A method and device for identifying a field bird net, a storage medium and an electronic device

Abstract

The application discloses a field bird catching net recognition method and device, a storage medium and electronic equipment, and belongs to the technical field of artificial intelligence. The method comprises the following steps: a rectangular marking box is used to mark the net surface area of the bird catching net in a target image and all support rods for fixing the bird catching net in the target image, so that a target data set is obtained; data in the target data set is used as training data, and in each iteration process of model training, a data enhancement operation is performed on the target image and the marking information in the training data, so that corresponding enhanced data is obtained; an initial target detection model is obtained; and the initial target detection model is fine-tuned based on the enhanced data, so that a fine-tuned target detection model is obtained, and the field bird catching net is recognized based on the fine-tuned target detection model. Through the recognition method, the field bird catching net can be accurately recognized.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and specifically relates to a method, device, storage medium and electronic device for identifying bird-catching nets in the wild. Background Technology

[0002] Migratory bird conservation is a crucial issue in global ecological protection. Illegal bird netting is one of the main human threats to migratory bird populations along several major flyways. In recent years, drone technology, due to its high mobility, wide field of view, and ability to carry various sensors, has been increasingly applied to wildlife conservation and anti-poaching patrols. Utilizing drones for aerial photography of key areas and using AI algorithms to automatically identify illegal facilities represents a technological trend aimed at improving conservation effectiveness.

[0003] There are two main types of existing implementation schemes: Option 1: Drone aerial photography + manual visual interpretation: Patrol personnel use drones to capture videos or photos, and after the flight, they review the footage on a computer and visually search for bird-catching nets in the images.

[0004] Efficiency bottleneck: the interpretation time is usually several times the flight time, and the human eye is easily fatigued, resulting in a very high rate of missed detection for small targets and targets that blend into the background.

[0005] Unable to provide real-time alerts: The inability to detect and lock onto targets in real time during flight may result in missing the best opportunity to stop illegal activities on the spot.

[0006] Option 2: An attempt at recognition based on a general object detection model: A few studies have attempted to directly use general detection models such as YOLO and Faster R-CNN to identify bird-catching nets. Their fatal flaw lies in: Ignoring the feature degradation problem in high-altitude oblique shots: During the shooting process, the mesh lines, with a diameter of only 1-2 mm, occupy less than 2 pixels in the image, appearing as extremely faint semi-transparent textures. If the model only uses the mesh as the detection target, its feature extraction layer can hardly learn any stable and distinguishable features, leading to model non-convergence or an extremely high false alarm rate.

[0007] Severe overfitting due to small sample size: It is extremely difficult to obtain samples of illegal bird-catching nets, and the amount of data is far from sufficient to support the target detection model to be trained from scratch. The model quickly overfits to the specific background at the time of training with limited samples. When the background is changed to a new woodland or shrubland, the recognition performance drops precipitously.

[0008] In conclusion, how to provide a method for identifying wild bird-catching nets that can be accurately identified is a problem that needs to be solved. Summary of the Invention

[0009] To address the aforementioned problems, this application provides a method, apparatus, storage medium, and electronic device for identifying bird-catching nets in the wild, which can accurately identify bird-catching nets in the wild.

[0010] In a first aspect, embodiments of this application provide a method for identifying bird-catching nets in the wild, the method comprising: Acquire target images under various conditions. The target images are images obtained by UAV aerial photography of key areas of migratory bird flyways. The target images include images of bird-catching net facilities. Based on preset annotation rules, the target image is annotated to obtain the corresponding rectangular annotation box. The preset annotation rules include: treating the bird-catching net surface and the support pole used to set up the bird-catching net as a whole target object, and defining the category of the target object as the bird-catching net facility. The target dataset is obtained by annotating the net area of ​​the bird-catching net in the target image and all the support poles used to fix the bird-catching net in the target image using the rectangular annotation box. The target dataset includes: the target image and the annotation file containing annotation information. All support poles include: bamboo poles, wooden sticks and metal pipes. The data in the target dataset is used as training data, and in each iteration of model training, data augmentation operations are performed on the target images and annotation information in the training data to obtain the corresponding augmented data; Obtain the initial object detection model; The initial target detection model is fine-tuned and trained based on the enhanced data to obtain a fine-tuned target detection model, which is then used to identify bird-catching nets in the wild.

[0011] Optionally, before fine-tuning the initial object detection model based on the enhanced data, the method further includes: The initial object detection model is loaded and pre-trained on a large-scale object detection dataset to obtain the backbone network weights; The backbone network weights are used as initial parameters to enable the initial object detection model to have general object feature extraction performance.

[0012] Optionally, the fine-tuning training of the initial target detection model based on the enhanced data includes: Based on the enhanced data, the following operations are performed sequentially to conduct the first phase of fine-tuning training: Freeze the weight parameters of the first N residual blocks or convolutional layers in the backbone network of the initial target detection model so that the model does not participate in gradient updates during training, where N is a natural number greater than or equal to 1. Only the parameters of the neck network and the head part of the initial target detection model are trained to adapt the model to the high-level semantic features corresponding to the bird-catching net facility; Based on the enhanced data, the following operations are performed sequentially to conduct the second phase of fine-tuning training: Once the loss function corresponding to the first stage is detected to have converged, the weight parameters of all layers of the backbone network are unfrozen, and a differentiated learning rate is set so that the model can learn the unique texture, geometry, and spatial combination relationship of the pole network facility while preserving general features.

[0013] Optionally, setting the differentiated learning rate includes: Set the learning rate corresponding to the shallow layer of the backbone network of the initial object detection model as the first learning rate; Set the learning rate corresponding to the deep layer of the ancient port network of the initial object detection model as the second learning rate; Set the learning rate of the neck network and the detection head of the initial object detection model as the third learning rate; Set the global learning rate, which finely adjusts all parameters, as the fourth learning rate; The fourth learning rate, the first learning rate, the second learning rate, and the third learning rate increase sequentially.

[0014] Optionally, performing data augmentation operations on the target images and annotation information in the training data includes: Perform data augmentation operations on the target images and annotation information in the training data. The data augmentation operations include at least one or more random combinations of the following: Perform geometry enhancement operations by executing random rotation, random scaling, random translation, and random horizontal flipping. A luminance enhancement operation is performed by randomly perturbing the brightness, contrast, and saturation values ​​with preset values ​​in the HSV color space. Weather simulation enhancement is performed by overlaying simulated fog layers of random concentration onto images based on an atmospheric scattering model. By randomly selecting one or more rectangular regions in the image using a random erasure method, and configuring their pixel values ​​as random noise or a constant, the operation of occlusion simulation enhancement is performed. The various occlusion scenarios include: the support pole being partially occluded by branches and shrubs, or the net being partially occluded by bird carcasses. Mosaic enhancement is performed by randomly selecting multiple training images with different backgrounds, scaling and cropping them respectively, and then stitching the scaled and cropped training images together to obtain a new image; and adjusting the coordinates of the corresponding bounding boxes based on the new image to enrich the background context in which the target appears.

[0015] Optionally, the step of marking the net area of ​​the bird-catching net in the target image and all the support rods in the target image used to fix the bird-catching net using the rectangular annotation box includes: If the same bird-catching net facility uses multiple support poles, then each of the multiple support poles is determined to serve the same bird-catching net, and each of the multiple support poles is included in the same label box.

[0016] Optionally, acquiring target image data under various conditions includes: Acquire target image data under different lighting conditions, including: sunny lighting conditions, cloudy lighting conditions, and dawn / dusk lighting conditions; or, Acquire target image data under different seasonal background conditions, including: background conditions during the green period and background conditions during the withered period; or... Acquire target image data under different terrain environmental conditions, including: forest environment conditions, reed marsh environment conditions, shrubland environment conditions, and farmland edge environment conditions.

[0017] Secondly, embodiments of this application provide a device for identifying bird-catching nets in the wild, the device comprising: The first acquisition module is used to acquire target images under various conditions. The target images are images obtained by UAV aerial photography of key areas of migratory bird flyways. The target images include images of bird-catching net facilities. The first annotation module is used to annotate the target image based on preset annotation rules to obtain the corresponding rectangular annotation box. The preset annotation rules include: treating the bird-catching net surface and the support pole used to set up the bird-catching net as a whole target object, and defining the category of the target object as the bird-catching net facility. The second annotation module is used to annotate the net area of ​​the bird-catching net in the target image and all the support poles in the target image used to fix the bird-catching net using the rectangular annotation box, so as to obtain the target dataset. The target dataset includes: the target image and the annotation file containing annotation information. All support poles include: bamboo poles, wooden sticks and metal pipes. The data augmentation module is used to use the data in the target dataset as training data, and to perform data augmentation operations on the target images and annotation information in the training data during each iteration of model training to obtain the corresponding augmented data. The second acquisition module is used to acquire the initial target detection model; The fine-tuning training module is used to fine-tune the initial target detection model based on the enhanced data to obtain a fine-tuned target detection model, which is then used to identify bird-catching nets in the wild.

[0018] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0019] Fourthly, an electronic device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method of the first aspect.

[0020] Compared with the prior art, this application has the following advantages: It can accurately identify bird-catching nets in the wild.

[0021] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a method for identifying bird-catching nets in the wild according to an embodiment of this application is shown; Figure 2 A schematic diagram of the structure of a bird-catching net identification device 200 according to an embodiment of this application is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] This application provides a method and apparatus for identifying bird-catching nets in the wild, a computer-readable medium, and an electronic device, which will be described below with reference to the accompanying drawings.

[0026] Please refer to Figure 1 It illustrates a flowchart of a method for identifying bird-catching nets in the wild, provided by some embodiments of this application, such as... Figure 1 As shown, the method for identifying bird-catching nets in the wild may include the following steps: Step S101: Acquire target images under various conditions. The target images are images obtained by the UAV taking aerial photos of key areas of the migratory bird flyway. The target images include images of bird-catching net facilities.

[0027] In one example, acquiring target image data under diverse conditions includes the following steps: Acquire target image data under different lighting conditions, including: sunny lighting conditions, cloudy lighting conditions, and dawn / dusk lighting conditions; or, Acquire target image data under different seasonal background conditions, including: background conditions during the green period and background conditions during the withered period; or... Acquire target image data under different terrain environmental conditions, including: forest environment conditions, reed marsh environment conditions, shrubland environment conditions, and farmland edge environment conditions.

[0028] In practical applications, the process of acquiring aerial image data from drones is described below: Using drones, aerial photography was conducted over key areas of migratory bird flyways at a height of 30-50 meters relative to the ground, a lens tilt angle of 45°, and fixed parameters without optical zoom to collect visible light images of bird-catching nets. The collected images should cover different lighting conditions (sunny, cloudy, dawn / dusk), different seasonal backgrounds (green season, withered season), and different terrain environments (woodland, reed beds, shrublands, farmland edges) to enhance sample diversity.

[0029] The above operations enable large-scale, high-efficiency automated patrols: drones take aerial photos at a height of 30-50 meters, with a 45° lens and without zoom, and transmit the images back to the algorithm server in real time for intelligent identification, replacing inefficient manual foot searches and significantly improving the detection efficiency and coverage area of ​​bird-catching nets along migratory bird flyways.

[0030] Step S102: Based on the preset annotation rules, the target image is annotated to obtain the corresponding rectangular annotation box. The preset annotation rules include: treating the bird-catching net surface and the support pole used to set up the bird-catching net as a whole target object, and defining the category of the target object as bird-catching net facility.

[0031] It should be noted that the bird-catching net surface and its support pole are defined as a single, integrated detection target for labeling and identification. By utilizing the significant visual features of the support pole and the spatial constraint relationship of the net surface, the concealed net body can be indirectly located using the features of the clearly visible support pole, thus indirectly achieving the location of low-visibility net body areas.

[0032] In the identification method provided in this application embodiment, the pole-net integrated target definition method binds the feature-prominent support pole to the net surface as an overall detection unit. By detecting the support pole and utilizing its spatial constraint relationship with the net surface, the net area is indirectly located. This fundamentally breaks through the physical bottleneck of visual detection of high-altitude bird-catching nets, making automated identification practically feasible, and effectively solving the problem of the net being invisible under high-altitude oblique shooting conditions.

[0033] It should be noted that the definition of the supporting structure in the "pole-net integration" is as follows: when the bird-catching net is not supported by an independent support pole, but is directly tied to tree trunks, shrub branches, or natural rock formations, the "tying point and the net edge in a stretched state due to stress" can be defined as the salient part of the combined target, replacing the role of the support pole in the aforementioned scheme. The labeling rules are adjusted accordingly: the bounding box must simultaneously include the visible net area and the aforementioned natural fixing structure.

[0034] Step S103: Label the net area of ​​the bird-catching net in the target image and all the support poles used to fix the bird-catching net in the target image using rectangular label boxes to obtain the target dataset. The target dataset includes: the target image and the label file containing the label information. All support poles include: bamboo poles, wooden sticks and metal pipes.

[0035] In one example, the net area of ​​the bird-catching net in the target image, as well as all the support poles used to fix the bird-catching net in the target image, are labeled using rectangular annotation boxes, including the following steps: If the same bird-catching net facility uses multiple support poles, then each of the multiple support poles must serve the same bird-catching net, and each of the multiple support poles must be included in the same label box.

[0036] Step S104: Use the data in the target dataset as training data, and in each iteration of model training, perform data augmentation operations on the target images and annotation information in the training data to obtain the corresponding augmented data.

[0037] In one example, data augmentation is performed on the target images and annotation information in the training data, including the following steps: Perform data augmentation operations on the target images and annotation information in the training data. The data augmentation operations include at least one or more random combinations of the following: Perform geometry enhancement operations by executing random rotation, random scaling, random translation, and random horizontal flipping. A luminance enhancement operation is performed by randomly perturbing the brightness, contrast, and saturation values ​​with preset values ​​in the HSV color space. Weather simulation enhancement is performed by overlaying simulated fog layers of random concentration onto images based on an atmospheric scattering model. By randomly selecting one or more rectangular regions in the image using a random erasure method, and configuring their pixel values ​​as random noise or a constant, the operation of occlusion simulation enhancement is performed. The various occlusion scenarios include: the support pole being partially occluded by branches and shrubs, or the net being partially occluded by bird carcasses. Mosaic enhancement is performed by randomly selecting multiple training images with different backgrounds, scaling and cropping them respectively, and then stitching the scaled and cropped training images together to obtain a new image; and adjusting the coordinates of the corresponding bounding boxes based on the new image to enrich the background context in which the target appears. There is no specific limit to the number of images; in a specific application scenario, four images can be used. Through the above mosaic enhancement operation, the model can be effectively prevented from overfitting to a specific background.

[0038] In practical applications, the specific geometric enhancement operations are as follows: random rotation (angle range ±15°), random scaling (ratio 0.8-1.2 times), random translation, and random horizontal flipping.

[0039] The specific steps for performing weather simulation enhancement operations are as follows: An atmospheric scattering model is used to overlay a simulated fog layer of random concentration onto the image. The formula is I'(x) = I(x)·t+ A·(1-t), where t is the transmittance and A is the atmospheric light value, simulating low visibility conditions such as morning fog and haze in the wild.

[0040] It should be noted that data augmentation methods can be substituted and combined as follows: In addition to haze simulation, the weather simulation augmentation steps mentioned above can be replaced or supplemented with rain streak overlay simulation, low-light enhancement reverse process simulation, motion blur simulation, etc.; occlusion simulation can use random grid occlusion (GridMask) instead of Cutout. Mosaic enhancement can be replaced with hybrid enhancement strategies such as MixUp or CutMix.

[0041] In each iteration batch of model training, online data augmentation is performed on the input raw images and labeled information to exponentially expand the effective sample space.

[0042] Step S105: Obtain the initial target detection model.

[0043] In a specific application scenario, the initial object detection model can be the YOLOv8 object detection model; this is just an example.

[0044] It should be noted that the object detection model can be replaced: In the recognition method provided in this application embodiment, the YOLOv8 model can be replaced with other models with similar end-to-end object detection capabilities, such as YOLOv9, YOLOv10, PP-YOLOE+, RT-DETR, D-FINE, etc. Any real-time object detection model based on convolutional neural networks or Transformer architecture, as long as it is combined with the "pole-network integrated" annotation strategy and few-shot training method described in the recognition method provided in this application embodiment, should be considered to fall within the scope of protection of this application.

[0045] Step S106: Fine-tune the initial target detection model based on the enhanced data to obtain the fine-tuned target detection model, and use the fine-tuned target detection model to identify bird-catching nets in the wild.

[0046] In one example, the initial object detection model is fine-tuned based on the enhanced data, including the following steps: Based on the augmented data, the following operations are performed sequentially to conduct the first phase of fine-tuning training: Freeze the weight parameters of the first N residual blocks or convolutional layers in the backbone network of the initial object detection model so that the model does not participate in gradient updates during training, where N is a natural number greater than or equal to 1. Only the parameters of the neck network and the head part of the initial target detection model are trained to adapt the model to the high-level semantic features corresponding to the bird-catching net facility; Based on the augmented data, the following operations are performed sequentially to conduct the second phase of fine-tuning training: Once the loss function corresponding to the first stage is detected to have converged, the weight parameters of all layers of the backbone network are unfrozen, and a differentiated learning rate is set so that the model can learn the unique texture, geometry, and spatial combination relationship of the pole network facility while preserving general features.

[0047] It should be noted that the fine-tuning training in the first stage enables the model to quickly adapt to the high-level semantic features of the specific target category of bird-catching net facilities, while avoiding the destruction of the underlying general features when the sample size is limited.

[0048] Through the fine-tuning training in the second stage, the model can fully learn the unique texture, geometry, and spatial combination relationships of pole and grid facilities while retaining general features.

[0049] In the identification method provided in this application embodiment, a two-stage fine-tuning strategy is adopted: in the first stage, some layers of the backbone network are frozen and only the neck and detection head are trained; in the second stage, all layers are unfrozen and global fine-tuning is performed using a differential learning rate to realize the training process of the bird net detection model under small sample conditions.

[0050] It should be noted that the transfer learning strategy corresponding to the first and second stages of fine-tuning training described above can be replaced by a layer-by-layer unfreezing fine-tuning strategy. That is, only one layer or one residual block of the backbone network is unfrozen at a time for training, progressing gradually from the top to the bottom. This strategy provides a more refined training process and is also suitable for small sample scenarios.

[0051] In one example, setting a differential learning rate includes the following steps: Set the learning rate corresponding to the shallow layer of the backbone network of the initial object detection model as the first learning rate; Set the learning rate corresponding to the deep layer of the ancient port network of the initial object detection model as the second learning rate; Set the learning rate of the neck network and the detection head of the initial object detection model as the third learning rate; Set the global learning rate, which finely adjusts all parameters, as the fourth learning rate; The fourth learning rate, the first learning rate, the second learning rate, and the third learning rate increase in that order.

[0052] In practical applications, the fourth learning rate is set to 1×10. -5 Up to 1×10 -4And set the first learning rate to 0.01-0.1 times the base learning rate, which is just an example.

[0053] In practical applications, the process of setting differentiated learning rates is as follows: Set differentiated learning rates—use lower learning rates for shallower layers of the backbone network (e.g., 0.01-0.1 times the base learning rate), medium learning rates for deeper layers of the backbone network, and higher learning rates for the neck network and detector head; use a smaller global learning rate (e.g., 1×10⁻⁶). -5 Up to 1×10 -4 Fine-tuning all parameters is performed; this is just an example.

[0054] In one example, before fine-tuning the initial object detection model based on the enhanced data, the recognition method provided in this application embodiment further includes the following steps: The initial object detection model is loaded and pre-trained on a large-scale object detection dataset to obtain the backbone network weights. The backbone network weights are used as initial parameters to enable the initial object detection model to have general object feature extraction performance.

[0055] It should be noted that large-scale object detection datasets can be large-scale datasets such as COCO.

[0056] In one example, the identification method provided in this application embodiment further includes the following steps: After identifying bird-catching nets in the wild based on the fine-tuned target detection model, the detection results are divided into three levels according to the target status category (normal / damaged / bird-caught) and confidence score output by the model: Level 1 alarm (background record), Level 2 alarm (labeled prompt), and Level 3 alarm (emergency push), and corresponding differentiated response actions are executed.

[0057] Through the above operations, tiered alarms can be implemented for bird net patrols.

[0058] It can also provide precise intelligence for law enforcement and protection operations: through the target status (normal / damaged / bird-hung) output by the model and the hierarchical alarm mechanism, the most suspicious and most dangerous target locations are automatically selected, and alarm information with geographical coordinates and on-site images is pushed to protection agencies or law enforcement personnel in real time to guide precise net clearing operations.

[0059] The method for identifying wild bird-catching nets provided in this application can accurately identify wild bird-catching nets. Furthermore, this method integrates multiple levels of online data augmentation techniques, such as weather simulation enhancement, occlusion simulation enhancement, and mosaic enhancement, combined with a two-stage transfer learning fine-tuning strategy based on COCO pre-trained weights. This allows for the training of a highly generalized detection model with only a few hundred real samples, demonstrating strong adaptability to practical deployment and significantly reducing the model's dependence on labeled data. Moreover, the method for identifying wild bird-catching nets provided in this application is fully automated from raw aerial image input to alarm event output, without relying on specific hardware platforms or communication architectures. It can be flexibly deployed on various ground servers, vehicle-mounted computing units, or cloud inference environments, exhibiting good portability and versatility, ultimately achieving end-to-end, independently deployable algorithms.

[0060] In the above embodiments, a method for identifying wild bird-catching nets is provided. Correspondingly, this application also provides a device for identifying wild bird-catching nets. The device for identifying wild bird-catching nets provided in this application can implement the above-described method for identifying wild bird-catching nets. This device can be implemented through software, hardware, or a combination of both. For example, the device can include integrated or separate functional modules or units to perform the corresponding steps in the above methods.

[0061] Please refer to Figure 2 This illustration shows a schematic diagram of an identification device for a bird-catching net in the wild, provided by some embodiments of this application. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple; relevant details can be found in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0062] like Figure 2 As shown, the identification device 200 for wild bird-catching nets may include: The first acquisition module 201 is used to acquire target images under various conditions. The target images are images obtained by the UAV taking aerial photos of key areas of the migratory bird flyway. The target images include images of bird-catching net facilities. The first annotation module 202 is used to annotate the target image based on preset annotation rules to obtain the corresponding rectangular annotation box. The preset annotation rules include: treating the bird-catching net surface and the support pole used to set up the bird-catching net as a whole target object, and defining the category of the target object as bird-catching net facility. The second annotation module 203 is used to annotate the net area of ​​the bird-catching net in the target image and all the support poles used to fix the bird-catching net in the target image through a rectangular annotation box to obtain the target dataset. The target dataset includes: the target image and the annotation file containing annotation information. All support poles include: bamboo poles, wooden sticks and metal pipes. The data augmentation module 204 is used to use the data in the target dataset as training data, and to perform data augmentation operations on the target images and annotation information in the training data during each iteration of model training to obtain the corresponding augmented data. The second acquisition module 205 is used to acquire the initial target detection model; The fine-tuning training module 206 is used to fine-tune the initial target detection model based on the enhanced data to obtain the fine-tuned target detection model, and then use the fine-tuned target detection model to identify bird-catching nets in the wild.

[0063] In some embodiments of this application, the identification device for the wild bird-catching net further includes: Loading modules (in) Figure 2 (not shown in the image) is used to: load the initial object detection model for pre-training on a large-scale object detection dataset to obtain the backbone network weights before fine-tuning the initial object detection model based on the enhanced data; The backbone network weights are used as initial parameters to enable the initial object detection model to have general object feature extraction performance.

[0064] In some embodiments of this application, the fine-tuning training module 206 is specifically used for: Based on the augmented data, the following operations are performed sequentially to conduct the first phase of fine-tuning training: Freeze the weight parameters of the first N residual blocks or convolutional layers in the backbone network of the initial object detection model so that the model does not participate in gradient updates during training, where N is a natural number greater than or equal to 1. Only the parameters of the neck network and the head part of the initial target detection model are trained to adapt the model to the high-level semantic features corresponding to the bird-catching net facility; Based on the augmented data, the following operations are performed sequentially to conduct the second phase of fine-tuning training: Once the loss function corresponding to the first stage is detected to have converged, the weight parameters of all layers of the backbone network are unfrozen, and a differentiated learning rate is set so that the model can learn the unique texture, geometry, and spatial combination relationship of the pole network facility while preserving general features.

[0065] In some embodiments of this application, the identification device for the wild bird-catching net further includes: Settings module (in) Figure 2 (Not shown in the image) is used to set the differential learning rate. Setting the differential learning rate specifically includes the following steps: Set the learning rate corresponding to the shallow layer of the backbone network of the initial object detection model as the first learning rate; Set the learning rate corresponding to the deep layer of the ancient port network of the initial object detection model as the second learning rate; Set the learning rate of the neck network and the detection head of the initial object detection model as the third learning rate; Set the global learning rate, which finely adjusts all parameters, as the fourth learning rate; The fourth learning rate, the first learning rate, the second learning rate, and the third learning rate increase in that order.

[0066] In some embodiments of this application, the identification device for the wild bird-catching net further includes: Data augmentation module 204 is specifically used for: Perform data augmentation operations on the target images and annotation information in the training data. The data augmentation operations include at least one or more random combinations of the following: Perform geometry enhancement operations by executing random rotation, random scaling, random translation, and random horizontal flipping. A luminance enhancement operation is performed by randomly perturbing the brightness, contrast, and saturation values ​​with preset values ​​in the HSV color space. Weather simulation enhancement is performed by overlaying simulated fog layers of random concentration onto images based on an atmospheric scattering model. By randomly selecting one or more rectangular regions in the image using a random erasure method, and configuring their pixel values ​​as random noise or a constant, the operation of occlusion simulation enhancement is performed. The various occlusion scenarios include: the support pole being partially occluded by branches and shrubs, or the net being partially occluded by bird carcasses. Mosaic enhancement is performed by randomly selecting multiple training images with different backgrounds, scaling and cropping them respectively, and then stitching the scaled and cropped training images together to obtain a new image; and adjusting the coordinates of the corresponding bounding boxes based on the new image to enrich the background context in which the target appears.

[0067] In some embodiments of this application, the second annotation module 203 is specifically used for: If the same bird-catching net facility uses multiple support poles, then each of the multiple support poles must serve the same bird-catching net, and each of the multiple support poles must be included in the same label box.

[0068] In some embodiments of this application, the first acquisition module 201 is specifically used for: Acquire target image data under different lighting conditions, including: sunny lighting conditions, cloudy lighting conditions, and dawn / dusk lighting conditions; or, Acquire target image data under different seasonal background conditions, including: background conditions during the green period and background conditions during the withered period; or... Acquire target image data under different terrain environmental conditions, including: forest environment conditions, reed marsh environment conditions, shrubland environment conditions, and farmland edge environment conditions.

[0069] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform the above-described method for identifying bird-catching nets in the wild.

[0070] According to another embodiment, an electronic device is also provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the above-described method for identifying bird-catching nets in the wild.

[0071] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this application can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0072] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying bird-catching nets in the wild, characterized in that, The method includes: Acquire target images under various conditions. The target images are images obtained by UAV aerial photography of key areas of migratory bird flyways. The target images include images of bird-catching net facilities. Based on preset annotation rules, the target image is annotated to obtain the corresponding rectangular annotation box. The preset annotation rules include: treating the bird-catching net surface and the support pole used to set up the bird-catching net as a whole target object, and defining the category of the target object as the bird-catching net facility. The target dataset is obtained by annotating the net area of ​​the bird-catching net in the target image and all the support poles used to fix the bird-catching net in the target image using the rectangular annotation box. The target dataset includes: the target image and the annotation file containing annotation information. All support poles include: bamboo poles, wooden sticks and metal pipes. The data in the target dataset is used as training data, and in each iteration of model training, data augmentation operations are performed on the target images and annotation information in the training data to obtain the corresponding augmented data; Obtain the initial object detection model; The initial target detection model is fine-tuned and trained based on the enhanced data to obtain a fine-tuned target detection model, which is then used to identify bird-catching nets in the wild.

2. The identification method according to claim 1, characterized in that, Before fine-tuning the initial object detection model based on the enhanced data, the method further includes: The initial object detection model is loaded and pre-trained on a large-scale object detection dataset to obtain the backbone network weights; The backbone network weights are used as initial parameters to enable the initial object detection model to have general object feature extraction performance.

3. The identification method according to claim 1, characterized in that, The fine-tuning and training of the initial target detection model based on the enhanced data includes: Based on the enhanced data, the following operations are performed sequentially to conduct the first phase of fine-tuning training: Freeze the weight parameters of the first N residual blocks or convolutional layers in the backbone network of the initial target detection model so that the model does not participate in gradient updates during training, where N is a natural number greater than or equal to 1. Only the parameters of the neck network and the head part of the initial target detection model are trained to adapt the model to the high-level semantic features corresponding to the bird-catching net facility; Based on the enhanced data, the following operations are performed sequentially to conduct the second phase of fine-tuning training: Once the loss function corresponding to the first stage is detected to have converged, the weight parameters of all layers of the backbone network are unfrozen, and a differentiated learning rate is set so that the model can learn the unique texture, geometry, and spatial combination relationship of the pole network facility while preserving general features.

4. The identification method according to claim 3, characterized in that, Setting the differentiated learning rate includes: Set the learning rate corresponding to the shallow layer of the backbone network of the initial object detection model as the first learning rate; Set the learning rate corresponding to the deep layer of the ancient port network of the initial object detection model as the second learning rate; Set the learning rate of the neck network and the detection head of the initial object detection model as the third learning rate; Set the global learning rate, which finely adjusts all parameters, as the fourth learning rate; The fourth learning rate, the first learning rate, the second learning rate, and the third learning rate increase sequentially.

5. The identification method according to claim 1, characterized in that, The data augmentation operation performed on the target images and annotation information in the training data includes: Perform data augmentation operations on the target images and annotation information in the training data. The data augmentation operations include at least one or more random combinations of the following: Perform geometry enhancement operations by executing random rotation, random scaling, random translation, and random horizontal flipping. A luminance enhancement operation is performed by randomly perturbing the brightness, contrast, and saturation values ​​with preset values ​​in the HSV color space. Weather simulation enhancement is performed by overlaying simulated fog layers of random concentration onto images based on an atmospheric scattering model. By randomly selecting one or more rectangular regions in the image using a random erasure method, and configuring their pixel values ​​as random noise or a constant, the operation of occlusion simulation enhancement is performed. The various occlusion scenarios include: the support pole being partially occluded by branches and shrubs, or the net being partially occluded by bird carcasses. Mosaic enhancement is performed by randomly selecting multiple training images with different backgrounds, scaling and cropping them respectively, and then stitching the scaled and cropped training images together to obtain a new image; and adjusting the coordinates of the corresponding bounding boxes based on the new image to enrich the background context in which the target appears.

6. The identification method according to claim 1, characterized in that, The step of annotating the net area of ​​the bird-catching net in the target image and all the support rods used to fix the bird-catching net in the target image using the rectangular annotation box includes: If the same bird-catching net facility uses multiple support poles, then each of the multiple support poles is determined to serve the same bird-catching net, and each of the multiple support poles is included in the same label box.

7. The identification method according to claim 1, characterized in that, The acquisition of target image data under various conditions includes: Acquire target image data under different lighting conditions, including: sunny lighting conditions, cloudy lighting conditions, and dawn / dusk lighting conditions; or, Acquire target image data under different seasonal background conditions, including: background conditions during the green period and background conditions during the withered period; or... Acquire target image data under different terrain environmental conditions, including: forest environment conditions, reed marsh environment conditions, shrubland environment conditions, and farmland edge environment conditions.

8. A device for identifying bird-catching nets used in the wild, characterized in that, The device includes: The first acquisition module is used to acquire target images under various conditions. The target images are images obtained by UAV aerial photography of key areas of migratory bird flyways. The target images include images of bird-catching net facilities. The first annotation module is used to annotate the target image based on preset annotation rules to obtain the corresponding rectangular annotation box. The preset annotation rules include: treating the bird-catching net surface and the support pole used to set up the bird-catching net as a whole target object, and defining the category of the target object as the bird-catching net facility. The second annotation module is used to annotate the net area of ​​the bird-catching net in the target image and all the support poles in the target image used to fix the bird-catching net using the rectangular annotation box, so as to obtain the target dataset. The target dataset includes: the target image and the annotation file containing annotation information. All support poles include: bamboo poles, wooden sticks and metal pipes. The data augmentation module is used to use the data in the target dataset as training data, and to perform data augmentation operations on the target images and annotation information in the training data during each iteration of model training to obtain the corresponding augmented data. The second acquisition module is used to acquire the initial target detection model; The fine-tuning training module is used to fine-tune the initial target detection model based on the enhanced data to obtain a fine-tuned target detection model, which is then used to identify bird-catching nets in the wild.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to perform the method described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method according to any one of claims 1 to 7.

Images