriminative power is stronger, and its credibility is higher, providing users (whether researchers or conservation managers) with a directly reliable basis for decision-making. Finally, the entire process, from perception, verification, prior constraints to final decision-making, constitutes a complete and interpretable computational reasoning chain. This not only significantly improves the accuracy and robustness of identification but also makes the system's judgment process transparent and traceable, greatly enhancing its application value in rigorous ecological monitoring and scientific research.

[0114] In some preferred embodiments, the above method further includes a similarity verification mechanism, which specifically includes:

[0115] If the difference between the maximum final confidence score Pfinal(s) and the remaining final confidence scores is less than a preset similarity threshold, a similarity verification mechanism is triggered. The similarity verification mechanism outputs the maximum final confidence score Pfinal(s) and the species labels whose difference with the maximum final confidence score Pfinal(s) is less than the preset similarity threshold for manual verification.

[0116] A "similarity verification" mechanism is introduced as an important supplement to the aforementioned fully automated decision-making process and as a final quality control measure. Its core design logic lies in identifying and handling special cases where there is high uncertainty within the system. When the difference between the maximum value of the final confidence score Pfinal(s) and other candidate values ​​is less than a preset similarity threshold, it indicates that even after integrating visual, spatiotemporal, and ecological evidence, the system still cannot make a high-confidence distinction between two or more species options. This typically occurs between species with highly similar visual features and overlapping ecological niches. This mechanism does not force the system to make potentially erroneous "hard decisions," but instead triggers a special output containing multiple candidate results, leaving the final decision to domain experts for manual verification. This design embodies the idea of ​​"human-machine collaboration and complementary advantages." While fully leveraging the algorithm's ability to process massive amounts of data, it respects the irreplaceable professional judgment of human experts in complex and ambiguous situations, effectively reducing the risk of misjudgment of highly similar species. By setting an "uncertainty threshold," it proactively identifies potential decision-making ambiguities, avoiding the algorithm forcibly outputting a potentially erroneous single result in a "dilemma," thus eliminating a specific type of error from the system design. Secondly, it significantly improves the reliability and credibility of the overall identification system. This mechanism clarifies the algorithm's capabilities, making the most difficult uncertainties transparent and leaving them to human judgment. This ensures the system's output has high confidence in scenarios where automatic judgment is possible, while remaining honest and prudent in ambiguous situations, thus gaining greater user trust. Finally, it greatly optimizes the workflow and efficiency of manual review. Compared to having experts review all results, this mechanism intelligently and precisely selects a very small number of truly challenging cases requiring human intervention for focused review. This allows valuable human resources to be concentrated on complex problems that machines cannot solve, achieving efficient collaboration between automated processes and expert wisdom, and improving the overall efficiency and accuracy of monitoring work.

[0117] In some other preferred embodiments, the method further includes adaptive updating of the regional correlation matrix, which includes:

[0118] Based on the new verified identification results, the regional correlation matrix is ​​updated again according to the preset update cycle;

[0119] During the update, a time decay mechanism is adopted to apply time decay weights to historical correlation data and add them together with the correlation calculated from historical monitoring data in the most recent update cycle.

[0120] This embodiment introduces an adaptive update mechanism for the regional correlation matrix W to ensure the system maintains high accuracy and applicability over the long term. Its core objective is to enable the knowledge base encoding animal spatiotemporal activity patterns—the regional correlation matrix W—to dynamically evolve and keep pace with the times. Specifically, this mechanism comprises two collaborative parts: First, periodic recalculation. The system, according to a preset cycle (e.g., weekly or monthly), inputs newly accumulated and verified recognition results as training data into the correlation calculation process, thereby generating updated correlation cognition based on the latest data. This ensures the system can learn and reflect new activity patterns and regional correlations arising from seasonal changes, environmental shifts, or population dynamics. Second, a time-decay weighting fusion strategy is introduced. When reusing historical correlation data, it is not simply overwritten or merged with equal weights, but rather an older data is given a weight that decreases over time. This means that older historical patterns have a smaller impact on the current matrix, while more recent observations have a greater influence. This design allows matrix W to smoothly "forget" outdated and potentially no longer applicable patterns, while reinforcing newly emerging patterns. This makes the system highly adaptable and robust to both long-term changes and short-term fluctuations in animal behavior patterns. The entire mechanism forms a closed loop of "perception-learning-application-re-optimization," transforming the entire recognition system from a static, historical snapshot-based tool into an intelligent perception system capable of continuous learning and self-improvement.

[0121] Furthermore, a computer vision-based animal feature recognition system is proposed, including:

[0122] The data storage module is used to store historical monitoring data of the target area and the regional correlation matrix W calculated based on the data. The matrix W represents the spatiotemporal correlation of animal identification results between any two geographic sub-regions. The matrix element W(Ri,Rj,s) represents the correlation degree of species s between region Ri and region Rj.

[0123] The image recognition module is used to acquire the image to be recognized of the target sub-region R0 at the current time, and call the benchmark animal recognition model to obtain at least one preliminary recognition result and its initial confidence P0(s);

[0124] The association information query module is used to query the regional association degree matrix W in the data storage module to obtain the identification information of the N associated regions with the highest association degree with the target sub-region R0.

[0125] The interactive verification module is used to interactively verify the preliminary recognition results obtained by the image recognition module based on the recognition information of the associated region obtained by the associated information query module, and obtain the confidence Pfused(s) after interactive verification.

[0126] The ecological prior calculation module is used to obtain the ecological factor vector F0 corresponding to the current identified event, and calculate the prior probability Pprior(s|F0) of the occurrence of species s under the current ecological conditions based on the preset ecological factor knowledge base.

[0127] The results fusion and output module is used to combine the confidence Pfused(s) after interactive verification with the prior probability Pprior(s|F0) to calculate the final confidence Pfinal(s), and output an enhanced identification result containing species label and confidence based on the final confidence Pfinal(s).

[0128] Furthermore, the method according to the embodiments of this application can also be achieved by means of... Figure 5 The architecture of the electronic device shown is used to implement this. For example... Figure 5 As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, ROM 503, RAM 504, a communication port 505 connected to a network, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as ROM 503 or hard disk 507, may store the computer vision-based animal feature recognition method provided in this application. The electronic device 500 may also include a user interface 508. Of course, Figure 5 The architecture shown is merely exemplary and can be omitted as needed when implementing different devices. Figure 5 One or more components in the illustrated electronic device.

[0129] Figure 6 This is a schematic diagram of a computer-readable storage medium structure provided in one embodiment of this application. Figure 6 The diagram illustrates a computer-readable storage medium 600 according to one embodiment of this application. The computer-readable storage medium 600 stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a computer vision-based animal feature recognition method according to an embodiment of this application, as described with reference to the above figures. The computer-readable storage medium 600 includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0130] In summary, the advantages of this invention are as follows: By creatively integrating multi-dimensional information, it significantly improves the accuracy and robustness of wildlife identification in complex open environments such as forests. First, by constructing and utilizing a regional correlation matrix representing the spatiotemporal patterns of animal activity, collaborative interactive verification across monitoring points is achieved. This effectively utilizes the migration and co-occurrence patterns of animal populations, overcoming the limitations of single-point identification being susceptible to accidental interference, such as brief occlusion and fluctuations in image quality, thus significantly enhancing the spatiotemporal consistency of the identification results. Second, by introducing prior probabilities of species occurrence calculated based on current ecological factors, domain ecological knowledge, such as species habitat preferences and activity rhythms, is used as prior constraints within a Bayesian framework. This is then fused with visual recognition results for decision-making, greatly improving the ability to distinguish species with similar visual characteristics or those appearing in unsuitable habitats, effectively reducing false alarms. Finally, the entire solution constitutes a dynamically enhanced intelligent sensing system. It not only integrates visual, spatiotemporal, and ecological evidence through a weighted fusion strategy to output more credible enhanced results, but its built-in adaptive update mechanism also enables the system knowledge to be continuously optimized over time, thus possessing stronger environmental adaptability and long-term application value. This provides core technical support for achieving highly reliable and automated wide-area ecological intelligent monitoring.

[0131] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for animal feature recognition based on computer vision, characterized in that, include: Based on historical monitoring data of the target area, calculate and store the regional correlation matrix W, which represents the spatiotemporal correlation of animal identification results between any two geographic sub-regions. The matrix element W(Ri, Rj, s) represents the correlation degree of species s between region Ri and region Rj, where Ri and Rj are the identifiers of geographic sub-regions, and s is the species identifier. Obtain the image of the target sub-region R0 to be identified at the current time, and use the benchmark animal identification model to obtain at least one preliminary identification result and its initial confidence P0(s); Query the region correlation matrix W to obtain the identification information of the N associated regions with the highest correlation to R0; Based on the identification information of the associated region, the preliminary identification result of R0 is interactively verified to obtain the confidence level Pfused(s) after interactive verification. Obtain the ecological factor vector corresponding to the current identified event. Based on a pre-defined ecological factor knowledge base, the prior probability Pprior(s|) of species s under current ecological conditions is calculated. ); Combining the post-interactive verification confidence Pfused(s) with the prior probability Pprior(s| ), calculate the final confidence level Pfinal(s), and output an enhanced identification result containing species label and confidence level based on the final confidence level Pfinal(s).

2. The method for animal feature recognition based on computer vision according to claim 1, characterized in that, The calculation and storage of the regional correlation matrix W, which represents the spatiotemporal correlation of animal identification results between any two geographic sub-regions based on historical monitoring data of the target area, specifically includes: Acquire historical monitoring data, including images of multiple geographic sub-regions over a period of time and verified animal identification tags; For species s, calculate the basic co-occurrence association C(Ri, Rj, s) between regions Ri and Rj: Where I(R, t, s) is an indicator function, which has a value of 1 if species s is identified in region R at time t, and a value of 0 otherwise. To prevent the smoothing factor from being divided by zero, the summation is performed across all times t in the historical time period to obtain the basic co-occurrence correlation C(Ri, Rj, s) between regions Ri and Rj. For species s, calculate the temporal transition correlation Atemp(Ri, Rj, s) between regions Ri and Rj: Indicates time lag The probability that species s will migrate from region Ri to region Rj. The time decay weighting function is... The preset time decay constant; Calculate the comprehensive correlation between regions Ri and Rj based on the basic co-occurrence correlation and temporal transition correlation between Ri and Rj: ， in, , The weighting coefficients are preset, and α+β=1.

3. The method for animal feature recognition based on computer vision according to claim 2, characterized in that, The specific steps for calculating the confidence level after interactive verification are as follows: From the regional correlation matrix W, select N regions Rk that make W(R0, Rk, s) greater than a preset threshold as the correlation regions, k=1, 2, ..., N; Obtain the identification confidence Pk(s) of species s in each associated region Rk within a similar time period; The weighted log-likelihood ratio fusion method was used to calculate the confidence Pfused(s) after cross-validation. For normalized correlation weights, The log-likelihood ratio represents the strength of evidence that the target is species s relative to the logarithmic form of the target not being species s.

4. The method for animal feature recognition based on computer vision according to claim 3, characterized in that, The step of obtaining the ecological factor array corresponding to the current identification event Based on a pre-defined ecological factor knowledge base, the prior probability Pprior(s|) of species s under current ecological conditions is calculated. Specifically, it includes: Constructing ecological factor vectors In this context, each dimension e represents an ecological factor, and m is the total number of ecological factors, including at least one of the following: season, day and night duration, temperature, rainfall, light intensity, altitude, slope, and distance from water source. Based on the ecological factor vector, for the target sub-region R0 at the current time, generate its corresponding ecological factor array. , Let R0 be the target sub-region at the current moment in terms of ecological factors. The above value; Based on the ecological factor vector, for a biological species s, generate its corresponding optimal ecological factor array. , For biological species s in the ecological factor dimension The upper range of required numerical values; The ecological factor requirements for species s are established based on historical data, with an emphasis on... For biological species s in the dimension of ecological factors The focus of demand; Calculate species s in a given ecological factor The prior probability of the following occurrence, Pprior(s| ); Let be the deviation of the j-th ecological factor, if Exceed The upper limit, then Depend on and The difference between the upper and lower limits is determined if Below The lower limit, then for and The difference between the lower limits is determined. Let be the degree of conformity of the j-th ecological factor.

5. The method for animal feature recognition based on computer vision according to claim 4, characterized in that, The combination of the post-verification confidence Pfused(s) and the prior probability Pprior(s| The final confidence level Pfinal(s) is calculated, and the enhanced identification result containing species label and confidence level is output based on the final confidence level Pfinal(s). Specifically, this includes: Based on the post-verification confidence Pfused(s) and the prior probability Pprior(s| The product of ) is used as the final confidence level Pfinal(s); The species characteristics corresponding to the maximum final confidence score Pfinal(s) are selected as the final species labels; The species label and final confidence score Pfinal(s) are output as the enhanced identification result.

6. The method for animal feature recognition based on computer vision according to claim 5, characterized in that, It also includes a similarity verification mechanism, which specifically includes: If the difference between the maximum final confidence score Pfinal(s) and the remaining final confidence scores is less than a preset similarity threshold, a similarity verification mechanism is triggered. The similarity verification mechanism outputs the maximum final confidence score Pfinal(s) and species labels whose difference with the maximum final confidence score Pfinal(s) is less than the preset similarity threshold for manual verification.

7. The method for animal feature recognition based on computer vision according to any one of claims 1-6, characterized in that, It also includes adaptive updating of the regional correlation matrix, which includes: Based on the new verified identification results, the regional correlation matrix is ​​updated again according to the preset update cycle; During the update, a time decay mechanism is adopted to apply time decay weights to historical correlation data and add them together with the correlation calculated from historical monitoring data in the most recent update cycle.

8. An animal feature recognition system based on computer vision, characterized in that, include: The data storage module is used to store historical monitoring data of the target area and the regional correlation matrix W calculated based on the data. The matrix W represents the spatiotemporal correlation of animal identification results between any two geographical sub-regions. The matrix element W(Ri,Rj,s) represents the correlation degree of species s between region Ri and region Rj. The image recognition module is used to acquire the image to be recognized of the target sub-region R0 at the current time, and call the benchmark animal recognition model to obtain at least one preliminary recognition result and its initial confidence P0(s); The association information query module is used to query the regional association degree matrix W in the data storage module to obtain the identification information of the N associated regions with the highest association degree with the target sub-region R0. The interactive verification module is used to perform interactive verification on the preliminary recognition result obtained by the image recognition module based on the recognition information of the associated region obtained by the associated information query module, and obtain the confidence level Pfused(s) after interactive verification. The ecological prior calculation module is used to obtain the ecological factor vector F0 corresponding to the current identified event, and calculate the prior probability Pprior(s|F0) of the occurrence of species s under the current ecological conditions based on the preset ecological factor knowledge base. The result fusion and output module is used to combine the post-interactive verification confidence Pfused(s) and the prior probability Pprior(s|F0) to calculate the final confidence Pfinal(s), and output an enhanced identification result containing species label and confidence based on the final confidence Pfinal(s).

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform a computer vision-based animal feature recognition method as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the computer vision-based animal feature recognition method according to any one of claims 1-7.

Images