Intelligent identification and expulsion method and system for birds at airports

Abstract

The present disclosure provides an intelligent identification and expulsion method and system for airport birds, which extracts a first image pixel representation carrier, a first image semantic representation carrier, and a first image visual representation carrier of an airport target area image, interacts the three to obtain a first interactive representation carrier, obtains the carrier space similarity between the first interactive representation carrier and a second interactive representation carrier corresponding to each airport bird reference image, determines the airport bird reference image corresponding to the airport target area image based on the carrier space similarity, formulates a target expulsion strategy based on the above, and according to the strategy, links a sonic wave repeller and other bird repellent terminals to perform bird repellent operations. This intelligent bird repellent method can extract image elements of different dimensions in an image, characterize the underlying image semantics of the airport target area image, and enable the first image visual representation carrier to reduce invalid information in the airport target area image, thereby increasing the accuracy of image recognition.

Description

Technical Field

[0001] The present disclosure relates to the fields of image processing and machine learning, and in particular to a method and system for intelligently identifying and driving away birds at an airport. Background Art

[0002] With the continuous development of aviation traffic, airport safety issues have received increasing attention. Among them, bird activities within the airport area pose a considerable threat to flight safety. The collision between birds and aircraft may lead to serious safety accidents. Therefore, intelligent identification and timely expulsion of birds at airports are particularly important.

[0003] Traditional methods of driving birds away from airports mostly rely on manual inspections and fixed driving equipment, but this method is inefficient and may not be able to detect and drive birds away in time in some cases. In recent years, with the rapid development of computer vision and machine learning technologies, the use of image recognition technology for automatic identification and driving away of birds at airports has gradually become a research hotspot. However, existing image recognition technology still faces many challenges in identifying birds at airports. Due to the complexity of the airport environment, birds in images may be affected by multiple factors such as lighting, angle, and occlusion, making identification more difficult. In addition, different species of birds have great similarities in appearance, which also makes accurate identification difficult. Improving the accuracy of airport bird identification is crucial to improving the safety management level of airports, reducing the risk of collision between birds and aircraft, and ensuring aviation safety. Summary of the invention

[0004] The purpose of the present disclosure is to provide a method and system for intelligent identification and expulsion of birds at an airport. The embodiments of the present disclosure are implemented as follows:

[0005] In a first aspect, an embodiment of the present disclosure provides a method for intelligently identifying and driving away birds at an airport, the method comprising:

[0006] Performing representation information mining on the airport target area image to obtain a first image pixel representation carrier, a first image semantic representation carrier, and a first image visual representation carrier, wherein the first image pixel representation carrier is used to carry the pixel feature representation of the airport target area image, the first image semantic representation carrier is used to carry the content feature representation of the airport target area image, and the first image visual representation carrier is used to carry the visual feature information of the airport target area image;

[0007] Interacting the first image pixel representation carrier, the first image semantic representation carrier, and the first image visual representation carrier to obtain a first interactive representation carrier;

[0008] Obtaining the vector space similarity between the first interactive representation vector and the second interactive representation vector corresponding to each airport bird control image; wherein the second interactive representation vector is obtained by mining and interacting the representation information of the airport bird control image;

[0009] Determining the airport bird control image corresponding to the airport target area image in each of the airport bird control images through the carrier space similarity;

[0010] A target driving strategy is generated according to the driving strategy corresponding to the determined airport bird control image.

[0011] Optionally, the performing characterization information mining on the airport target area image to obtain the first image pixel characterization carrier includes:

[0012] A plurality of image pixel clusters are obtained through the airport target area image, each image pixel cluster includes a plurality of image pixels, and at least one image pixel having an intersection is included between two consecutive image pixel clusters;

[0013] Determining a node distribution result of each of the plurality of image pixel clusters in a visual knowledge graph, and determining a frequency of existence of each of the plurality of image pixel clusters in the airport target area image, wherein the visual knowledge graph includes a plurality of candidate image pixel clusters;

[0014] The first image pixel representation carrier is obtained through the node distribution results of each image pixel group in the visual knowledge graph and the existence frequency of each image pixel group in the airport target area image.

[0015] Optionally, the performing representation information mining on the airport target area image to obtain the first image semantic representation carrier includes:

[0016] Acquire an image pixel matrix corresponding to the airport target area image, wherein the image pixel matrix includes a plurality of image pixels obtained by extracting pixel values ​​of the airport target area image and arranging them according to original positions;

[0017] The mutual information of adjacent elements of each image pixel in the image pixel matrix is ​​extracted to obtain the first image semantic representation carrier.

[0018] Optionally, the performing representation information mining on the airport target area image to obtain a first image visual representation carrier includes:

[0019] Acquire visual feature information corresponding to each image pixel in the airport target area image;

[0020] A plurality of visual feature information sets are obtained through the visual feature information corresponding to each image pixel in the airport target area image, each visual feature information set includes visual feature information corresponding to a plurality of image pixels, and at the same time, two consecutive visual feature information sets contain at least one visual feature information corresponding to an image pixel having an intersection;

[0021] Determine a node distribution result of each of the several visual feature information sets in a visual feature information visual knowledge graph, and determine an existence frequency of each of the several visual feature information sets in the airport target area image, wherein the visual feature information visual knowledge graph includes several candidate visual feature information sets;

[0022] The first image visual representation carrier is obtained through the node distribution results of each visual feature information set in the visual feature information visual knowledge graph and the existence frequency of each visual feature information set in the airport target area image.

[0023] Optionally, before obtaining the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image, the method further includes:

[0024] filtering the airport bird control images in the airport bird control image library based on the preliminary similarity between the airport target area image and the airport bird control images in the airport bird control image library to obtain a candidate airport bird control image library, wherein the candidate airport bird control image library includes a plurality of airport bird control images determined from the airport bird control image library;

[0025] The obtaining of the vector space similarity between the first interactive representation vector and the second interactive representation vector corresponding to each airport bird control image includes:

[0026] The carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image in the candidate airport bird control image library is obtained.

[0027] Optionally, the airport target area image is subjected to representation information mining based on a first representation information mining algorithm, and the airport bird control image is subjected to representation information mining based on a second representation information mining algorithm, and the first representation information mining algorithm and the second representation information mining algorithm are two different representation information mining algorithms; wherein the first representation information mining algorithm and the second representation information mining algorithm have the same algorithm composition, but correspond to different algorithm parameters;

[0028] The first characterization information mining algorithm and the second characterization information mining algorithm are trained based on the following operations:

[0029] Acquire a plurality of training image template groups, each of which includes an airport target area image template and an airport bird control image template;

[0030] Based on the first representation information mining algorithm, the representation information mining is performed on the airport target area image template to obtain a first image pixel representation carrier, a first image semantic representation carrier and a first image visual representation carrier;

[0031] The first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier are interacted to obtain a first interactive representation carrier; wherein the first image pixel representation carrier is used to carry the pixel feature representation of the airport target area image template, the first image semantic representation carrier is used to carry the content feature representation of the airport target area image template, and the first image visual representation carrier is used to carry the visual feature information of the airport target area image template;

[0032] Based on the second representation information mining algorithm, the representation information mining is performed on the airport bird control image template to obtain a second image pixel representation carrier, a second image semantic representation carrier and a second image visual representation carrier;

[0033] The second image pixel representation carrier, the second image semantic representation carrier and the second image visual representation carrier are interacted to obtain a second interactive representation carrier; wherein the second image pixel representation carrier is used to carry the pixel feature representation of the airport bird comparison image template, the second image semantic representation carrier is used to carry the content feature representation of the airport bird comparison image template, and the second image visual representation carrier is used to carry the visual feature information of the airport bird comparison image template;

[0034] Performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain an algorithm training cost;

[0035] The parameters of the first characterization information mining algorithm and the second characterization information mining algorithm are corrected through the algorithm training cost, and the trained first characterization information mining algorithm is used to perform characterization information mining and interaction on the airport target area image to determine the airport bird control image corresponding to the airport target area image.

[0036] Optionally, the plurality of training image template groups include: a training image true template group, a simple training image false template group and a complex training image false template group; wherein the airport bird control image template included in the training image true template group is a labeled airport bird control image of the airport target area image template;

[0037] The airport bird control image template included in the simple training image false template group is randomly sampled from an airport bird control image library and is not a labeled airport bird control image of the airport target area image template;

[0038] The representation carrier spatial similarity between the airport bird control image template and the airport target area image template included in the complex training image false template group is not less than a preset spatial similarity coefficient, and is not a labeled airport bird control image of the airport target area image template, wherein the representation carrier spatial similarity is the carrier spatial similarity between the first interactive representation carrier of the airport target area image template and the second interactive representation carrier of the airport bird control image template.

[0039] Optionally, the method further comprises:

[0040] Acquire the representation carrier space similarity between the airport target area image template and each airport bird control image in the airport bird control image library;

[0041] The M airport bird control images other than the annotated airport bird control images of the airport target area image template, which have the highest spatial similarity in the representation carrier, are used as airport bird control image templates to form M false template groups of the complex training image together with the airport target area image template, where M is a non-zero natural number.

[0042] Optionally, performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain an algorithm training cost includes:

[0043] Performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain a first cost variable;

[0044] performing similarity learning on the first image pixel representation carrier and the second image pixel representation carrier to obtain a second cost variable;

[0045] performing similarity learning on the first image semantic representation carrier and the second image semantic representation carrier to obtain a third cost variable;

[0046] performing similarity learning on the first image visual representation carrier and the second image visual representation carrier to obtain a fourth cost variable;

[0047] The algorithm training cost is obtained through the first cost variable, the second cost variable, the third cost variable and the fourth cost variable.

[0048] In a second aspect, the present disclosure provides a computer system, comprising:

[0049] one or more processors;

[0050] Memory;

[0051] One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method provided in the first aspect above is implemented.

[0052] The beneficial effects of the present disclosure include at least: the airport bird intelligent identification and expulsion method and system provided by the present disclosure, by extracting the first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier of the airport target area image, the three are interacted to obtain the first interactive representation carrier, the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image is obtained, and the airport bird control image corresponding to the airport target area image is determined based on the carrier space similarity. In this way, image elements of different dimensions in the image can be extracted, the bottom image semantics of the airport target area image can be represented based on the extraction of the first image pixel representation carrier, the context and hidden upper semantics of the airport target area image can be represented based on the extraction of the first image semantic representation carrier, and the extraction of the first image visual representation carrier can reduce the invalid information in the airport target area image, so that the accuracy of image recognition is increased. The first interactive representation carrier of the airport target area image is obtained by interacting with each of the above representation carriers, and the carrier space similarity between the airport target area image and the airport bird control image is compared at three levels, so as to further improve the accuracy of bird recognition and improve the bird expulsion effect.

[0053] Other features will be described in part in the following description. Those skilled in the art will discover these features in part upon inspection of the following content and drawings, or may learn these features through production or use. The features of the present application may be implemented and obtained by practicing or using various aspects of the methods, tools, and combinations listed in the detailed examples described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings required for describing the embodiments of the present disclosure are briefly introduced below.

[0055] Figure 1It is a flow chart of an intelligent method for identifying and driving away birds at an airport provided by an embodiment of the present disclosure.

[0056] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present disclosure.

[0057] Reference numerals: 100 - computer system; 101 - processor; 102 - bus; 103 - memory; 104 - transceiver. DETAILED DESCRIPTION

[0058] The following describes the embodiments of the present disclosure in conjunction with the drawings in the embodiments of the present disclosure. The terms used in the implementation method part of the embodiments of the present disclosure are only used to explain the specific embodiments of the present disclosure, and are not intended to limit the present disclosure.

[0059] The executor of the method for intelligent identification and expulsion of birds at airports in the embodiments of the present disclosure is a computer system, including but not limited to servers, personal computers, laptops, tablet computers, smart phones, etc. The server includes but is not limited to a single network server, a server group consisting of multiple network servers, or a cloud consisting of a large number of computers or network servers in cloud computing, wherein cloud computing is a type of distributed computing, a super virtual computer consisting of a group of loosely coupled computer sets. The computer system can be operated alone to implement the present disclosure, or it can be connected to the network and implement the present disclosure through interactive operations with other computer systems in the network. The network in which the computer system is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, VPN network, etc.

[0060] The embodiment of the present disclosure provides an intelligent identification and expulsion method for birds at an airport, which is applied to a computer system, such as Figure 1 As shown, the method includes:

[0061] Step S100: Perform representation information mining on the airport target area image to obtain a first image pixel representation carrier, a first image semantic representation carrier and a first image visual representation carrier, wherein the first image pixel representation carrier is used to carry the pixel feature representation of the airport target area image, the first image semantic representation carrier is used to carry the content feature representation of the airport target area image, and the first image visual representation carrier is used to carry the visual feature information of the airport target area image.

[0062] In step S100, specifically, the computer system first acquires an image of the airport target area, which is usually achieved by installing cameras in key areas of the airport. Once the image is acquired, the computer system begins to mine representation information, which can be divided into three main parts: pixel feature extraction, semantic feature extraction, and visual feature extraction.

[0063] In terms of pixel feature extraction, the computer system analyzes each pixel of the image to capture basic information such as color, brightness, contrast, etc. For example, if there is a group of white pigeons on a green lawn in the image, the pixel features will include a large number of green and white pixels and their distribution patterns. These pixel features can be represented as a vector or matrix, the first image pixel representation carrier, which records in detail the intensity distribution of each color channel in the image.

[0064] Next comes semantic feature extraction, which is a more complex step that involves understanding the content of the image. Computer systems can use deep learning models, such as convolutional neural networks (CNNs), to identify objects in an image and their relationships. In an airport environment, this may include identifying key elements such as runways, airplanes, birds, buildings, etc. With a well-trained model, the computer system is able to extract high-level features in the image, which constitute the first image semantic representation carrier, which describes the entities in the image and their attributes.

[0065] Finally, visual feature extraction involves low-level features of the image such as texture, shape, and edges. These features may be particularly important for distinguishing different types of birds, as different birds differ in feather texture, body shape, and other aspects. Computer systems use image processing techniques, such as the Sobel operator or Canny edge detection, to extract these visual features. These features are encoded as first-image visual representation carriers that provide detailed information about the appearance of objects in the image.

[0066] Step S100 is a multi-level feature extraction process that combines traditional image processing techniques and advanced machine learning algorithms to fully mine the representation information of the airport target area image. This information provides a solid foundation for the subsequent bird identification and driving away strategy formulation.

[0067] Step S200: interacting the first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier to obtain a first interactive representation carrier.

[0068] Step S200 interacts or fuses the three different representation carriers extracted in the previous step to generate a comprehensive representation carrier, namely, a first interactive representation carrier.

[0069] In this step, the computer system uses a specific algorithm or model to integrate the first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier extracted from the airport target area image. These representation carriers carry the pixel feature, content feature and visual feature information of the image respectively.

[0070] Specifically, the first image pixel representation vector may include an array or matrix describing the color and brightness of each pixel in the image. For example, in an image of an airport runway, this vector can capture information such as the color of the runway, the position and color of the marking line, etc.

[0071] The first image semantic representation carrier is more abstract. It may be extracted through a deep learning model (such as a convolutional neural network CNN) and reflects the entities and concepts in the image. In an airport scene, this carrier may contain the recognition results of objects such as airplanes, birds, runways, buildings, and the relationships between them.

[0072] The first image visual representation carrier focuses on the visual features of the image, such as texture, shape, and edges. These features can be extracted through image processing technology and are particularly important for identifying different types of birds, because the visual features of birds such as feather texture and body shape are the key to their classification.

[0073] In step S200, the computer system uses a fusion algorithm, such as feature concatenation, weighted averaging, or a more complex machine learning model, to fuse these different levels of features together. The purpose of this is to create a more comprehensive and information-rich representation carrier, namely the first interactive representation carrier. This new representation carrier can comprehensively consider the pixel details, semantic content, and visual features of the image, providing strong support for bird identification and driving away strategy formulation in subsequent steps.

[0074] For example, if there is a bird parked on a runway in the image, through the fusion processing of step S200, the first interactive representation carrier will be able to comprehensively reflect the feather color, body outline and scene (runway) of the bird, thereby helping the computer system to more accurately identify the type of the bird in subsequent steps and formulate corresponding driving away strategies.

[0075] Step S300: obtaining the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image; wherein the second interactive representation carrier is obtained by performing representation information mining and interaction on the airport bird control image.

[0076] Step S300 calculates the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier of the pre-stored airport bird control image, that is, the similarity between the two in the feature space, such as the similarity in the vector space, so as to determine the bird species in the target image.

[0077] In this step, the computer system first accesses a database containing a variety of comparison images of common birds at airports. Each image in this database has undergone the same processing flow as steps S100 and S200, that is, representation information mining and interaction have been performed to generate corresponding second interactive representation carriers. These second interactive representation carriers are structurally consistent with the first interactive representation carriers, and both integrate feature information at the three levels of pixels, semantics, and vision. Next, the computer system uses a specific similarity calculation algorithm, such as cosine similarity, Euclidean distance, etc., to compare the similarity between the first interactive representation carrier and each second interactive representation carrier in the database one by one. This similarity calculation process is performed in a high-dimensional feature space, in which each dimension represents a feature of a certain aspect of the image.

[0078] For example, if the target image captures a pigeon walking on an airport runway, the first interactive representation carrier obtained after processing in steps S100 and S200 will contain characteristic information such as the color, texture, shape, and scene of the pigeon. In step S300, the computer system compares the first interactive representation carrier with the second interactive representation carriers of all birds in the database one by one. By calculating the similarity, the computer system can find the reference image that is most similar to the pigeon in the target image, thereby determining the type of the pigeon.

[0079] The key to this process lies in the accuracy and efficiency of similarity calculation. In order to improve accuracy, the computer system can use a combination of multiple similarity calculation methods, or further optimize and select features. At the same time, in order to improve efficiency, the computer system can use some fast search algorithms or parallel computing technologies to accelerate the matching process. Finally, through the similarity calculation and matching of step S300, the computer system can accurately identify the bird species in the airport target area image, providing a key information basis for the subsequent expulsion strategy formulation.

[0080] Step S400: determining the airport bird reference image corresponding to the airport target area image from each of the airport bird reference images through the carrier space similarity.

[0081] Step S400 determines the bird reference image that best matches the target area image from the plurality of airport bird reference images based on the carrier spatial similarity calculated in the previous step. In this step, the computer system has calculated the similarity between the first interactive representation carrier and the second interactive representation carrier of each airport bird reference image through step S300. These similarity values ​​reflect the similarity between the birds in the target area image and the birds in the reference image.

[0082] Now, the computer system needs to make a decision based on these similarity values ​​to determine which reference image is most similar to the bird in the target area image. This is usually done by comparing the similarity values. The computer system goes through all the similarity values ​​and finds the maximum value. The reference image corresponding to this maximum value is the reference image of the bird that best matches the target area image.

[0083] For example, if the similarity value calculated in step S300 shows that the birds in the target area image have the highest similarity to a pigeon reference image in the database, then in step S400, the computer system determines that the pigeon reference image is the airport bird reference image corresponding to the target area image.

[0084] The key to this process is the accurate calculation and comparison of similarity values. Only when the features of the image are accurately extracted and the similarity is correctly calculated in the previous steps can the correct decision be made in step S400. Through the decision of step S400, the computer system can accurately identify the bird species in the airport target area image and find the most matching control image for it. This provides important reference information for the subsequent expulsion strategy formulation, allowing airport managers to take corresponding measures according to actual conditions to ensure the safe operation of the airport.

[0085] Step S500: generating a target driving away strategy according to the driving away strategy corresponding to the determined airport bird control image.

[0086] Step S500 selects the corresponding driving strategy according to the determined airport bird control image, and finally generates a specific driving plan for the target bird. In this step, the computer system first accesses a pre-set driving strategy database. This database stores effective driving methods for different types of birds, which may be sound, light, physical barriers or other forms of driving means. Each type of bird may correspond to one or more effective driving strategies, which are based on previous practical experience and scientific research.

[0087] Once the bird species in the airport target area image is determined (through the previous step S400), the computer system searches for the corresponding optimal repelling method in the repelling strategy database according to the characteristics of the bird. For example, if the identified bird is a pigeon, the computer system may choose to play a sound of a certain frequency to repel it, because this sound has been proven to have a repelling effect on pigeons.

[0088] Next, the computer system generates a specific target expulsion strategy based on the expulsion strategy found and the actual conditions at the airport (such as weather, time, airport layout, etc.). This strategy will specify what expulsion means to use, when and where to implement them, and the frequency and intensity of implementation.

[0089] For example, if a group of sparrows is identified gathering near an airport runway, the computer system can generate a targeted repellent strategy that includes the following key points: 1) Use a high-frequency sound wave repellent; 2) Set up the repellent near the area where the sparrows gather; 3) Turn on the repellent at regular intervals every day for a certain period of time; 4) Adjust the frequency and intensity of the repellent according to the sparrows' reactions.

[0090] Finally, the computer system outputs the generated target repelling strategy to airport managers for their execution. In this way, the airport can take scientific and effective repelling measures according to the actual situation, reduce the impact of birds on airport operations, and ensure flight safety. For example, managers can link sonic repellents and other bird repelling terminals to carry out bird repelling operations based on this strategy.

[0091] As an implementation manner, in step S100, performing representation information mining on the airport target area image to obtain a first image pixel representation carrier includes:

[0092] Step S110: obtaining a plurality of image pixel clusters through the airport target area image, each image pixel cluster includes a plurality of image pixels, and at least one image pixel having an intersection is contained between two consecutive image pixel clusters.

[0093] Step S110 is a pre-processing step in the image processing process of the airport bird intelligent identification and expulsion scheme. In this step, the computer system analyzes the airport target area image to form a number of image pixel groups.

[0094] Specifically, the computer system first scans the entire airport target area image and divides it into multiple small image areas, which are called image pixel clusters. Each image pixel cluster is composed of several image pixels, which have certain similarities in color. For example, one image pixel cluster may contain a group of grass pixels with similar colors and shapes, while another image pixel cluster may be composed of a group of runway pixels with similar brightness and orderly arrangement. The division of these image pixel clusters is not completely independent, but there is a certain intersection, that is, overlap. In other words, there will be at least one or more overlapping image pixels between two consecutive image pixel clusters. This design is to ensure the continuity and integrity of image information and avoid the loss of important detail information during the division process.

[0095] For example, assume that the airport target area image includes an airplane taking off and the surrounding runway, grass and other environments. In step S110, the computer system can divide the wings, fuselage, runway and grass of the airplane into different image pixel groups. At the same time, in order to ensure the continuity of information, there will be overlapping pixels between the image pixel groups of adjacent parts such as the wing and fuselage, the fuselage and the runway.

[0096] In this way, the computer system can provide richer and more accurate image information for subsequent feature extraction and recognition steps, thereby improving the accuracy and efficiency of bird identification at airports.

[0097] Step S120: Determine the node distribution result of each of the several image pixel clusters in the visual knowledge graph, and determine the existence frequency of each of the several image pixel clusters in the airport target area image, wherein the visual knowledge graph includes several candidate image pixel clusters.

[0098] Step S120 connects the original image information with subsequent feature extraction and recognition. In this step, the computer system needs to complete two main tasks: one is to determine the node distribution results of each image pixel group in the Visual Knowledge Graphs (VKG), and the other is to count the frequency of each image pixel group in the airport target area image. Regarding the visual knowledge graph, it is a pre-built data structure that contains a large number of candidate image pixel groups and their relationships. These candidate image pixel groups may be extracted from a large amount of image data and represent a variety of different image features and patterns. Each node in the graph corresponds to a specific image pixel group, and the connection between the nodes reflects the association and similarity between these image pixel groups.

[0099] In step S120, the computer system matches each image pixel cluster obtained in step S110 with the candidate image pixel cluster in the visual knowledge graph. This matching process may be based on the comparison of image color features. Once a matching node is found, it means that the image pixel cluster has found the corresponding position in the graph, thereby determining its node distribution result in the graph.

[0100] For example, suppose there is an image pixel cluster in the airport target area image that represents a patch of grass. In the visual knowledge graph, the computer system can find a matching node that also represents grass and is connected to other nodes representing natural elements such as trees and flowers. Secondly, the computer system also counts the frequency of each image pixel cluster in the airport target area image, which can be counted by the number of occurrences. This frequency can reflect the importance and commonness of the image pixel cluster in the image. For example, if an image pixel cluster (such as a runway) appears frequently in an image, then its frequency of existence will be relatively high.

[0101] Step S120 not only helps the computer system understand the semantic information of each part in the image (through the node distribution of the visual knowledge graph), but also provides statistical data on the frequency of occurrence of these parts in the image. This information will provide strong support for subsequent feature extraction, bird identification and driving away strategy formulation.

[0102] Step S130: Obtain the first image pixel representation carrier through the node distribution results of each image pixel group in the visual knowledge graph and the existence frequency of each image pixel group in the airport target area image.

[0103] In step S130, the computer system will use the information obtained in the previous step, that is, the node distribution results of each image pixel group in the visual knowledge graph and the frequency of each image pixel group in the airport target area image, to generate a first image pixel representation carrier. The first image pixel representation carrier is a data structure that can fully reflect the image features. It integrates the semantic information and frequency information of the image pixel group, and provides an important data basis for the subsequent bird identification and driving away strategy formulation.

[0104] Specifically, the computer system extracts the semantic features associated with each image pixel cluster based on the node distribution in the visual knowledge graph. For example, if a certain image pixel cluster corresponds to the "grassland" node in the graph, the computer system will extract features related to the grass, such as color. These features can help the system better understand the image content and improve the accuracy of bird recognition.

[0105] At the same time, the computer system also considers the frequency of each image pixel cluster in the image. Image pixel clusters with high frequencies often correspond to the main content or important features in the image, so they are given greater weight when generating the representation carrier. For example, in an image of an airport runway, the image pixel clusters of the runway appear more frequently, so in the representation carrier, features related to the runway will be highlighted.

[0106] Finally, by comprehensively considering the semantic features and the frequency of existence of the image pixel clusters, the computer system is able to generate a comprehensive and accurate first image pixel representation vector, which not only contains the visual information of the image, but also reflects the importance and relationship between the various parts of the image.

[0107] As an implementation manner, in step S100, performing representation information mining on the airport target area image to obtain a first image semantic representation carrier includes:

[0108] Step S140: obtaining an image pixel matrix corresponding to the airport target area image, wherein the image pixel matrix includes a plurality of image pixels obtained by extracting pixel values ​​of the airport target area image and arranging them according to original positions;

[0109] Step S150: extracting the mutual information of adjacent elements of each image pixel in the image pixel matrix to obtain the first image semantic representation carrier.

[0110] Step S140 is one of the key steps in the airport bird intelligent identification and expulsion scheme, which involves obtaining an image pixel matrix of the airport target area image. The purpose of this step is to convert the image into a format that can be more easily processed and analyzed by a computer.

[0111] In step S140, the computer system first reads the image data of the airport target area. This image can be a real-time picture captured by a surveillance camera, or it can be an archived image extracted from an image database. In either case, the computer system scans the image point by point and extracts the pixel value of each pixel. The pixel value is the basic component unit of the image, which represents the color and brightness information of each point in the image. In digital image processing, the pixel value is usually represented by a number or a group of numbers. For example, in a grayscale image, the value of each pixel may be an integer between 0 (black) and 255 (white). In a color image, the value of each pixel may be composed of three components (red, green, and blue), and the value range of each component is also 0 to 255.

[0112] After extracting all the pixel values, the computer system arranges them into a two-dimensional array, the image pixel matrix, according to their positions in the original image. Each element of this matrix corresponds to a pixel in the image, and the value of the element is the pixel value of the pixel.

[0113] For example, suppose we have a simple grayscale image of an airport runway with only black and white colors in the image, then its image pixel matrix is ​​as follows:

[0114] [255, 255, 255, 0, 0],

[0115] [255, 255, 0, 0, 255],

[0116] [0, 0, 0, 255, 255],

[0117] [0, 0, 255, 255, 0],

[0118] [255, 255, 255, 0, 0]

[0119] In this matrix, the value 255 represents white pixels (for example, the runway part), and the value 0 represents black pixels (for example, the background or other non-runway parts). Through such a matrix, the computer system can easily perform various complex mathematical operations and analyses on the image, thereby providing basic data for subsequent bird identification and driving away strategies.

[0120] As an implementation manner, in step S100, performing representation information mining on the airport target area image to obtain a first image visual representation carrier includes:

[0121] Step S160: Obtain visual feature information corresponding to each image pixel in the airport target area image.

[0122] The goal of step S160 is to extract the visual feature information corresponding to each image pixel in the airport target area image. In this step, the computer system scans the input airport target area image pixel by pixel and extracts its visual feature information for each pixel. These visual features include but are not limited to texture, color and brightness. Specifically: texture features describe the surface structure and arrangement rules of pixels and their surrounding areas. For example, on an airport runway, the surface texture of the runway may be completely different from the texture of grass or apron. The computer system can extract texture features by calculating grayscale co-occurrence matrix, autocorrelation function and other methods within the pixel neighborhood. Color features are a common feature in image recognition, which describes the color distribution and properties of objects in the image. In the airport bird recognition scene, different types of birds may have different feather colors. The computer system can convert the image from RGB color space to other color spaces (such as HSV or Lab color space) to more accurately describe and extract color features. Brightness features reflect the brightness of pixels in the image. In an airport environment, brightness information may help distinguish between birds in the sky, on the ground and in flight. The computer system can calculate the brightness value of each pixel and generate features such as brightness histogram or brightness gradient.

[0123] For example, suppose the computer system is processing an image of a flock of birds flying over an airport runway. In step S160, the computer system analyzes the pixels in the image one by one. For the pixels in the bird area, the computer system extracts the texture (such as the fineness of the feathers), color (such as the multiple colors of the bird feathers and their distribution) and brightness (such as the light and dark changes of the bird flock in the sun) of these pixels. These features will be used for subsequent bird identification and the formulation of driving away strategies.

[0124] By extracting visual feature information in step S160, the computer system can obtain rich image description data, providing a solid foundation for subsequent image analysis and processing.

[0125] Step S170: obtaining a plurality of visual feature information sets through the visual feature information corresponding to each image pixel in the airport target area image, each visual feature information set including visual feature information corresponding to a plurality of image pixels, and at the same time, two consecutive visual feature information sets contain visual feature information corresponding to at least one image pixel with an intersection.

[0126] In step S170, the computer system groups the visual feature information corresponding to each image pixel in the airport target area image to form a plurality of visual feature information sets.

[0127] Specifically, each visual feature information set includes feature information corresponding to a number of image pixels with similar visual features. These visual features may include but are not limited to texture, color, brightness, etc., which are extracted in step S160. By grouping pixels with similar features, the computer system can more effectively perform local analysis on the image and identify different image regions and objects.

[0128] In addition, step S170 also specifically points out that two consecutive visual feature information sets should contain at least one visual feature information corresponding to an image pixel with an intersection. This design is to ensure that every part of the image can be fully covered and analyzed to avoid missing any important information. At the same time, this overlapping set design also helps to improve the consistency and accuracy of image analysis.

[0129] To illustrate, assume that the computer system is processing an image of an airport runway with a flock of birds flying on it. In step S170, the computer system groups the pixels into different sets of visual feature information based on the visual features of each pixel (such as color, texture, etc.). For example, pixels on the surface of the runway may be grouped into one set because of similar texture and color features, while flying birds are grouped into another set because of visual features different from the runway (such as feather color and texture). At the same time, since the edge pixels of the birds may contain features of both the runway and the birds, the feature information of these pixels will appear in multiple sets, forming an intersection between the sets. Through this grouping method, the computer system can more accurately identify different objects and areas in the image, providing strong support for subsequent bird identification and expulsion.

[0130] Step S180: Determine the node distribution result of each of the several visual feature information sets in the visual feature information visual knowledge graph, and determine the existence frequency of each of the several visual feature information sets in the airport target area image, wherein the visual feature information visual knowledge graph includes several candidate visual feature information sets.

[0131] In step S180, the computer system mainly completes two tasks: determining the node distribution results of each visual feature information set in the visual feature information visual knowledge graph, and calculating the frequency of each set in the airport target area image. First, the computer system uses a pre-built visual feature information visual knowledge graph, which contains several candidate visual feature information sets. This knowledge graph can be regarded as a large database or network, in which each node represents a specific visual feature information set, and the connection between nodes reflects the relationship and similarity between these sets.

[0132] When the computer system processes to step S180, it will match each visual feature information set obtained in step S170 with the nodes in the knowledge graph. This process is similar to searching for the node in the graph that is most similar to the current set, thereby determining the distribution or location of the set in the graph. This matching can be achieved by calculating the similarity or distance between sets, such as using methods such as cosine similarity and Euclidean distance.

[0133] For example, if a set of visual feature information contains features related to bird feathers (such as specific color distribution and texture patterns), then the set can be matched with a node representing "bird feather features" in the knowledge graph. Secondly, the computer system calculates the frequency of existence (or counts the number of occurrences) of each visual feature information set in the airport target area image. This can be achieved by counting the number or proportion of pixels in the image that belong to the set. Sets with high frequencies of existence often correspond to significant features or important areas in the image. Continuing with the above example, if there are multiple birds in the image and their feathers occupy a larger part of the image, then the visual feature information set related to bird feathers will have a higher frequency of existence.

[0134] Step S180 provides important data support for subsequent bird identification and expulsion strategy formulation by determining the node distribution of each visual feature information set in the visual knowledge graph and the frequency of existence in the image. This information helps the computer system to more accurately identify the birds in the image and formulate corresponding expulsion measures based on their characteristics.

[0135] Step S190: Obtain the first image visual representation carrier through the node distribution results of each visual feature information set in the visual feature information visual knowledge graph and the existence frequency of each visual feature information set in the airport target area image.

[0136] Step S190 performs comprehensive processing on the visual feature information sets extracted and analyzed in the previous steps to generate a first image visual representation carrier. In step S190, the computer system constructs a first image visual representation carrier based on the node distribution results of each visual feature information set in the visual feature information visual knowledge graph, and the frequency of existence of these sets in the airport target area image. This representation carrier is actually a description that integrates various visual features in the image, which can more comprehensively reflect the content and information of the image. Specifically, the computer system first analyzes the node distribution of each visual feature information set in the knowledge graph. These nodes represent different visual features, and their distribution reflects the relative importance and relevance of various features in the image. For example, if a node is associated with multiple visual feature information sets, then this feature may be more prominent in the image.

[0137] Next, the computer system considers the frequency of each set of visual feature information in the image. Feature sets with high frequencies tend to correspond to the main content or salient areas in the image. For example, in the scenario of airport bird recognition, if there are multiple birds in the image and their feather features appear frequently in the image, then these features will occupy a more important position in the representation carrier.

[0138] By integrating the above two aspects of information, the computer system can generate a comprehensive and accurate first image visual representation carrier, which not only includes various visual features in the image, but also reflects the relative relationship and importance between these features.

[0139] For example, suppose there is a group of birds flying in the airport target area image, and their feathers show specific colors and textures. In step S190, the computer system analyzes the node distribution of these feather features in the knowledge graph and counts their frequency of occurrence in the image. Then, the computer system generates a first image visual representation carrier based on this information, which contains key visual features such as the color and texture of the bird's feathers, as well as the importance and relevance of these features in the image. This representation carrier will provide strong support for the subsequent bird identification and expulsion strategy formulation.

[0140] As an implementation manner, before obtaining the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image in step S300, the method further includes:

[0141] Step S290: filtering the airport bird control images in the airport bird control image library based on the preliminary similarity between the airport target area image and the airport bird control images in the airport bird control image library to obtain a candidate airport bird control image library, wherein the candidate airport bird control image library includes a plurality of airport bird control images determined from the airport bird control image library;

[0142] Then, the obtaining of the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image includes: obtaining the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image in the candidate airport bird control image library.

[0143] The core purpose of step S290 is to preliminarily screen the airport bird reference image library to reduce the amount of calculation for subsequent recognition and improve recognition efficiency. In this step, the computer system first calculates the preliminary similarity between the airport target area image and each image in the airport bird reference image library. This similarity calculation can be based on the overall characteristics of the image, such as color distribution, texture characteristics, etc. By setting a similarity threshold, the computer system can filter out those bird reference images that are obviously not matched with the target area image, thereby obtaining a candidate airport bird reference image library.

[0144] For example, assuming that a red bird is captured in the airport target area image, then in step S290, the computer system can filter out those reference images that are obviously not red birds, such as blue parrots or black petrels, etc. In this way, the candidate airport bird reference image library only includes bird images that are relatively close to the target area image in terms of color and other features.

[0145] The purpose of step S30 is to further search for the bird image that best matches the target area image in the candidate airport bird reference image library. To achieve this, the computer system needs to obtain the vector space similarity between the first interactive representation vector (representing the features of the target area image) and the second interactive representation vector (representing the features of the reference image) corresponding to each image in the candidate airport bird reference image library.

[0146] This similarity calculation can be based on more detailed feature matching, such as the shape, posture, feather details of the bird, etc. By comparing the similarity of these features between the first interactive representation carrier and the second interactive representation carrier, the computer system can determine which control image is the most matched to the target area image.

[0147] Continuing with the above example, in step S300, the computer system compares the similarity between the target area image of red birds and each red bird image in the candidate airport bird reference image library. Through fine feature matching, the computer system can eventually determine the bird image that is most similar to the target area image, thereby completing the bird recognition process.

[0148] In general, step S290 and step S300 together constitute an efficient bird identification process. Through the combination of preliminary screening and fine matching, this process can effectively reduce the complexity of calculation and improve the real-time performance of identification while ensuring the accuracy of identification.

[0149] As an implementation mode, the airport target area image is characterized by information mining based on a first characterization information mining algorithm, and the airport bird control image is characterized by information mining based on a second characterization information mining algorithm, and the first characterization information mining algorithm and the second characterization information mining algorithm are two different characterization information mining algorithms; wherein, the first characterization information mining algorithm and the second characterization information mining algorithm have the same algorithm composition, but correspond to different algorithm parameters.

[0150] In the application scenario of intelligent identification and expulsion of birds at airports, the computer system needs to accurately identify the birds in the target area of ​​the airport so that timely expulsion measures can be taken to ensure flight safety. To achieve this goal, the computer system performs representation information mining, i.e., feature extraction, on the airport target area image and the airport bird control image. In this process, the computer system uses two different representation information mining algorithms: the first representation information mining algorithm and the second representation information mining algorithm. Although the algorithm composition, i.e., the algorithm architecture, of these two algorithms is the same, for example, they both use the same neural network layer, but they correspond to different algorithm parameters, such as weights, biases, etc.

[0151] Specifically, the first representation information mining algorithm is used to process the airport target area image. Through specific parameter settings, the algorithm can extract key features in the target area image, such as the shape, color, texture, etc. of the bird. These features will help in the subsequent matching and identification with the airport bird control image.

[0152] The second representation information mining algorithm is used to process images in the airport bird reference image library. Unlike the first algorithm, the second algorithm uses another set of parameter settings to extract features from the reference images. The purpose of this is to make the features of the reference images more prominent and distinguishable, thereby improving the matching accuracy with the airport target area images.

[0153] For example, suppose that both the first representation information mining algorithm and the second representation information mining algorithm adopt the architecture of convolutional neural network (CNN). In the first algorithm, the computer system may adjust the weights and bias parameters of the convolution layer so that the network pays more attention to extracting the overall shape and color features of birds in the airport target area image. In the second algorithm, the computer system may set different parameters so that the network pays more attention to extracting the detailed features in the airport bird control image, such as the feather texture and eye shape of the birds.

[0154] By adopting different parameter settings of the representation information mining algorithm, the computer system can more comprehensively extract the characteristic information in the airport target area image and the airport bird control image, thereby improving the accuracy and efficiency of bird identification. This dual algorithm strategy has important application value in the airport bird intelligent identification and expulsion program.

[0155] Then, the present disclosure also provides a training method for a first characterization information mining algorithm and a second characterization information mining algorithm, which specifically includes the following steps:

[0156] Step S10: obtaining a plurality of training image template groups, each of which includes an airport target area image template and an airport bird control image template.

[0157] In step S10, the task of the computer system is to obtain a number of training image template groups, which will be used as training data for subsequent optimization of the representation information mining algorithm.

[0158] Specifically, each training image template group contains two main parts: one is the airport target area image template, and the other is the airport bird control image template. These two types of images will be processed as sample pairs during the training process.

[0159] The airport target area image template is an actual scene image obtained from the airport's surveillance camera or other image acquisition equipment, which may contain birds. The characteristics of these images are that they reflect the real situation of the airport environment, including various complex factors such as light, angle, and background.

[0160] Airport bird reference image templates are clear, standardized bird images. These images are usually taken in a controlled environment or obtained from a professional image library to ensure image quality and accuracy. The purpose of the reference image is to provide a clear reference to help the algorithm learn and identify birds in images of the airport target area.

[0161] For example, if the computer system is being trained to identify a particular type of bird, it can collect a series of airport target area images containing this type of bird, as well as corresponding high-definition, standardized airport bird control images. These images will be combined into multiple training image template groups for subsequent training processes.

[0162] In this way, the computer system was able to use the large amount of training data to optimize its representation information mining algorithm, thereby improving its ability to accurately identify birds in complex airport environments. This is an important step in ensuring airport safety and effectively repelling birds.

[0163] Step S20: Based on the first representation information mining algorithm, representation information mining is performed on the airport target area image template to obtain a first image pixel representation carrier, a first image semantic representation carrier and a first image visual representation carrier.

[0164] In the application scenario of intelligent identification and expulsion of birds at airports, step S20 is the key link for the computer system to process images and extract features. In this step, the computer system uses the first representation information mining algorithm to conduct an in-depth analysis of the airport target area image template to obtain multi-level information in the image.

[0165] Specifically, the first representation information mining algorithm may be a complex deep learning model, such as a convolutional neural network (CNN), which can extract rich feature information from images. When the algorithm is applied to the airport target area image template, it mines the representation information of the image from different angles.

[0166] First, through the processing of the algorithm, the computer system can obtain the first image pixel representation carrier. This carrier contains the pixel-level features of the image, such as color, brightness, contrast, etc. Taking the airport target area image as an example, the pixel representation carrier may capture the subtle texture of the bird's feathers, color distribution, and contrast with the background. Secondly, the algorithm will also extract the first image semantic representation carrier. This carrier reflects the content and contextual information in the image, such as objects (such as birds), scenes (such as airport runways) in the image, and the relationship between them. In the scene of airport bird recognition, the semantic representation carrier may encode important information such as the bird's species, posture, behavior, and interaction with the surrounding environment.

[0167] Finally, through the processing of the algorithm, the first image visual representation carrier can also be obtained. This carrier captures the visual style and structural information of the image, which helps the computer system understand the overall layout and composition of the image. When identifying birds at the airport, the visual representation carrier may highlight the visual features of the bird, such as its shape, outline, and relative position to other objects.

[0168] For example, suppose the airport target area image template contains a pigeon moving on the runway. Through the processing of the first representation information mining algorithm, the computer system can obtain a pixel representation carrier containing information such as the color and brightness of the pigeon's feathers, a semantic representation carrier reflecting the pigeon's type and posture, and a visual representation carrier highlighting the pigeon's shape and relative position to the runway. These rich representation information will provide strong support for subsequent image matching and bird recognition.

[0169] Step S30: The first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier are interacted to obtain a first interactive representation carrier; wherein the first image pixel representation carrier is used to carry the pixel feature representation of the airport target area image template, the first image semantic representation carrier is used to carry the content feature representation of the airport target area image template, and the first image visual representation carrier is used to carry the visual feature information of the airport target area image template.

[0170] Step S30 is a feature fusion step in the computer system, which can refer to the aforementioned step S200. In this step, the computer system will interact with the first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier extracted in the previous step, that is, to fuse the features at different levels to form a more comprehensive and integrated feature representation, namely, the first interactive representation carrier.

[0171] Specifically, the first image pixel representation carrier carries the pixel feature representation of the airport target area image template, such as the feather texture, color and other details of the birds in the image. These pixel features are the basis of image recognition and can provide rich visual clues.

[0172] The first image semantic representation carrier carries the content feature representation of the image, which reflects the entities, scenes and their relationships in the image. In the context of airport bird recognition, this includes information such as the bird's species, behavior, and relationship with other objects. These semantic features help the computer system understand the connotation and context of the image.

[0173] The first image visual representation carrier carries the visual feature information of the image, such as shape, outline and structure. These features help capture the overall visual style and layout of the image, which is particularly important for identifying different types of birds.

[0174] In step S30, the computer system effectively fuses these features of different dimensions through a specific algorithm or model, such as a feature fusion layer or a multimodal fusion method in deep learning. The purpose of this is to integrate the advantages of various features and improve the accuracy and robustness of subsequent bird identification.

[0175] For example, suppose the system is processing an image of a bird on an airport runway. Through pixel representation, the computer system can capture the delicate texture of the bird's feathers; through semantic representation, the computer system can understand that this is a bird walking on the runway; and through visual representation, the computer system can grasp the overall shape and posture of the bird. In step S30, these different levels of features will be fused together to form a more complete, multi-dimensional feature representation, namely the first interactive representation carrier. This carrier will provide strong support for subsequent classification, recognition and other tasks.

[0176] Step S40: Based on the second representation information mining algorithm, representation information mining is performed on the airport bird control image template to obtain a second image pixel representation carrier, a second image semantic representation carrier and a second image visual representation carrier.

[0177] It can be understood that the process of step S40 can refer to step S20. Similar to step S20, this step also involves mining of representation information, but the target object is the airport bird reference image template.

[0178] In this step, the computer system uses a second representation information mining algorithm, which may be similar to or different from the first representation information mining algorithm in step S20, depending on the design and requirements of the system. Typically, these algorithms are based on advanced deep learning techniques, such as convolutional neural networks (CNNs) or other image processing techniques.

[0179] By processing the airport bird control image template, the computer system obtains three different types of representation carriers: second image pixel representation carrier, second image semantic representation carrier and second image visual representation carrier. The second image pixel representation carrier captures the pixel-level features of the bird image, such as the color, texture and distribution of feathers. These features are essential for distinguishing different species of birds. The second image semantic representation carrier reveals semantic information such as the specific species and behavior patterns of the birds in the image. This is essential for the computer system to understand the birds in the image and respond accordingly.

[0180] Finally, the second image visual representation carrier provides comprehensive information about the bird's shape, posture, and visual structure. This helps the computer system identify and track the bird more accurately.

[0181] For example, if the airport bird reference image template is a high-definition photo of a pigeon, then in step S40, the computer system will extract various features of the photo through the second representation information mining algorithm. The second image pixel representation carrier may include the gray and white texture and specific color distribution of the pigeon's feathers; the second image semantic representation carrier may reveal that it is a pigeon and its behavior pattern of resting or foraging; and the second image visual representation carrier may highlight the overall outline of the pigeon, the shape of its wings, and the head features and other visual information.

[0182] Step S50: The second image pixel representation carrier, the second image semantic representation carrier and the second image visual representation carrier are interacted to obtain a second interactive representation carrier; wherein the second image pixel representation carrier is used to carry the pixel feature representation of the airport bird control image template, the second image semantic representation carrier is used to carry the content feature representation of the airport bird control image template, and the second image visual representation carrier is used to carry the visual feature information of the airport bird control image template.

[0183] It can be understood that the process of step S50 can refer to step S30. In step S50, the computer system interacts the second image pixel representation carrier, the second image semantic representation carrier and the second image visual representation carrier previously extracted from the airport bird control image template, that is, fuses these features to form a comprehensive feature representation, namely, the second interactive representation carrier.

[0184] Specifically, the second image pixel representation carrier carries the pixel feature representation of the airport bird reference image template. These features may include detailed information such as the bird's feather color and texture, which are basic elements in image recognition. For example, the specific feather color or texture of a certain bird can serve as its unique identification mark.

[0185] The second image semantic representation carrier reflects semantic information such as the bird species and behavior in the image. This information helps the computer system understand what the bird in the image is and what it is doing. In the context of bird expulsion at airports, identifying the species and behavior of birds is crucial to determining whether expulsion measures need to be taken and what methods should be used.

[0186] The second image visual representation carrier captures the overall visual features and layout of the bird, such as the bird's shape, posture, etc. These features help the computer system grasp the overall appearance of the bird, thereby improving the accuracy of recognition.

[0187] In step S50, the computer system fuses these features through a specific algorithm to generate a second interactive representation carrier. This process is similar to combining multiple clues or evidence together to form a more comprehensive and accurate judgment. For example, if the computer system relies only on pixel features, it may not be able to accurately distinguish between two birds with similar appearances, but by combining semantic features and visual features, the computer system can make a more accurate judgment. In general, the fusion process of step S50 is to improve the accuracy and robustness of bird identification. By integrating different types of features, the computer system is able to more comprehensively understand the birds in the image, thereby providing a reliable basis for subsequent expulsion strategies. In the scenario of bird expulsion at the airport, this accurate identification is crucial because it is directly related to the effectiveness and safety of the expulsion measures.

[0188] Step S60: performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain an algorithm training cost.

[0189] Step S60 performs similarity learning on the features extracted in the previous step so that the system can accurately identify the birds in the airport target area. In this step, the computer system performs similarity learning on the first interactive representation carrier (the feature fusion result extracted from the airport target area image template) and the second interactive representation carrier (the feature fusion result extracted from the airport bird control image template), also known as comparative learning. The purpose of this learning is to enable the system to distinguish different image features and identify similar features, so as to accurately determine the bird species in the target area.

[0190] Specifically, similarity learning is achieved through a machine learning algorithm, which can be a relatively simple distance measurement method, such as Euclidean distance, cosine similarity, etc., or a more complex machine learning model, such as support vector machine (SVM), deep learning network, etc. In this process, the computer system calculates the similarity between the first interaction representation carrier and the second interaction representation carrier, and then adjusts the model parameters according to the preset label (i.e., whether they are similar or the same bird).

[0191] For example, if the first interactive representation vector represents the characteristics of a flying pigeon on an airport runway, and the second interactive representation vector represents the characteristics of a static pigeon in the bird database, then through similarity learning, the computer system should be able to determine that the two vectors are similar because they both represent pigeons. To do this, the computer system needs to constantly adjust its internal parameters to minimize the prediction error, which is the algorithm training cost (also called loss or error).

[0192] The algorithm training cost is an important indicator to measure the accuracy of model prediction. In the process of similarity learning, the computer system constantly tries different parameter combinations to find the set of parameters that minimizes the training cost. This cost is usually obtained by calculating the difference between the model prediction result and the actual label. In the scenario of airport bird identification, a lower training cost means that the system can more accurately identify the birds in the target area. In general, the similarity learning in step S60 is a key learning process performed by the computer system in order to accurately identify the birds in the target area of ​​the airport. Through this process, the computer system can learn how to distinguish different bird characteristics and accurately identify the bird species in the target area, thereby providing accurate decision-making basis for subsequent expulsion measures.

[0193] Step S70: Correct the parameters of the first characterization information mining algorithm and the second characterization information mining algorithm through the algorithm training cost, and use the trained first characterization information mining algorithm to perform characterization information mining and interaction on the airport target area image to determine the airport bird control image corresponding to the airport target area image.

[0194] In step S70, the computer system uses the algorithm training cost calculated in the previous step to modify the parameters of the first representation information mining algorithm and the second representation information mining algorithm. These representation information mining algorithms may be deep learning networks, convolutional neural networks (CNNs) or other image processing algorithms, which are used to extract useful feature information from images.

[0195] The algorithm training cost reflects the accuracy of the model's predictions. If the cost is high, it means that there is a large difference between the model predictions and the actual labels, and the model needs to be optimized by adjusting the algorithm parameters. In the scenario of airport bird identification, this means that the system needs to more accurately mine the feature information in the image in order to more accurately identify the birds. By correcting the parameters of the representation information mining algorithm, the computer system is able to improve its ability to extract features from images. For example, if the system finds that the previously extracted pixel features are not effective enough to distinguish different birds, it can adjust the relevant parameters to extract more discriminative features. Similarly, if the system finds that the way semantic features or visual features are extracted needs to be improved, it will also adjust the parameters accordingly.

[0196] The trained first representation information mining algorithm will be used to perform representation information mining and interaction on the airport target area image. This means that when the system receives a new airport target area image, it will use this optimized algorithm to extract feature information from the image and match it with known airport bird reference images. For example, assume that after training and optimization in step S70, the system can accurately extract key features such as bird feather color, texture, and shape from the image. When the airport surveillance camera captures a new target area image, the computer system uses the trained first representation information mining algorithm to extract these features and compares them with the airport bird reference images in the database. In this way, the computer system can quickly and accurately determine the bird species in the target area.

[0197] In general, step S70 is an important step in optimizing the representation information mining algorithm in the computer system. By using the algorithm training cost to adjust the algorithm parameters, the computer system can improve its ability and accuracy in extracting features from images, thereby better identifying birds in the airport target area. This is crucial for subsequent repellent measures because it ensures that the system can detect potential bird threats in the first place and take corresponding actions.

[0198] As an embodiment, the several training image template groups include: a training image true template group, a simple training image false template group and a complex training image false template group; wherein the airport bird control image template included in the training image true template group is a labeled airport bird control image of the airport target area image template; the airport bird control image template included in the simple training image false template group is arbitrarily sampled from an airport bird control image library, and is not a labeled airport bird control image of the airport target area image template; the representation carrier spatial similarity between the airport bird control image template included in the complex training image false template group and the airport target area image template is not less than a preset spatial similarity coefficient, and is not a labeled airport bird control image of the airport target area image template, wherein the representation carrier spatial similarity is the carrier spatial similarity between the first interactive representation carrier of the airport target area image template and the second interactive representation carrier of the airport bird control image template.

[0199] In the application scenario of intelligent identification and expulsion of birds at airports, the computer system uses a specific training image template group for model training to improve the accuracy of identifying birds at airports. This training image template group includes three types: training image true template group (Positive Samples), simple training image false template group (Easy NegativeSamples) and complex training image false template group (Hard Negative Samples).

[0200] The airport bird control image templates in the training image true template set are the annotated airport bird control images of the airport target area image templates. This means that these images are exact matches to the birds in the airport target area image, serving as positive examples during training. For example, if there is a pigeon in the airport target area image, then the training image true template set will contain an annotated image corresponding to the pigeon.

[0201] The airport bird control image templates in the simple training image fake template group are randomly sampled from the airport bird control image library, but they are not annotated airport bird control images of the airport target area image template. These images are used as simple negative samples in the training process to help the model distinguish other birds that are not similar to the target bird. For example, if the target area image is a pigeon, the simple training image fake template group may contain images of other types of birds such as sparrows and magpies.

[0202] The airport bird control image templates in the complex training image false template group have a certain representation carrier spatial similarity with the airport target area image template, but this similarity is not less than a preset spatial similarity coefficient, and these images are not labeled airport bird control images of the airport target area image template. These images are used as complex negative samples in the training process to help the model more finely distinguish other birds that are similar in appearance to the target bird. For example, if the target area image is a bald eagle, the complex training image false template group may contain images of other raptors, such as golden eagles or red-tailed hawks, because these birds have a certain similarity in appearance to the bald eagle.

[0203] Through this carefully designed set of training image templates, the computer system is able to train a more accurate and robust bird recognition model. This model can not only accurately identify birds that fully match the target area image, but also effectively distinguish other birds that are not similar to the target bird, and other birds that are similar in appearance to the target bird. This is crucial for the airport bird intelligent identification and expulsion system because it ensures that the system can accurately identify and expel potential bird threats in actual applications.

[0204] As an embodiment, the method further includes:

[0205] Step S1: obtaining the representation carrier space similarity between the airport target area image template and each airport bird reference image in the airport bird reference image library;

[0206] Step S2: The M airport bird control images with the highest spatial similarity of the representation carrier other than the annotated airport bird control images of the airport target area image template are used as airport bird control image templates to form M false template groups of the complex training image together with the airport target area image template, where M is a non-zero natural number.

[0207] Step S1 obtains the spatial similarity of the representation carrier between the airport target area image template and each airport bird control image in the airport bird control image library. This step helps the system understand which images in the bird control image library the birds in the target area image are most similar to. Specifically, the computer system first extracts the features of the airport target area image template, which include but are not limited to the shape, color, texture, etc. of the bird. These features are converted into a feature vector for subsequent similarity calculations. For example, a feature vector may contain information on multiple dimensions such as the shape of the bird's wings, the distribution of feather color, and the shape of the beak.

[0208] Next, the computer system performs the same feature extraction process on each image in the airport bird reference image library to generate a corresponding feature vector. The computer system then uses a similarity metric, such as cosine similarity or Euclidean distance, to calculate the similarity between the feature vector of the airport target area image template and the feature vector of each image in the bird reference image library.

[0209] For example, if the airport target area image template contains a red-headed gull, the computer system will search for the image that is most similar to the red-headed gull in the airport bird control image library. By calculating the similarity between feature vectors, the computer system can find a similarity-ranked list of other bird images that are most similar to the red-headed gull. The output of this step is a similarity metric that represents the degree of similarity between the airport target area image template and each image in the airport bird control image library. These similarity metrics will be used in subsequent steps to construct a complex training image false template group to improve the accuracy of airport bird identification. In this way, the computer system can more accurately identify birds in the airport target area and take appropriate repellent measures to ensure flight safety.

[0210] In the application scenario of intelligent identification and expulsion of birds at airports, step S2 is the key step in constructing a set of false templates for complex training images. In this step, the computer system selects the M airport bird control images with the highest similarity, in addition to the labeled airport bird control images of the airport target area image template, based on the spatial similarity of the representation carrier calculated in step S1. These selected images will be used as airport bird control image templates, and together with the airport target area image template, form a set of M false templates for complex training images.

[0211] Specifically, if the annotated bird in the airport target area image template is an egret, then in step S1, the computer system has calculated the similarity between the egret and all images in the airport bird reference image library. In step S2, the computer system excludes the annotated images that are exactly the same as the egret in the target area image template, and then selects the M images that are most similar to the egret from the remaining images.

[0212] These selected images may be other birds that are very similar to the target egret in color, shape or texture, such as white cranes, white storks, etc. These images serve as false samples in the false template group of complex training images, and together with the target area image template, they constitute a set of challenging training data.

[0213] For example, if M is set to 3, the computer system selects three other bird images that are most similar to the target egret. These three images, together with the target egret image, form three complex training image false template groups. These template groups will be used to train the machine learning model so that it can still accurately identify the target bird when faced with other birds that are highly similar to the target bird.

[0214] In this way, the computer system is able to use these complex training image pseudo template groups to enhance the generalization ability and recognition accuracy of the model. In practical applications, this helps to improve the accuracy of bird identification at airports, thereby more effectively driving away birds and ensuring the safe operation of airports.

[0215] As an implementation mode, the step S60, performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain the algorithm training cost, includes: performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain a first cost variable; performing similarity learning on the first image pixel representation carrier and the second image pixel representation carrier to obtain a second cost variable; performing similarity learning on the first image semantic representation carrier and the second image semantic representation carrier to obtain a third cost variable; performing similarity learning on the first image visual representation carrier and the second image visual representation carrier to obtain a fourth cost variable; obtaining the algorithm training cost through the first cost variable, the second cost variable, the third cost variable and the fourth cost variable.

[0216] Specifically, the computer system performs similarity learning on the first interaction representation carrier and the second interaction representation carrier. The interaction representation carrier may include the behavior, posture, and interaction information between the bird in the image and the environment. By comparing these interaction features, the computer system can learn how to identify the characteristics of different birds in different behavior modes such as flying, foraging, and perching. The first cost variable obtained in this step reflects the similarity or difference between the two interaction representation carriers.

[0217] The computer system performs similarity learning on the first image pixel representation vector and the second image pixel representation vector. The pixel representation vector focuses on the detailed information of the image, such as color, texture, etc. In the scenario of airport bird recognition, this helps the system distinguish birds with similar shapes but different colors or textures. Through this step, the computer system can obtain a second cost variable, which reflects the degree of similarity between the two pixel representation vectors.

[0218] The computer system learns the similarity between the first image semantic representation carrier and the second image semantic representation carrier. The semantic representation carrier may contain high-level information such as the type and habits of the bird. In the airport environment, this helps the system identify birds that are similar in appearance but have different habits. The third cost variable obtained in this step reflects the similarity between the two semantic representation carriers.

[0219] The computer system learns the similarity between the visual representation carrier of the first image and the visual representation carrier of the second image. The visual representation carrier may focus on the overall visual features of the image, such as shape, outline, etc. In airport bird recognition, this helps the system quickly capture the basic shape features of the bird and perform rough classification. The fourth cost variable obtained in this step reflects the similarity between the two visual representation carriers.

[0220] The computer system combines the cost variables obtained in the previous four steps (first cost variable, second cost variable, third cost variable and fourth cost variable) to calculate the final algorithm training cost. This cost function is the key to machine learning model optimization. It guides the model to continuously adjust parameters during training to better adapt to and identify various birds in the airport environment.

[0221] For example, suppose the system is learning how to distinguish between egrets and white cranes in an airport environment. The two birds are similar in appearance, but different in color, texture, and habits. Through the similarity learning of steps S61 to S64, the computer system can capture the differences between egrets and white cranes in interaction, pixels, semantics, and visual features, respectively. Then, in step S65, the computer system calculates the training cost based on these differences, thereby guiding the model to more accurately identify the two birds. In practical applications, this helps to improve the accuracy of bird identification at airports, thereby improving the efficiency of expulsion and ensuring the safe operation of airports.

[0222] The following is an exemplary explanation of the calculation formulas for each cost variable and training cost:

[0223] First, for the first cost variable C1 (similarity cost of interaction representation carrier):

[0224] If cosine similarity is used to measure the similarity between two interactive representation carriers, the formula can be as follows:

[0225] ;

[0226] in, and They are the vector representations of two interaction representation vectors.

[0227] Then, for the second cost variable C2 (the similarity cost of the pixel representation carrier):

[0228] For pixel representation carriers, the mean square error (MSE) can be used to measure the pixel-level difference between two images:

[0229] ;

[0230] Among them, xi and yi are the values ​​of the corresponding pixels in the two pixel representation vectors, and n is the total number of pixels.

[0231] Again, for the third cost variable C3 (similarity cost of semantic representation carrier):

[0232] For semantic representation carriers, if they are in the form of category labels, cross entropy loss can be used to measure similarity:

[0233] ;

[0234] Where C is the number of categories, yc is the one-hot encoding of the true label, and pc is the probability of the category predicted by the model.

[0235] Finally, for the fourth cost variable (similarity cost of visual representation carrier):

[0236] The similarity of visual representation carriers can also be measured by cosine similarity, similar to the calculation method of the first cost variable.

[0237] ;

[0238] in, and are the vectors of two visual representation carriers respectively.

[0239] The final algorithm training cost CT can be the weighted sum of the above four cost variables:

[0240] ;

[0241] Among them, w1, w2, w3, and w4 are the weights of each cost variable, which can be adjusted according to the importance of the task and experience.

[0242] The present disclosure also provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present disclosure.

[0243] Processor 101 may be a CPU, a general-purpose processor, a GPU, a DSP, an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of the present disclosure. Processor 101 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0244] The bus 102 may include a path to transmit information between the above components. The bus 102 may be a PCI bus or an EISA bus, etc. The bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0245] The memory 103 may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an EEPROM, a CD-ROM or other optical disk storage, an optical disk storage (including a compressed optical disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.

[0246] The memory 103 is used to store application code for executing the solution of the present disclosure, and the execution is controlled by the processor 101. The processor 101 is used to execute the application code stored in the memory 103 to implement the content shown in any of the above method embodiments.

[0247] An embodiment of the present disclosure provides a computer system, and the computer system in the embodiment of the present disclosure includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above method is implemented.

[0248] An embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon. When the computer program runs on a processor, the processor can execute corresponding contents in the aforementioned method embodiment.

[0249] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0250] The above description is only a partial implementation mode of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present disclosure. These improvements and modifications should also be regarded as the protection scope of the present disclosure.

Claims

1. An intelligent method for identifying and driving away birds at an airport, characterized in that: The method comprises: Performing representation information mining on the airport target area image to obtain a first image pixel representation carrier, a first image semantic representation carrier, and a first image visual representation carrier, wherein the first image pixel representation carrier is used to carry the pixel feature representation of the airport target area image, the first image semantic representation carrier is used to carry the content feature representation of the airport target area image, and the first image visual representation carrier is used to carry the visual feature information of the airport target area image; Interacting the first image pixel representation carrier, the first image semantic representation carrier, and the first image visual representation carrier to obtain a first interactive representation carrier; Obtaining the vector space similarity between the first interactive representation vector and the second interactive representation vector corresponding to each airport bird control image; wherein the second interactive representation vector is obtained by mining and interacting the representation information of the airport bird control image; Determining the airport bird control image corresponding to the airport target area image in each of the airport bird control images through the carrier space similarity; Generate a target driving strategy according to the driving strategy corresponding to the determined airport bird control image; The method of mining the representation information of the airport target area image to obtain the first image pixel representation carrier includes: A plurality of image pixel clusters are obtained through the airport target area image, each image pixel cluster includes a plurality of image pixels, and at least one image pixel having an intersection is included between two consecutive image pixel clusters; Determining a node distribution result of each of the plurality of image pixel clusters in a visual knowledge graph, and determining a frequency of existence of each of the plurality of image pixel clusters in the airport target area image, wherein the visual knowledge graph includes a plurality of candidate image pixel clusters; Obtaining the first image pixel representation carrier through the node distribution result of each image pixel group in the visual knowledge graph and the existence frequency of each image pixel group in the airport target area image; The method of mining the representation information of the airport target area image to obtain the first image semantic representation carrier includes: Acquire an image pixel matrix corresponding to the airport target area image, wherein the image pixel matrix includes a plurality of image pixels obtained by extracting pixel values ​​of the airport target area image and arranging them according to original positions; Extracting mutual information of adjacent elements of each image pixel in the image pixel matrix to obtain the first image semantic representation carrier; The method of mining the representation information of the airport target area image to obtain the first image visual representation carrier includes: Acquire visual feature information corresponding to each image pixel in the airport target area image; A plurality of visual feature information sets are obtained through the visual feature information corresponding to each image pixel in the airport target area image, each visual feature information set includes visual feature information corresponding to a plurality of image pixels, and at the same time, two consecutive visual feature information sets contain at least one visual feature information corresponding to an image pixel having an intersection; Determine a node distribution result of each of the several visual feature information sets in a visual feature information visual knowledge graph, and determine an existence frequency of each of the several visual feature information sets in the airport target area image, wherein the visual feature information visual knowledge graph includes several candidate visual feature information sets; The first image visual representation carrier is obtained through the node distribution results of each visual feature information set in the visual feature information visual knowledge graph and the existence frequency of each visual feature information set in the airport target area image.

2. The method according to claim 1, characterized in that Before obtaining the carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image, the method further includes: filtering the airport bird control images in the airport bird control image library based on the preliminary similarity between the airport target area image and the airport bird control images in the airport bird control image library to obtain a candidate airport bird control image library, wherein the candidate airport bird control image library includes a plurality of airport bird control images determined from the airport bird control image library; The obtaining of the vector space similarity between the first interactive representation vector and the second interactive representation vector corresponding to each airport bird control image includes: The carrier space similarity between the first interactive representation carrier and the second interactive representation carrier corresponding to each airport bird control image in the candidate airport bird control image library is obtained.

3. The method according to claim 1 or 2, characterized in that The airport target area image is subjected to representation information mining based on a first representation information mining algorithm, and the airport bird control image is subjected to representation information mining based on a second representation information mining algorithm, wherein the first representation information mining algorithm and the second representation information mining algorithm are two different representation information mining algorithms; wherein the first representation information mining algorithm and the second representation information mining algorithm have the same algorithm composition but correspond to different algorithm parameters; The first characterization information mining algorithm and the second characterization information mining algorithm are trained based on the following operations: Acquire a plurality of training image template groups, each of which includes an airport target area image template and an airport bird control image template; Based on the first representation information mining algorithm, the representation information mining is performed on the airport target area image template to obtain a first image pixel representation carrier, a first image semantic representation carrier and a first image visual representation carrier; The first image pixel representation carrier, the first image semantic representation carrier and the first image visual representation carrier are interacted to obtain a first interactive representation carrier; wherein the first image pixel representation carrier is used to carry the pixel feature representation of the airport target area image template, the first image semantic representation carrier is used to carry the content feature representation of the airport target area image template, and the first image visual representation carrier is used to carry the visual feature information of the airport target area image template; Based on the second representation information mining algorithm, the representation information mining is performed on the airport bird control image template to obtain a second image pixel representation carrier, a second image semantic representation carrier and a second image visual representation carrier; The second image pixel representation carrier, the second image semantic representation carrier and the second image visual representation carrier are interacted to obtain a second interactive representation carrier; wherein the second image pixel representation carrier is used to carry the pixel feature representation of the airport bird comparison image template, the second image semantic representation carrier is used to carry the content feature representation of the airport bird comparison image template, and the second image visual representation carrier is used to carry the visual feature information of the airport bird comparison image template; Performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain an algorithm training cost; The parameters of the first characterization information mining algorithm and the second characterization information mining algorithm are corrected through the algorithm training cost, and the trained first characterization information mining algorithm is used to perform characterization information mining and interaction on the airport target area image to determine the airport bird control image corresponding to the airport target area image.

4. The method according to claim 3, characterized in that The plurality of training image template groups include: a training image true template group, a simple training image false template group and a complex training image false template group; wherein the airport bird control image template included in the training image true template group is a labeled airport bird control image of the airport target area image template; The airport bird control image template included in the simple training image false template group is randomly sampled from an airport bird control image library and is not a labeled airport bird control image of the airport target area image template; The representation carrier spatial similarity between the airport bird control image template and the airport target area image template included in the complex training image false template group is not less than a preset spatial similarity coefficient, and is not a labeled airport bird control image of the airport target area image template, wherein the representation carrier spatial similarity is the carrier spatial similarity between the first interactive representation carrier of the airport target area image template and the second interactive representation carrier of the airport bird control image template.

5. The method according to claim 4, characterized in that The method further comprises: Acquire the representation carrier space similarity between the airport target area image template and each airport bird control image in the airport bird control image library; The M airport bird control images other than the annotated airport bird control images of the airport target area image template, which have the highest spatial similarity in the representation carrier, are used as airport bird control image templates to form M false template groups of the complex training image together with the airport target area image template, where M is a non-zero natural number.

6. The method according to claim 3, characterized in that The performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain an algorithm training cost includes: Performing similarity learning on the first interaction representation carrier and the second interaction representation carrier to obtain a first cost variable; performing similarity learning on the first image pixel representation carrier and the second image pixel representation carrier to obtain a second cost variable; performing similarity learning on the first image semantic representation carrier and the second image semantic representation carrier to obtain a third cost variable; performing similarity learning on the first image visual representation carrier and the second image visual representation carrier to obtain a fourth cost variable; The algorithm training cost is obtained through the first cost variable, the second cost variable, the third cost variable and the fourth cost variable.

7. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 6 is implemented.

Images