A method and apparatus for processing images for wildlife monitoring

Abstract

This invention provides a method and apparatus for processing images used in wildlife monitoring. The method includes: acquiring wildlife image data; the wildlife image data is obtained through an infrared camera; filtering the wildlife image data to obtain valid image data; obtaining wildlife identification results based on the valid image data and a trained wildlife identification model; the wildlife identification model is trained on a preset network model using historical wildlife images; visualizing the wildlife identification results to obtain visualized data; and obtaining wildlife monitoring results based on the valid image data, the wildlife identification results, the visualized data, and historical monitoring data. This invention can improve the accuracy of wildlife identification and monitoring.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and also to a method and apparatus for processing images for wildlife monitoring. Background Technology

[0002] Continuous research on biodiversity resources in nature reserves helps to understand the basic status and dynamic changes of biodiversity within these reserves, and facilitates the development of scientific conservation and management plans. Systematic and long-term monitoring of rare and endangered wildlife is not only fundamental to achieving wildlife diversity protection, but also a crucial indicator for evaluating the management and conservation efforts of protected areas. However, how to efficiently manage and monitor these areas to ensure the effective protection of rare and endangered wildlife remains a pressing issue. Traditional wildlife monitoring relies on manual field surveys, which are costly and inefficient. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and apparatus for processing images for wildlife monitoring, so as to improve the efficiency of wildlife monitoring and reduce monitoring costs.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect of the present invention provides a method for processing images used in wildlife monitoring, comprising: Acquire wildlife image data; the wildlife image data is obtained through infrared cameras. The wildlife image data is filtered to obtain valid image data; Based on the effective image data and the trained wildlife recognition model, the wildlife recognition result is obtained; the wildlife recognition model is obtained by training a preset network model based on collected historical wildlife images; The wildlife identification results are visualized to obtain visualized data; Based on the wildlife identification results, the visualization data, and the historical monitoring data, the wildlife monitoring results are obtained.

[0005] Optionally, the wildlife image data is filtered to obtain valid image data, including: Based on the grayscale variance and a preset variance threshold of the wildlife image data, the wildlife image data is filtered once to obtain the first image data. Based on the pixel difference of the first image data and a preset difference threshold, the first image data is filtered a second time to obtain the second image data; Based on the target contour data of the second image data, the second image data is filtered for effective images to obtain effective image data.

[0006] Optionally, based on the effective image data and the trained wildlife recognition model, the wildlife recognition result is obtained, including: The valid image data is preprocessed to obtain preprocessed image data; The preprocessed image data is input into the first processing layer of the wildlife recognition model to obtain the first processing result; The first processing result is input into the second processing layer of the wildlife recognition model to obtain the second processing result; The second processing result is input into the third processing layer of the wildlife recognition model to obtain the wildlife recognition result.

[0007] Optionally, the first processing result is input into the second processing layer of the wildlife recognition model to obtain a second processing result, including: pass The second processing result is obtained, in which, The second processing result is the probability predicted by the model that the current target belongs to the f-th class of wild animal species. This represents the deep feature output value corresponding to the f-th species in the first processing result, where F is the total number of all monitored wild animal species in the protected area. This is the deep feature output value corresponding to the g-th species in the first processing result.

[0008] Optionally, the training process of the wildlife recognition model includes: Acquire historical images of wild animals; The historical wildlife images are subjected to data annotation processing to obtain annotated image data; Based on the quality assessment results of the labeled image data, the sample data is determined; The sample data is augmented to obtain training samples; The preset network model is trained according to the preset training strategy and the training samples to obtain the wildlife recognition model.

[0009] Optionally, visualization processing is performed based on the valid image data and the wildlife identification results to obtain visualized data, including: The effective image data is subjected to image enhancement processing to obtain enhanced image data; The enhanced image data and the wildlife identification results are visualized according to a preset visualization scheme to obtain visualized data.

[0010] Optionally, based on the wildlife identification results, the visualization data, and historical monitoring data, wildlife monitoring results are obtained, including: Obtain preset early warning conditions; Based on the preset warning conditions and the wildlife identification results, the warning data is determined; Statistical data were obtained based on the wildlife identification results and historical monitoring data. Based on the early warning data, the visualization data, and the statistical data, the results of wildlife monitoring are obtained.

[0011] A second aspect of the present invention provides a wildlife monitoring image processing apparatus, comprising: The acquisition module is used to acquire wildlife image data; the wildlife image data is obtained through infrared camera acquisition. The processing module is used to perform effective image filtering on the wildlife image data to obtain effective image data; to obtain wildlife identification results based on the effective image data and a trained wildlife identification model; the wildlife identification model is obtained by training a preset network model based on collected historical wildlife images; to perform visualization processing on the effective image data and the wildlife identification results to obtain visualized data; and to obtain wildlife monitoring results based on the wildlife identification results, the visualized data, and historical monitoring data.

[0012] A third aspect of the present invention provides a computing device, comprising: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described in the first aspect.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in the first aspect.

[0014] The above-described solution of the present invention has at least the following beneficial effects: The above-described solution of the present invention acquires wildlife image data through an infrared camera, performs effective image filtering on the wildlife image data to obtain effective image data, and obtains wildlife identification results based on the effective image data and a trained wildlife identification model. The wildlife identification model is trained on a preset network model based on historical wildlife images. The effective image data and the wildlife identification results are then visualized to obtain visualized data. Finally, wildlife monitoring results are obtained based on the wildlife identification results, the visualized data, and historical monitoring data. This not only enables multi-species, high-precision identification but also helps improve the accuracy of wildlife monitoring results. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the wildlife monitoring image processing method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the wildlife monitoring image processing device in an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0017] like Figure 1 As shown, an embodiment of the present invention proposes a method for processing images for wildlife monitoring, comprising the following steps: Step 101: Acquire wildlife image data; the wildlife image data is obtained through infrared camera acquisition. Step 102: Perform effective image filtering on the wildlife image data to obtain effective image data; Step 103: Based on the valid image data and the trained wildlife recognition model, obtain the wildlife recognition result; the wildlife recognition model is obtained by training a preset network model based on collected historical wildlife images; Step 104: Perform visualization processing on the valid image data and the wildlife identification results to obtain visualized data; Step 105: Obtain wildlife monitoring results based on the wildlife identification results, the visualization data, and the historical monitoring data.

[0018] The wildlife monitoring image processing method of this invention acquires wildlife image data through an infrared camera, performs effective image filtering on the wildlife image data to obtain effective image data, obtains wildlife identification results based on the effective image data and a trained wildlife identification model, wherein the wildlife identification model is trained on a preset network model based on historical wildlife images, performs visualization processing based on the effective image data and the wildlife identification results to obtain visualization data, and finally obtains wildlife monitoring results based on the wildlife identification results, the visualization data, and historical monitoring data. This method not only enables multi-species, high-precision identification but also helps improve the accuracy of wildlife monitoring results.

[0019] In an optional embodiment of the present invention, step 101, acquiring wildlife image data, may include: Step 1011: Obtain target area data; Specifically, the target area data includes key areas within the target monitoring protected area (i.e., wildlife protected area), such as animal habitats, water sources, and migration routes. The target area can be set according to the actual situation to facilitate the subsequent deployment of infrared cameras based on the target area.

[0020] Step 1012: Deploy infrared cameras according to preset deployment conditions and target area data, and acquire wildlife image data through the infrared cameras.

[0021] Specifically, the preset deployment conditions include the infrared camera deployment density, which needs to be determined based on the area of ​​the target region, the frequency of animal activity, detection priority, etc. Here, through... The infrared camera deployment density D is obtained, where D is the infrared camera deployment density, S is the total area of ​​a single zone of the target monitoring and protection area (square kilometers), and B is the total number of infrared cameras to be deployed within the target area, in units of cameras. K is the regional monitoring priority correction coefficient, and A is the effective monitoring coverage area of ​​a single infrared camera (square kilometers / camera), rounded up to ensure full coverage. Here, the total area S of a single zone of the target monitoring protected area refers to the planar land area of ​​a small independent monitoring operation area separately divided from the entire large nature reserve according to habitat type, topography, management boundary, and functional zoning; the effective monitoring coverage area A of a single infrared camera is calculated based on the 50-meter effective trigger monitoring radius of the infrared camera, and the standard effective coverage area of ​​a single camera is fixed at 0.2 square kilometers / camera, which is reduced to 0.12 square kilometers / camera in complex forests, mountains, and other obstructed areas; the regional monitoring priority correction coefficient K is determined as follows: K=1.5 to 2.0 for core key areas (such as core habitats, water sources, and migration routes), K=1.0 for general protection monitoring areas, and K=0.6 to 0.8 for edge buffer transition areas; the total number of infrared cameras to be deployed in the target area is calculated and rounded up to ensure that the number of cameras deployed meets the minimum coverage requirements and eliminates monitoring blind spots.

[0022] Here, the infrared camera must meet the following core technical parameters to adapt to the complex lighting and temperature differences in the protected area: support 4G / 5G / Wi-Fi dual-mode transmission, enabling real-time data upload, local storage during network outages (maximum storage capacity ≥128GB), and automatic data synchronization after network recovery; minimum illumination ≤0.001 lux, supporting infrared night vision mode (infrared distance ≥50 meters), clearly capturing images of animal activity at night; image resolution ≥4K (3840×2160), video resolution ≥1080P, frame rate ≥25fps, ensuring the capture of animal morphology and behavioral details; standby power consumption ≤50mW, supporting solar power supply, with a battery life ≥3 months, adapting to environments without power supply in the wild; dynamic range ≥120dB, adapting to complex lighting scenarios such as strong light, backlight, and shadows, avoiding overexposure or underexposure of images.

[0023] Based on the above conditions, infrared cameras were selected and deployed in the target area according to the calculated deployment density. Wildlife image data was then acquired using these cameras. It should be noted that the infrared cameras used a dual-mode approach combining passive triggering and timed acquisition when acquiring wildlife images. Specifically, when an animal enters the monitoring range (trigger distance ≤ 50 meters), the camera is passively triggered to capture photos (3 frames per second) and videos (10 to 30 seconds per segment). Simultaneously, environmental images were acquired at fixed times each day (e.g., 2-4 AM and 1-3 PM) to aid in analyzing the correlation between animal activity patterns and the environment.

[0024] In an optional embodiment of the present invention, step 102, which involves filtering the wildlife image data to obtain valid image data, may include: Step 1021: Based on the grayscale variance and preset variance threshold of the wildlife image data, the wildlife image data is filtered once to obtain the first image data; Specifically, through The image grayscale variance is calculated, where, Let M be the image grayscale variance, M be the image pixel width, and N be the image pixel height. This represents the grayscale value (ranging from 0 to 255) of the pixel at image coordinates (x, y) in the wildlife image data. This represents the average grayscale value of all pixels in the entire image from the wildlife image data. When the image grayscale variance of an image in the wildlife image data is less than a preset variance threshold (e.g., 500), the image is determined to be a blurry image and is removed. The remaining images after removing the blurry images from the wildlife image data constitute the first image data.

[0025] Step 1022: Based on the pixel difference of the first image data and the preset difference threshold, perform a second filtering on the first image data to obtain the second image data; Specifically, through ,in, The global average pixel difference in the entire image in the first image data represents the overall motion amplitude of the image. It represents the grayscale difference between adjacent frames of a single pixel. , Let be the grayscale value of the pixel in frame t. The pixel grayscale value of the (t-1)th frame; when the global average pixel difference of the entire image in the first image data is less than the preset difference threshold (e.g., 20), it is determined that there is no moving target (empty shot, static scene) in the image and it is removed. After removing the images without moving targets in the first image data, the remaining images are the second image data.

[0026] Step 1023: Based on the target contour data of the second image data, perform effective image filtering on the second image data to obtain effective image data.

[0027] Specifically, firstly, target contours are extracted from the second image data to obtain target contour data, including the target contour width and height. This extraction can be achieved through the following steps: converting the color image in the second image data into a single-channel grayscale image to simplify subsequent contour calculations and eliminate color interference; setting a fixed grayscale threshold to convert the grayscale image into a pure black and white binary image, with the foreground suspected target white and the background environment black, thus forcibly separating the target from the background; using a combination of erosion and dilation operations to remove minor noise and fragmented pseudo-contours, filling in small gaps in the target contour to ensure contour integrity; traversing the binary image from top to bottom and from outside to inside to extract all continuously closed edge pixel sets, forming several independent target contour sets; discarding fragmented and invalid contours with too small an area, retaining only valid contours with entity target features for subsequent aspect ratio and area determination calculations. Then, the aspect ratio of the target contour is obtained using R=W / H, where R is the aspect ratio, W is the target contour width in the target contour data, and H is the target contour height in the target contour data; through... The area of ​​the region is obtained, where, The area of ​​the region (the area of ​​the region enclosed by a closed outline). Let be the coordinates of the i-th edge pixel in the target contour data. Let be the coordinates of the (i+1)th edge pixel of the contour in the target contour data, and n be the total number of pixels on the contour edge. The coordinates of the first and last edges are connected in a closed loop. The second image data is filtered for valid images based on the aspect ratio of the target contour, the area of ​​the region, the preset aspect ratio range (e.g., when the wild animal is a deer, the preset aspect ratio range is 1.5 to 2.8), and the preset area range of the target contour (e.g., when the wild animal is a deer, the preset area range of the target contour is 5000 to 20000 pixels). Here, images in the second image data whose target contour aspect ratio conforms to the preset aspect ratio range and whose area conforms to the preset area range of the target contour are considered valid image data.

[0028] In an optional embodiment of the present invention, step 103, obtaining the wildlife identification result based on the effective image data and the trained wildlife identification model, may include: Step 10311: Preprocess the valid image data to obtain preprocessed image data; Specifically, the effective image data is normalized using the same normalization method as the training samples. The size of the effective image data is uniformly adjusted to 640×640 pixels, and the pixel values ​​are normalized to the [0,1] range to improve the model processing efficiency.

[0029] Step 10312: Input the preprocessed image data into the first processing layer of the wildlife recognition model to obtain the first processing result; Specifically, through Obtain the pixel values ​​of the convolutional feature map. represents the pixel values ​​of the feature map after convolution, which are the deep features extracted from the preprocessed image data, including shape features (such as body shape, fur color, and limb structure) and behavioral features (such as limb posture). K is the size of the convolution kernel (3×3 convolution kernel). For convolution kernel weights, is the pixel value of the image in the preprocessed image data, and b is the bias term; Then through The first processing result is obtained, in which, The first processing result is the final feature pixel value after pooling operation compression and purification, where P is the pooling kernel size (2×2 standard pooling kernel). The output pixel values ​​of the feature map after the previous convolution operation are denoted as , i and j are the coordinates of the new small feature map after pooling, and k and l are the sliding offsets within the pooling kernel (ranging from 0 to P-1). First, a deep feature map containing the physical and behavioral characteristics of wild animals is generated through convolution operations. Then, pooling operations perform dimensionality reduction, compression, noise reduction, and purification on the convolution output features, retaining the core effective features without adding or generating any new features. Only existing features are simplified and enhanced, reducing the computational cost of subsequent model calculations.

[0030] Step 10313: Input the first processing result into the second processing layer of the wildlife recognition model to obtain the second processing result; Specifically, through The second processing result is obtained, in which, The second processing result is the probability predicted by the model that the current target belongs to the f-th class of wild animal species. This represents the deep feature output value corresponding to the f-th species in the first processing result, where F is the total number of all monitored wild animal species in the protected area. This is the deep feature output value corresponding to the g-th species in the first processing result.

[0031] Step 10314: Input the second processing result into the third processing layer of the wildlife recognition model to obtain the wildlife recognition result.

[0032] Specifically, the species with the highest probability from the second processing results is selected as the final wildlife identification result.

[0033] In an optional embodiment of the present invention, step 103, the training process of the wildlife recognition model includes: Step 10321: Obtain historical images of wild animals; Specifically, historical wildlife images can be created by combining animal photos and videos captured by infrared cameras over a period of time from various protected areas with publicly available datasets of rare and endangered wild animals collected online. These images need to cover individuals of different species, postures (standing, foraging, resting), environments (day / night, strong light / backlight, sunny / rainy), and age groups to ensure the diversity and representativeness of the dataset.

[0034] Step 10322: Perform data annotation processing on the historical wildlife images to obtain annotated image data; Specifically, open-source image annotation tools are used to annotate historical wildlife images. Annotations include: species category (e.g., giant panda, golden monkey, deer, etc.), individual outline (boundary box coordinates), and behavioral category (foraging, resting, mating, migration). Audio file format is used for annotation, ensuring an accuracy rate of ≥98%. Furthermore, images can be named and managed according to their source, allowing for filtering, classification, and annotation based on criteria such as whether the image is clear (1 indicates animal presence, 0 indicates absence, 2 indicates human presence, 3 indicates both animal and human presence, daytime or nighttime (1 indicates daytime, 0 indicates nighttime), and animal name).

[0035] Step 10323: Determine sample data based on the quality assessment results of the labeled image data; Specifically, the completeness and accuracy of the labeled image data can be assessed. If the assessment results meet the preset qualification criteria—that is, ensuring that all types of data have been fully collected and can comprehensively reflect the actual situation of wild animals within the nature reserve—then the qualified labeled image data can be used as sample data. Here, techniques such as cross-validation can be used to assess the accuracy of the labeled image data to ensure its accuracy.

[0036] Here, the completeness and accuracy of the labeled image data can be evaluated through the following steps: The total number of labeled image data samples is counted, and the actual number of labeled samples corresponding to each type of wild animal is counted for each category. The actual number of samples in each category is compared with the preset minimum number of samples. If the number of samples in any category is lower than the minimum threshold, it is determined that the data is incomplete and marked as missing samples to be supplemented. The labeled image data is classified and statistically analyzed in multiple dimensions according to the shooting light intensity, day and night time, season and climate, target distance scale, target occlusion degree, and target pose angle. The sample distribution under each scenario is checked to see if it meets the preset ratio requirements. If the sample proportion of a certain key scenario is insufficient or the scene coverage is missing, it is determined that the scene dimension completeness does not meet the standard. Iterate through all labeled images in the labeled image data one by one, and check whether each image completely contains all the required annotation fields such as target category label, target bounding box coordinates, target key points, occlusion status annotation, and image quality marker; check for empty annotations, missing fields, and missing annotation content, and determine that the field completeness is unqualified if there are any missing fields. The integrity assessment is based on three factors: the number of samples, scene distribution, and labeled fields. If all indicators meet the preset qualification conditions, the integrity of the labeled image data is deemed qualified; if any one of them is not met, the integrity assessment fails, and the corresponding missing samples need to be supplemented and the labeled content needs to be improved. Each labeled image in the image data is checked to ensure that the wildlife category labels in the image annotations are consistent with the actual target species in the image, and problems such as incorrect category labeling, species confusion, and incorrect label selection are investigated; labeled samples whose category labels do not match the actual targets are marked as incorrectly labeled and unqualified samples. Check the degree of fit between the bounding boxes in the labeled image data and the actual outline of the wild animal target, and calculate the intersection-union ratio (IOU) between the labeled boxes and the standard manual verification reference boxes; if the IOU is lower than the preset threshold, or the labeled boxes are too offset, select too much background, or do not completely wrap the target, the location labeling is judged to be inaccurate. Each image in the labeled image data is checked to see if all real wildlife targets in the image have been labeled. If a target is not labeled, it is considered a missing label. If there are multiple duplicate label boxes for the same target, it is considered a redundant label. Missing or duplicate labeled samples are marked as data with unacceptable accuracy. When all labeled sample categories are correctly labeled, the bounding box accuracy meets the standard, there are no missing or duplicate labels, and all quantitative indicators meet the preset error allowable range, the accuracy of the labeled data is deemed qualified; otherwise, it is returned for correction until the verification is qualified.

[0037] Step 10324: Perform data augmentation on the sample data to obtain training samples; Specifically, image enhancement processing is performed on the sample data to obtain enhanced image data; image normalization processing is then performed on the enhanced image data to obtain training samples. Image enhancement processing can expand the scale of the sample data (increasing the data volume by 3 to 5 times), avoiding model overfitting. Here, any of the following methods can be used to perform image enhancement processing on the sample data, or other image enhancement processing methods can be selected according to the actual situation. The following are just examples: A cropping window is randomly generated on the original image of the sample data. The window size is randomly selected between 0.6 and 1.0 times the size of the original image. The image is cropped according to the randomly generated upper left corner coordinates. The cropped image is then interpolated and scaled back to the preset base size to complete the random cropping enhancement.

[0038] Randomly generate probability values ​​between 0 and 1; when the probability value is greater than 0.5, perform a horizontal flip on the original image of the sample data; optionally, perform a vertical flip (adapted to the pose of the wild animal target); retain the flipped image as augmented image data; A scaling factor is randomly generated, ranging from 0.5 to 1.5 times; the original image of the sample data is scaled proportionally according to the scaling factor; if the scaled size is larger than the base size, the center is cropped to the target size; if the scaled size is smaller than the base size, the edges are padded with zeros / black padding and then scaled to the target size to obtain the enhanced image data. A brightness adjustment coefficient is randomly generated, ranging from 0.7 to 1.3. Each pixel value of the original image of the sample data is adjusted according to the following formula: New pixel value = Original pixel value × Brightness coefficient. Pixel values ​​that exceed the range of 0 to 255 are truncated to ensure that the pixel values ​​are valid, thus obtaining the enhanced image data.

[0039] Image normalization adjusts the size of the enhanced image to 640×640 pixels, and normalizes the pixel values ​​to the [0,1] range. The formula is as follows: ,in, For normalized images, To enhance the original pixel values ​​of the image, =0, It is 255.

[0040] Step 10325: Train the preset network model according to the preset training strategy and the training samples to obtain the wildlife recognition model.

[0041] Specifically, the preset training strategy includes: loss function Learning rate The batch size is 32, the total training epochs are 100, the AdamW optimizer (adaptive moment estimator) is used, the momentum is 0.9, and the weight decay is 0.0005. For loss function, , , These are the weighting coefficients. For species classification loss, the cross-entropy loss function is used to optimize the accuracy of species identification. C is the number of training samples. The true label for the sample (0 or 1). To predict the probability that the sample belongs to a certain species for the model, For bounding box regression loss, , CIoU loss function is used to optimize the localization accuracy of the target contour bounding box. To measure the intersection-union ratio (IUU) between the measured bounding box and the ground truth bounding box, is the squared Euclidean distance between the centers of the predicted and ground truth bounding boxes, c is the diagonal length of the smallest bounding rectangle enclosing both the predicted and ground truth bounding boxes, a is the balance coefficient, and v is a parameter that measures the consistency of the aspect ratio between the predicted and ground truth bounding boxes. For behavior classification loss, the cross-entropy loss function is used to optimize the accuracy of animal behavior recognition. The formula is the same as... Consistent, except the label is replaced with a behavior category label. Among them, For learning rate, To minimize the learning rate, The initial learning rate is 0.001. For the current training round, This refers to the total number of training rounds.

[0042] Here, the loss function is used to ensure that the model simultaneously optimizes the accuracy of three tasks (species classification, bounding box regression, and behavior classification).

[0043] The training samples were divided into a training set (70%), a validation set (20%), and a test set (10%) in a ratio of 7:2:1. The training set was used for training model parameters, the validation set was used for parameter adjustment during the training process, and the test set was used for the final performance evaluation of the model.

[0044] When training the pre-defined network model, the training set is input into the model, and the model parameters (only the parameters of the first processing layer, for the first 50 rounds) are updated by calculating the gradient of each learnable parameter through backward differentiation of the loss function layer by layer. After each training round, the model performance is evaluated using a validation set (mean precision, accuracy, and recall are calculated). If the validation set accuracy does not improve for 5 consecutive rounds, an early stopping strategy is adopted to stop training and avoid overfitting. In the last 50 rounds, the first processing layer is unfrozen, and the parameters of all layers are fine-tuned to further improve the model accuracy. After training, the model performance is evaluated using a test set. The core evaluation metrics include: Species recognition accuracy: Number of correctly identified species / total number of test set samples, target ≥ 95%; Average precision: Average precision of all species, target ≥ 90%; Behavior recognition accuracy: Number of correctly identified animal behaviors / number of behavior samples in the test set, target ≥ 88%; Recognition speed: Single image recognition time ≤ 30ms, meeting real-time recognition requirements. If the model performance does not meet the target, the following optimization measures will be adopted: increase dataset diversity (supplementing samples of specific environments and poses), adjust loss function weights, optimize learning rate strategy, add data augmentation methods, and retrain the model until the target performance is achieved. The final trained model will be used as the wildlife recognition model.

[0045] In an optional embodiment of the present invention, step 104, which involves performing visualization processing based on the valid image data and the wildlife identification results to obtain visualization data, may include: Step 1041: Perform image enhancement processing on the effective image data to obtain enhanced image data; Specifically, through Image enhancement processing is performed on blurred images in the valid image data to obtain enhanced image data, where, To enhance image data, i.e., to enhance the image's frequency domain representation, Let be the complex conjugate of the fuzzy degeneracy function. For the image power spectrum, For fuzzy degradation function, The noise power spectrum, The power spectrum of the original sharp image in the effective image data. This represents the frequency domain of a blurred image.

[0046] Step 1042: Perform visualization processing on the enhanced image data and the wildlife identification results according to the preset visualization scheme to obtain visualized data.

[0047] Specifically, the preset visualization scheme includes map visualization, which involves marking the deployment locations (latitude and longitude) of infrared cameras and the specific locations where wild animals are spotted (latitude and longitude, obtained from the infrared camera's GPS module) on an electronic map, using different colored markers to distinguish different species. Clicking on a marker allows users to view real-time images and species information for that location. In practice, the wild animal identification results, enhanced image data, and the deployment locations of the infrared cameras are bound together to obtain a binding result. The binding result is then visualized according to the preset visualization scheme to obtain visualized data. Here, the binding result is marked on an electronic map, using different colored markers to distinguish different species. Clicking on a marker allows users to view real-time images and species information for that location, providing a convenient and intuitive view of the monitoring results.

[0048] In an optional embodiment of the present invention, step 105, obtaining the wildlife monitoring result based on the wildlife identification result, the visualization data, and the historical monitoring data, may include: Step 1051: Obtain preset early warning conditions; Specifically, preset early warning conditions may include: when rare wild animals or abnormal behaviors (such as injuries or abnormal gatherings) are identified, an early warning (such as an audible warning and / or a platform pop-up warning) is issued, and the early warning information (species, location, time, and images) is pushed to the mobile phones and computers of management personnel in real time for timely handling.

[0049] Step 1052: Determine the warning data based on the preset warning conditions and the wildlife identification results; Specifically, when the wildlife identification result meets any of the preset warning conditions, such as the species identification result in the wildlife identification result matching the rare wildlife in the preset warning conditions, the warning data includes: warning method (such as sound warning and / or platform pop-up warning), warning content (which may include: species, location, time, image, etc.).

[0050] Step 1053: Obtain statistical data based on the wildlife identification results and historical monitoring data; Specifically, based on the wildlife identification results, the species name, quantity, location (latitude and longitude), and time of occurrence can be determined. The wildlife identification results and the species name, quantity, location (latitude and longitude), and time of occurrence in historical monitoring data can be compared in the form of charts to form statistical data, which facilitates comparison and analysis based on real-time monitoring results and historical monitoring data.

[0051] Step 1054: Obtain wildlife monitoring results based on the early warning data, the visualization data, and the statistical data.

[0052] Specifically, wildlife monitoring results can be wildlife monitoring reports, which are generated by filling early warning data, visualization data, and statistical data into a preset report template.

[0053] A specific embodiment of the wildlife monitoring image processing method of the present invention includes: Step 111: Acquire wildlife image data; the wildlife image data is obtained through infrared camera acquisition; First, determine the area data that needs to be monitored, and then deploy infrared cameras in the target area according to the calculated infrared camera deployment density to collect wildlife image data.

[0054] Step 112: Perform effective image filtering on the wildlife image data to obtain effective image data; After removing blurry images and images without moving targets from the wildlife image data, the effective image data is obtained through effective image filtering.

[0055] Step 113: Based on the valid image data and the trained wildlife recognition model, obtain the wildlife recognition result; the wildlife recognition model is obtained by training a preset network model based on collected historical wildlife images; After acquiring historical wildlife images, data annotation and quality assessment are performed. Once the quality is deemed satisfactory, the pre-set network model is trained using the historical wildlife images to obtain a wildlife recognition model.

[0056] Step 114: Perform visualization processing on the valid image data and the wildlife identification results to obtain visualized data; After image enhancement processing is performed on the effective image data, the enhanced image data and wildlife identification results are bound together according to the preset visualization scheme and marked on the electronic map to achieve visualization processing and obtain visualized data.

[0057] Step 115: Obtain wildlife monitoring results based on the wildlife identification results, the visualization data, and the historical monitoring data.

[0058] The warning data is determined by comparing the preset warning conditions with the wildlife identification results; the warning data, visualization data and statistical data are then filled into the preset report template to obtain the wildlife monitoring results.

[0059] The wildlife monitoring image processing method of this invention utilizes an infrared camera with characteristics such as low illumination, low power, and wide dynamic range, making it suitable for the complex environment of protected areas. The trained wildlife recognition model supports recognition of low-light and blurred images, with a wide range of applications. Through this model, multi-species, high-precision identification is achieved, with a species recognition accuracy of ≥95% and a behavior recognition accuracy of ≥88%, effectively avoiding misidentification and missed identification. Real-time output of monitoring results allows managers to quickly and intuitively understand the dynamics of protected areas and the population status of wildlife, enabling effective monitoring of wildlife.

[0060] like Figure 2 As shown, an embodiment of the present invention provides a wildlife monitoring image processing device 200, comprising: The acquisition module 201 is used to acquire wildlife image data; the wildlife image data is obtained through infrared camera acquisition. The processing module 202 is used to perform effective image filtering on the wildlife image data to obtain effective image data; to obtain wildlife identification results based on the effective image data and a trained wildlife identification model; the wildlife identification model is obtained by training a preset network model based on collected historical wildlife images; to perform visualization processing on the effective image data and the wildlife identification results to obtain visualized data; and to obtain wildlife monitoring results based on the wildlife identification results, the visualized data, and historical monitoring data.

[0061] Optionally, acquire wildlife image data, including: Obtain data for the target area; Infrared cameras are deployed according to preset deployment conditions and target area data, and wildlife image data are acquired through the infrared cameras.

[0062] Optionally, the wildlife image data is filtered to obtain valid image data, including: Based on the grayscale variance and a preset variance threshold of the wildlife image data, the wildlife image data is filtered once to obtain the first image data. Based on the pixel difference of the first image data and a preset difference threshold, the first image data is filtered a second time to obtain the second image data; Based on the target contour data of the second image data, the second image data is filtered for effective images to obtain effective image data.

[0063] Optionally, based on the effective image data and the trained wildlife recognition model, the wildlife recognition result is obtained, including: The valid image data is preprocessed to obtain preprocessed image data; The preprocessed image data is input into the first processing layer of the wildlife recognition model to obtain the first processing result; The first processing result is input into the second processing layer of the wildlife recognition model to obtain the second processing result; The second processing result is input into the third processing layer of the wildlife recognition model to obtain the wildlife recognition result.

[0064] Optionally, the first processing result is input into the second processing layer of the wildlife recognition model to obtain a second processing result, including: pass The second processing result is obtained, in which, The second processing result is the probability predicted by the model that the current target belongs to the f-th class of wild animal species. This represents the deep feature output value corresponding to the f-th species in the first processing result, where F is the total number of all monitored wild animal species in the protected area. This is the deep feature output value corresponding to the g-th species in the first processing result.

[0065] Optionally, the training process of the wildlife recognition model includes: Acquire historical images of wild animals; The historical wildlife images are subjected to data annotation processing to obtain annotated image data; Based on the quality assessment results of the labeled image data, the sample data is determined; The sample data is augmented to obtain training samples; The preset network model is trained according to the preset training strategy and the training samples to obtain the wildlife recognition model.

[0066] Optionally, visualization processing is performed based on the valid image data and the wildlife identification results to obtain visualized data, including: The effective image data is subjected to image enhancement processing to obtain enhanced image data; The enhanced image data and the wildlife identification results are visualized according to a preset visualization scheme to obtain visualized data.

[0067] Optionally, based on the wildlife identification results, the visualization data, and historical monitoring data, wildlife monitoring results are obtained, including: Obtain preset early warning conditions; Based on the preset warning conditions and the wildlife identification results, the warning data is determined; Statistical data were obtained based on the wildlife identification results and historical monitoring data. Based on the early warning data, the visualization data, and the statistical data, the results of wildlife monitoring are obtained.

[0068] The wildlife monitoring image processing device of this invention acquires wildlife image data through an infrared camera, performs effective image filtering on the wildlife image data to obtain effective image data, and obtains wildlife identification results based on the effective image data and a trained wildlife identification model. The wildlife identification model is trained on a preset network model based on historical wildlife images. The effective image data and the wildlife identification results are then visualized to obtain visualized data. Finally, the wildlife monitoring results are obtained based on the wildlife identification results, the visualized data, and historical monitoring data. This device not only enables multi-species, high-precision identification but also helps improve the accuracy of wildlife monitoring results.

[0069] It should be noted that this device corresponds to the method described above, and all implementations in the method embodiments described above are applicable to the embodiments of this device and can achieve the same technical effect. Further details will not be provided in this embodiment.

[0070] This invention also provides a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method as described in any of the above embodiments. All implementations in the above method embodiments are applicable to the embodiments of this device and can achieve the same technical effects. Further details are omitted in this embodiment.

[0071] This invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in any of the above embodiments. All implementations in the above method embodiments are applicable to the embodiments of this device and can achieve the same technical effects. Further details are omitted in this embodiment.

[0072] It should be noted that in the apparatus and method of the present invention, the components or steps can obviously be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Furthermore, the steps for performing the above series of processes can naturally be performed in the order described and in chronological order, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel, overlapping, or independently of each other.

[0073] It should be noted that in the above embodiments, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments described above is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0074] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for processing images used in wildlife monitoring, characterized in that, include: Acquire wildlife image data; the wildlife image data is obtained through infrared cameras. The wildlife image data is filtered to obtain valid image data; Based on the effective image data and the trained wildlife recognition model, the wildlife recognition result is obtained; the wildlife recognition model is obtained by training a preset network model based on collected historical wildlife images; Visualization data is obtained by performing visualization processing on the valid image data and the wildlife identification results; Based on the wildlife identification results, the visualization data, and the historical monitoring data, the wildlife monitoring results are obtained.

2. The wildlife monitoring image processing method according to claim 1, characterized in that, The wildlife image data is filtered to obtain valid image data, including: Based on the grayscale variance and a preset variance threshold of the wildlife image data, the wildlife image data is filtered once to obtain the first image data. Based on the pixel difference of the first image data and a preset difference threshold, the first image data is filtered a second time to obtain the second image data; Based on the target contour data of the second image data, the second image data is filtered for effective images to obtain effective image data.

3. The wildlife monitoring image processing method according to claim 1, characterized in that, Based on the effective image data and the trained wildlife recognition model, the wildlife recognition results are obtained, including: The valid image data is preprocessed to obtain preprocessed image data; The preprocessed image data is input into the first processing layer of the wildlife recognition model to obtain the first processing result; The first processing result is input into the second processing layer of the wildlife recognition model to obtain the second processing result; The second processing result is input into the third processing layer of the wildlife recognition model to obtain the wildlife recognition result.

4. The wildlife monitoring image processing method according to claim 3, characterized in that, The first processing result is input into the second processing layer of the wildlife recognition model to obtain the second processing result, including: pass The second processing result is obtained, in which, The second processing result is the probability predicted by the model that the current target belongs to the f-th class of wild animal species. This represents the deep feature output value corresponding to the f-th species in the first processing result, where F is the total number of all monitored wild animal species in the protected area. This is the deep feature output value corresponding to the g-th species in the first processing result.

5. The wildlife monitoring image processing method according to claim 1, characterized in that, The training process of the wildlife recognition model includes: Acquire historical images of wild animals; The historical wildlife images are subjected to data annotation processing to obtain annotated image data; Based on the quality assessment results of the labeled image data, the sample data is determined; The sample data is augmented to obtain training samples; The preset network model is trained according to the preset training strategy and the training samples to obtain the wildlife recognition model.

6. The wildlife monitoring image processing method according to claim 1, characterized in that, Visualization data is obtained by performing visualization processing on the valid image data and the wildlife identification results, including: The effective image data is subjected to image enhancement processing to obtain enhanced image data; The enhanced image data and the wildlife identification results are visualized according to a preset visualization scheme to obtain visualized data.

7. The wildlife monitoring image processing method according to claim 1, characterized in that, Based on the wildlife identification results, the visualization data, and historical monitoring data, the wildlife monitoring results are obtained, including: Obtain preset early warning conditions; Based on the preset warning conditions and the wildlife identification results, the warning data is determined; Statistical data were obtained based on the wildlife identification results and historical monitoring data. Based on the early warning data, the visualization data, and the statistical data, the results of wildlife monitoring are obtained.

8. A wildlife monitoring image processing device, characterized in that, include: The acquisition module is used to acquire wildlife image data; the wildlife image data is obtained through infrared camera acquisition. The processing module is used to perform effective image filtering on the wildlife image data to obtain effective image data; to obtain wildlife identification results based on the effective image data and a trained wildlife identification model; the wildlife identification model is obtained by training a preset network model based on collected historical wildlife images; to perform visualization processing on the effective image data and the wildlife identification results to obtain visualized data; and to obtain wildlife monitoring results based on the wildlife identification results, the visualized data, and historical monitoring data.

9. A computing device, characterized in that, include: A processor, a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images