A method and system for detecting flying animals

Abstract

The application is suitable for the technical field of flying animal detection, and particularly relates to a flying animal detection method and system. The method comprises the following steps: receiving a detection request uploaded by a user, extracting a picture that needs to be detected, tracing back to a source device, finding an edge node, deploying a pre-constructed global target detection model to the edge node, inputting the picture into the global target detection model, and outputting a plurality of flying animal instances and candidate boxes; calculating the brightness value of each pixel point by using a brightness formula, drawing a two-dimensional frequency base graph composed of base functions, and dividing the picture into a plurality of blocks. By determining the intersection and union set, the application can verify the suspected position, avoid missed detection, effectively filter the misdetected area, construct a verification task, and label the flying animal through the man-machine cooperative mode, thereby reducing the professional threshold, increasing the public participation, and greatly improving the labeling efficiency and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flying animal detection, and particularly relates to a flying animal detection method and system. BACKGROUND

[0002] In aspects such as environmental impact assessment, pest and disease monitoring, and wildlife protection, monitoring equipment is often deployed to collect image data of flying animals in a region, so as to monitor the types, quantities, and activity patterns of flying animals in the target region, and the detection task of flying animals is generally completed by a target detection model.

[0003] However, the traditional target detection model is prone to false detection and missed detection in a complex background, and has limited ability to distinguish different types of flying animals, especially flying animals with similar shapes or small sizes.

[0004] Therefore, how to use the target detection model to detect the same picture multiple times is a technical problem to be solved by the present application. SUMMARY

[0005] The present application aims to provide a flying animal detection method and system to solve the problem of how to use the target detection model to detect the same picture multiple times in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A flying animal detection method, the method comprising:

[0008] receiving a detection request uploaded by a user, extracting a picture that needs to be detected, tracing back to a source device, finding an edge node, deploying a pre-constructed global target detection model to the edge node, inputting the picture into the global target detection model, and outputting a plurality of flying animal instances and candidate boxes;

[0009] calculating the brightness value of each pixel point using a brightness formula, drawing a two-dimensional frequency base graph composed of basis functions, dividing the picture into a plurality of blocks, quantizing the blocks using the two-dimensional frequency base graph, obtaining transformation coefficients, traversing a high-frequency signal region using a preset brightness matrix and a chrominance matrix, and performing interpolation magnification;

[0010] creating a local target detection model and deploying it to a cloud server, inputting the high-frequency signal region into the local target detection model, outputting a plurality of flying animal instances and verification boxes, comparing the candidate boxes and the verification boxes, calculating the intersection over union, obtaining the intersection and the union, and inserting the intersection into the flying animal instances;

[0011] The edge node and the cloud server are integrated to build a distributed processing architecture, the verification frame is uploaded to the cloud server, a segment corresponding to the verification frame is intercepted from the picture to generate a verification task, the verification task is issued when a detection request is received, a verification result is obtained, the flying animal instance and the verification result are marked in the picture, and the picture is sent to a preset terminal.

[0012] Further, the step of backtracking the source device, finding the edge node, and deploying the pre-constructed global target detection model to the edge node comprises:

[0013] A scheduling platform of the edge node is constructed, and evaluation indexes of each edge node are collected, wherein the evaluation indexes at least include: delay and available bandwidth;

[0014] The weight value of each evaluation index is configured, the real-time value of the evaluation index is collected, the evaluation score is calculated, and the evaluation score is uploaded to the scheduling platform and dynamically updated.

[0015] Further, the method further comprises:

[0016] Based on the evaluation score, the edge nodes are clustered into ordinary nodes and super nodes, and the global target detection model is deployed to a plurality of super nodes;

[0017] The load of each super node is identified, and a load balancing mechanism is integrated into the scheduling platform.

[0018] Further, the step of dividing the picture into a plurality of blocks, quantizing the blocks by using a two-dimensional frequency base graph, and obtaining the transform coefficient comprises:

[0019] The basic range of each block is set, and the edge pixels in the picture are traversed by using an edge detection algorithm;

[0020] The number of edge pixels in each basic range is counted, and the basic range is adjusted based on the number, wherein the adjustment includes: expansion and reduction.

[0021] Further, the step of calculating the intersection over union, obtaining the intersection and the union, and inserting the intersection into the flying animal instance comprises:

[0022] The intersection over union of each picture is calculated, the first set value and the second set value are selected, the picture with the intersection over union less than the first set value is defined as a negative sample, and the picture with the intersection over union greater than the second set value is defined as a positive sample;

[0023] The positive sample and the negative sample are integrated to generate a training set, and the global target detection model and the local target detection model are trained.

[0024] Further, the step of issuing a verification task and obtaining a verification result when receiving a detection request comprises:

[0025] A plurality of distribution channels of the verification task are set, and a confidence score of each verification frame is calculated based on the verification result;

[0026] The verification result with a confidence score greater than a threshold value is marked in the picture.

[0027] Further, the system comprises:

[0028] The receiving module is configured to receive a detection request uploaded by a user, extract a picture requiring detection, trace a source device, find an edge node, deploy a pre-constructed global object detection model to the edge node, input the picture into the global object detection model, and output a plurality of flying animal instances and candidate frames;

[0029] The magnification module is configured to calculate a brightness value of each pixel point using a brightness formula, draw a two-dimensional frequency base graph composed of a base function, divide the picture into a plurality of blocks, quantize the blocks using the two-dimensional frequency base graph to obtain transformation coefficients, and traverse a high-frequency signal region using a preset brightness matrix and a chroma matrix, and perform interpolation magnification.

[0030] The insertion module is configured to create a local object detection model and deploy the local object detection model to a cloud server, input the high-frequency signal region into the local object detection model, output a plurality of flying animal instances and verification frames, compare the candidate frames and the verification frames, calculate an intersection over union, obtain an intersection and a union, and insert the intersection into the flying animal instances.

[0031] The sending module is configured to integrate the edge node and the cloud server, construct a distributed processing architecture, upload the verification frames to the cloud server, cut out a segment corresponding to the verification frames from the picture, generate a verification task, issue the verification task when receiving the detection request, obtain a verification result, mark the flying animal instances and the verification result in the picture, and send the picture to a preset terminal.

[0032] Further, the receiving module comprises:

[0033] The collection unit is configured to construct a scheduling platform of the edge node, and collect evaluation indexes of each edge node, wherein the evaluation indexes at least include a delay and an available bandwidth.

[0034] The updating unit is configured to configure a weight value of each evaluation index, collect a real-time value of the evaluation index, calculate an evaluation score, upload the evaluation score to the scheduling platform, and dynamically update the evaluation score.

[0035] Further, the magnification module comprises:

[0036] The traversal unit is configured to set a basic range of each patch, traverse edge pixels in the picture by using an edge detection algorithm, and output the edge pixels.

[0037] The statistical unit is configured to count a number of the edge pixels in each basic range, and adjust the basic range based on the number, wherein the adjustment comprises expansion and reduction.

[0038] Further, the insertion module comprises:

[0039] The calculation unit is configured to calculate an intersection over union of each picture, select a first set value and a second set value, define a picture with an intersection over union less than the first set value as a negative sample, and define a picture with an intersection over union greater than the second set value as a positive sample.

[0040] The training unit is configured to integrate the positive sample and the negative sample, generate a training set, and train the global target detection model and the local target detection model.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] By deploying the global target detection model into the edge node, the data transmission pressure can be reduced, the flying animals can be quickly screened out, and the suspected positions can be marked out. By determining the high-frequency signal region, the edge and texture features of the flying animals can be highlighted, the distinction between the flying animals and the background can be enhanced, the detection accuracy of the flying animals can be improved, the identifiable of the small targets can be enhanced by interpolation magnification, the edge detail expression ability is greatly improved, the detection accuracy is further improved, the picture can be detected twice by constructing the local target detection model, the identification efficiency of the small flying animals is improved, the suspected positions can be verified by determining the intersection and the union, the missed detection can be avoided, the false detection region can be effectively filtered, the flying animals can be labeled by the man-machine cooperative mode by constructing the verification task, the public participation is increased, and the labeling efficiency and accuracy are greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of a flying animal detection method provided by the embodiment of the present application is shown in the figure.

[0044] Figure 2 A first sub-flowchart of the flying animal detection method provided by the embodiment of the present application is shown in the figure.

[0045] Figure 3 A second sub-flowchart of the flying animal detection method provided by the embodiment of the present application is shown in the figure.

[0046] Figure 4 A third sub-flowchart of the flying animal detection method provided by the embodiment of the present application is shown in the figure.

[0047] Figure 5 A fourth sub-flow block diagram of the flying animal detection method provided by the embodiment of the present application is shown in the figure;

[0048] Figure 6 A composition block diagram of the flying animal detection system provided by the embodiment of the present application is shown in the figure;

[0049] Figure 7 A composition block diagram of the receiving module in the flying animal detection system provided by the embodiment of the present application is shown in the figure;

[0050] Figure 8 A composition block diagram of the amplifying module in the flying animal detection system provided by the embodiment of the present application is shown in the figure;

[0051] Figure 9 A composition block diagram of the inserting module in the flying animal detection system provided by the embodiment of the present application is shown in the figure;

[0052] Figure 10 A composition block diagram of the sending module in the flying animal detection system provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0054] In the embodiment 1, Figure 1 The flying animal detection method implementation process provided by the embodiment of the present application is shown in the figure, and the following is described in detail as follows:

[0055] S100: receiving a detection request uploaded by a user, extracting a picture to be detected, tracing back to a source device, finding an edge node, deploying a pre-constructed global target detection model to the edge node, inputting the picture to the global target detection model, and outputting a plurality of flying animal instances and candidate boxes.

[0056] The detection request uploaded by the user is received, and a picture to be detected is determined, wherein the picture can also be a video snapshot; for example, an automatic monitoring camera in a nature reserve collects image data at a fixed time every day and uploads it to a cloud platform, on a certain day, an administrator initiates a detection request through a management system to find out the flying animals appearing in the nature reserve on that day or in the recent period of time; if the data collected by the monitoring camera is video data, a plurality of snapshots are intercepted therefrom, and the flying animal detection is performed by using the snapshots.

[0057] The source device of the picture is determined, and according to the deployment position and network topology architecture of the source device, an edge node capable of being accessed by the source device is found, wherein the edge node can be an intelligent network or an edge server; a global target detection model is constructed, wherein the global target detection model is a target detection model for detecting flying animals in the picture in the prior art, the target detection model is constructed based on a convolutional neural network, and is iteratively trained through flying animal features; the global target detection model is deployed to the edge node, and the picture is input to the global target detection model, flying animals in the picture are identified through feature extraction and multi-layer depth convolution on the picture, and are output in the form of instances, wherein the instances refer to specific targets or individual objects in the picture, for example, people, trees and vehicles; the global target detection model also generates a plurality of candidate boxes, and the candidate boxes demarcate position regions in the picture that may contain flying animals.

[0058] S200: Calculate the brightness value of each pixel point by using the brightness formula, draw a two-dimensional frequency base graph composed of base functions, divide the picture into a plurality of blocks, quantize the blocks by using the two-dimensional frequency base graph, obtain the transform coefficients, and traverse the high-frequency signal region by using the preset brightness matrix and the chroma matrix, and perform interpolation magnification.

[0059] The brightness formula (Y=0.299R+0.587G+0.114B, Y represents the brightness value of a pixel, R, G and B are the intensity values of the red, green and blue color channels of the pixel, respectively) is used to calculate the brightness value of each pixel in the picture, and the picture is converted into a grayscale image; a two-dimensional frequency base function in the prior art is selected to decompose the grayscale image, and a corresponding two-dimensional frequency base graph is obtained, which shows the intensity distribution of the picture at different spatial frequencies; the picture is divided into a plurality of fixed-size blocks (such as 8x8 pixel blocks), a two-dimensional frequency base graph is used to perform discrete transformation on each block, a corresponding transform coefficient matrix is obtained, and the transform coefficients of each block are quantized by combining the brightness matrix and the chroma matrix in the prior art. The region with a transform coefficient greater than a threshold value or a non-“0” value is defined as a high-frequency signal region, and these regions often contain a large amount of detail information and edge features, which are the key to identifying smaller and more hidden flying animals; interpolation magnification technology is used to locally magnify the high-frequency signal region.

[0060] S300: Create a local target detection model and deploy it to a cloud server, input the high-frequency signal region into the local target detection model, output a plurality of flying animal instances and verification boxes, compare the candidate boxes and the verification boxes, calculate the intersection over union, obtain the intersection and the union, and insert the intersection into the flying animal instances.

[0061] A detection model dedicated to small-scale target recognition, i.e., a local target detection model, is constructed, wherein the global target detection model and the local target detection model are both constructed by a target detection model, and the difference between the two lies in the different training methods and training data. The local target detection model is trained by using the high-frequency signal region after interpolation magnification, so that the local target detection model has stronger edge feature perception ability and small target recognition precision. The local target detection model is deployed in a cloud server, and the high-frequency signal region is input into the local target detection model, and a plurality of flying animal instances and verification boxes are output. The verification box is similar to the candidate box, and refers to a position region that may contain a flying animal.

[0062] The candidate box and the verification box are compared, the region where the two coincide is defined as an intersection, the total region after merging is defined as a union, and an intersection-union ratio is calculated, wherein the intersection-union ratio refers to the number of pixels of the intersection divided by the number of pixels of the union; the instance in the region where the intersection is located is inserted into the flying animal instance.

[0063] S400: The edge node and the cloud server are integrated, a distributed processing architecture is constructed, the verification box is uploaded to the cloud server, the segment corresponding to the verification box is intercepted from the picture to generate a verification task, when a detection request is received, the verification task is issued, a verification result is obtained, the flying animal instance and the verification result are marked in the picture, and the picture is sent to a preset terminal.

[0064] The edge node and the cloud server are integrated, and a distributed processing architecture with cooperative computing capability is constructed. In the processing architecture, the edge node is mainly used to preliminarily detect the flying animals in the picture, and the cloud server is used to undertake tasks such as brightness calculation, high-frequency region extraction, interpolation magnification, candidate box and verification box comparison, and fine-grained detection of the picture. Further, after the edge node completes the preliminary processing, it uploads the detected suspected target region and the corresponding verification box to the cloud server. The cloud server has stronger computing resources and model library, which facilitates the deployment of a local target detection model with higher precision. The verification box from the edge node is received and subjected to processing operations such as depth recognition, intersection-union ratio analysis, target fusion and instance confirmation, and the processed verification box is uploaded to the cloud server to identify the flying animal instances therein again. The segment corresponding to the verification box is intercepted in the picture, and a verification task is generated, wherein the verification task refers to integrating the segment into the same thumbnail and writing the instruction: “Please find the flying animals in the following pictures” (similar to a picture verification code). When a detection request is received, the verification task is sent to the user, and the user selects the flying animals from the verification task. The selection result of the user is the verification result. All the flying animal instances and the verification result are marked in the picture and sent to a preset terminal, which is the device terminal of the user.

[0065] In embodiment 2,Figure 2 The implementation flow of the flying animal detection method provided by the embodiment of the application is shown. The steps of tracing the source device, finding the edge node, and deploying the pre-constructed global target detection model into the edge node are described in detail as follows.

[0066] S101: A scheduling platform of the edge node is constructed, and evaluation indexes of each edge node are collected, wherein the evaluation indexes at least include delay and available bandwidth.

[0067] The scheduling platform is constructed, which is mainly used for unified management and scheduling of all edge nodes. The scheduling platform continuously collects and dynamically updates the evaluation indexes of each edge node, wherein the evaluation indexes include communication delay, available bandwidth, node load, CPU utilization, and the like of the edge node.

[0068] S102: The weight value of each evaluation index is configured, the real-time value of the evaluation index is collected, the evaluation score is calculated, uploaded to the scheduling platform, and dynamically updated.

[0069] According to the importance of each evaluation index, the corresponding weight value is configured, which can be set according to the actual application scene requirement or formulated by the management personnel of the scheduling platform. The real-time value of the evaluation index of each edge node is collected, the real-time value of the evaluation index is multiplied by the weight value, and the calculation results of all are superimposed to obtain the evaluation score. The evaluation score is marked in the scheduling platform and dynamically updated. The higher the evaluation score is, the more excellent the performance of the corresponding edge node is. When the global target detection model is deployed, the edge node with a higher evaluation score is preferentially selected.

[0070] In the embodiment 3, different from the embodiment 1, the method further includes:

[0071] Based on the evaluation score, the edge nodes are clustered into ordinary nodes and super nodes, and the global target detection model is deployed into several super nodes.

[0072] The load of each super node is identified, and a load balancing mechanism is integrated into the scheduling platform.

[0073] According to the evaluation score, the edge node is divided into a common node and a super node, wherein the super node is the edge node with a higher evaluation score in the above description, and the global target detection model is deployed in the super node; according to the load, a plurality of super nodes can also be selected, and a global target detection model is deployed in each super node; the CPU usage rate, the memory occupation rate, the task queue length and the concurrent connection number are used as the load indexes, and the load state of the edge node is judged according to the load indexes; when the load of the edge node exceeds the threshold value, the new task is no longer preferentially assigned; the load balancing mechanism is integrated into the scheduling platform, and the load balancing mechanism refers to preferentially assigning the new task to the edge node with a smaller load; it should be noted that the task here refers to identifying the flying animal instance by using the global target detection model in the edge node.

[0074] In embodiment 4, Figure 3 The flying animal detection method implementation process provided by the embodiment of the application is shown, and the step of cutting the picture into a plurality of blocks, quantizing the blocks by using a two-dimensional frequency base graph, and obtaining the transform coefficient is described in detail as follows.

[0075] S201: Set the basic range of each block, and traverse the edge pixels in the picture by using an edge detection algorithm.

[0076] The basic range of each block is set, wherein the basic range can be 8*8, and the edge pixels in each picture are determined by using an edge detection algorithm.

[0077] S202: Count the number of edge pixels in each basic range, and adjust the basic range based on the number, wherein the adjustment includes: expansion and reduction.

[0078] The number of edge pixels in each basic range is calculated, when the number is greater than a threshold value, the basic range is reduced, that is, the basic range is adjusted to 4*4, otherwise, if the number is less than or equal to the threshold value, the basic range is expanded to 16*16.

[0079] In this embodiment, by expanding the basic range, the data processing complexity can be reduced and the processing efficiency can be improved, and by reducing the basic range, the local features can be captured and the detail richness can be improved.

[0080] In embodiment 5, Figure 4 The flying animal detection method implementation process provided by the embodiment of the application is shown, and the step of calculating the intersection-over-union, obtaining the intersection and the union, and inserting the intersection into the flying animal instance is described in detail as follows.

[0081] S301: Calculate the intersection over union of each picture, select a first set value and a second set value, define the picture with the intersection over union less than the first set value as a negative sample, and define the picture with the intersection over union greater than the second set value as a positive sample.

[0082] In each picture, determine the candidate box and the verification box, and calculate the intersection over union, wherein each picture corresponds to an intersection over union; according to the intersection over union, select two set values, i.e., a first set value and a second set value, define the picture with the intersection over union less than the first set value as a negative sample, and define the picture with the intersection over union greater than the second set value as a positive sample.

[0083] For example, assuming that the first set value is 0.4 and the second set value is 0.5, if the intersection over union of a certain picture is less than 0.4, it indicates that the global target detection model or the local target detection model has a low recognition accuracy for the picture, and there are more false detections; if the intersection over union is greater than 0.5, it indicates that the model (the target detection model and the local target detection model are collectively referred to as the model) has a high recognition accuracy, and when the intersection over union of the picture is between 0.4 and 0.5, the corresponding picture is not used for training, in other words, using such a picture for training cannot improve the recognition ability of the model.

[0084] S302: Integrate the positive samples and the negative samples to generate a training set, and train the global target detection model and the local target detection model.

[0085] The positive samples and the negative samples are used to generate a training set, and the training is performed; in the actual model training process, the positive samples are used as targets to enable the model to learn "what is a target", and the negative samples are also used to train the model to recognize "what is not a target".

[0086] In embodiment 6, Figure 5 The implementation process of the flying animal detection method provided in the embodiment of the application is shown, and the following describes the step of issuing a verification task and obtaining a verification result when a search request is received:

[0087] S401: Set a plurality of distribution channels for the verification task, and calculate the confidence score of each verification box based on the verification result.

[0088] A plurality of distribution channels for the verification task are set, each distribution channel represents a specific use scenario or platform (such as a web page, a mobile terminal, a third-party interface, etc.), the verification task is distributed to the user through the distribution channel; that is, a thumbnail is generated using the segment corresponding to the verification box, and is sent to the user in the form of a verification code; the verification result of the user is obtained, the behavior data of the user in the verification process is collected, and the confidence score of each verification box is determined, for example, when the verification box is selected, 1 score value is added to the confidence score of the corresponding verification box.

[0089] S402: Mark the verification result with the confidence score greater than the threshold to the picture.

[0090] After a plurality of users perform the verification code processing, the confidence score of each verification frame is calculated, and the verification result with the confidence score greater than the threshold is marked to the picture.

[0091] For example, an automatic monitoring camera (number A) in a nature reserve collects image data at a fixed time every day and uploads it to a cloud platform. On a certain day, administrator A initiates a detection request through a management system to find out the flying animals appearing in the nature reserve in the last 5 hours; at this time, A needs to verify through a verification code; the verification code is a thumbnail of the verification frame of the picture collected by other monitoring cameras or devices. It should be noted that the verification code is not composed of the verification frame of the picture collected by A. If the verification frame in the picture collected by A is not verified in a timely manner through the verification code (the reason may be that the manual verification resource is valid, the network is congested, etc.) when A initiates the detection request, the user is sent a picture only marked with a flying animal instance (without marking the verification frame and the verification result).

[0092] Figure 6 The composition structure block diagram of the flying animal detection system provided by the embodiment of the application is shown, and the flying animal detection system 1 comprises:

[0093] The receiving module 11 is configured to receive a search request uploaded by a user, extract a picture that needs to be detected, trace back to a source device, find out an edge node, deploy a pre-constructed global target detection model to the edge node, input the picture into the global target detection model, and output a plurality of flying animal instances and candidate frames;

[0094] The magnification module 12 is configured to calculate the brightness value of each pixel point by using a brightness formula, draw a two-dimensional frequency base graph composed of a base function, divide the picture into a plurality of blocks, quantize the blocks by using the two-dimensional frequency base graph, obtain a transformation coefficient, traverse a high-frequency signal region by using a preset brightness matrix and a chroma matrix, and perform interpolation magnification;

[0095] The insertion module 13 is configured to create a local target detection model and deploy the local target detection model to a cloud server, input the high-frequency signal region into the local target detection model, output a plurality of flying animal instances and verification frames, compare the candidate frames and the verification frames, calculate an intersection over union, obtain an intersection and a union, and insert the intersection into the flying animal instances;

[0096] The sending module 14 is configured to integrate the edge node and the cloud server, construct a distributed processing architecture, upload the verification frame to the cloud server, extract a segment corresponding to the verification frame from the picture, generate a verification task, distribute the verification task when receiving a search request, obtain a verification result, mark the flying animal instance and the verification result to the picture, and send the picture to a preset terminal.

[0097] Figure 7 The receiving module 11 is shown in the component structure block diagram of the flying animal detection system provided by the embodiment of the application.

[0098] The acquisition unit 111 is configured to construct a scheduling platform of the edge node, and acquire an evaluation index of each edge node, wherein the evaluation index at least includes a delay and an available bandwidth.

[0099] The updating unit 112 is configured to configure a weight value of each evaluation index, acquire a real-time value of the evaluation index, calculate an evaluation score, upload the evaluation score to the scheduling platform, and dynamically update.

[0100] Figure 8 The amplification module 12 is shown in the component structure block diagram of the flying animal detection system provided by the embodiment of the application.

[0101] The traversal unit 121 is configured to set a basic range of each block, and traverse an edge pixel in the picture by using an edge detection algorithm.

[0102] The statistical unit 122 is configured to count a number of the edge pixels in each basic range, and adjust the basic range based on the number, wherein the adjustment includes expansion and reduction.

[0103] Figure 9 The insertion module 13 is shown in the component structure block diagram of the flying animal detection system provided by the embodiment of the application.

[0104] The calculation unit 131 is configured to calculate an intersection over union of each picture, select a first set value and a second set value, define a picture with an intersection over union less than the first set value as a negative sample, and define a picture with an intersection over union greater than the second set value as a positive sample.

[0105] The training unit 132 is configured to integrate the positive sample and the negative sample, generate a training set, and train the global target detection model and the local target detection model.

[0106] Figure 10 The sending module 14 is shown in the component structure block diagram of the flying animal detection system provided by the embodiment of the application.

[0107] The setting unit 141 is configured to set a plurality of distribution channels of the verification task, and calculate a confidence score of each verification frame based on the verification result;

[0108] The marking unit 142 is configured to mark the verification result with the confidence score greater than a threshold to the picture.

[0109] The receiving module 11 is mainly configured to complete step S100, the amplifying module 12 is mainly configured to complete step S200, the inserting module 13 is mainly configured to complete step S300, and the sending module 14 is mainly configured to complete step S400;

[0110] The collecting unit 111 is mainly configured to complete step S101, and the updating unit 112 is mainly configured to complete step S102;

[0111] The traversing unit 121 is mainly configured to complete step S201, and the counting unit 122 is mainly configured to complete step S202;

[0112] The calculating unit 131 is mainly configured to complete step S301, and the training unit 132 is mainly configured to complete step S302;

[0113] The setting unit 141 is mainly configured to complete step S401, and the marking unit 142 is mainly configured to complete step S402.

[0114] The above-described embodiments are merely representative of the technical features of the present application, and the description is relatively specific and detailed, but should not be construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

[0115] The above-described embodiments merely express several embodiments of the present application, which are described in detail, but should not be construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

[0116] The above-described embodiments merely express several embodiments of the present application, which are described in detail, but should not be construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for detecting flying animals, characterized in that, The method includes: The system receives detection requests uploaded by users, extracts the images to be detected, traces back to the source device, finds edge nodes, deploys the pre-built global object detection model to the edge nodes, inputs the images into the global object detection model, and outputs several flying animal instances and candidate boxes. Using the luminance formula, the luminance value of each pixel is calculated, a two-dimensional frequency base map composed of basis functions is drawn, the image is divided into several blocks, the blocks are quantized using the two-dimensional frequency base map to obtain the transformation coefficients, and the high-frequency signal region is traversed using the preset luminance matrix and chrominance matrix, and interpolation amplification is performed. A local target detection model is created and deployed to a cloud server. The high-frequency signal region is input into the local target detection model, and several flying animal instances and verification boxes are output. The candidate boxes and verification boxes are compared, the intersection-union ratio is calculated, and the intersection and union are obtained. The intersection is then inserted into the flying animal instances. By integrating edge nodes and cloud servers to build a distributed processing architecture, the verification box is uploaded to the cloud server. The corresponding fragment of the verification box is extracted from the image to generate a verification task. When a detection request is received, the verification task is issued, the verification result is obtained, the flying animal instance and the verification result are marked on the image and sent to the preset terminal. The steps of tracing back to the source device, finding the edge nodes, and deploying the pre-built global object detection model to the edge nodes include: Build a scheduling platform for edge nodes and collect evaluation metrics for each edge node, wherein the evaluation metrics include at least: latency and available bandwidth; Configure the weight value of each evaluation indicator, collect the real-time value of the evaluation indicator, calculate the evaluation score, upload it to the scheduling platform, and update it dynamically.

2. The method for detecting flying animals according to claim 1, characterized in that, The method further includes: Based on the evaluation scores, edge nodes are clustered into ordinary nodes and super nodes, and the global object detection model is deployed to several super nodes; Identify the load of each supernode and integrate a load balancing mechanism into the scheduling platform.

3. The method for detecting flying animals according to claim 1, characterized in that, The steps of dividing the image into several blocks and quantizing the blocks using a two-dimensional frequency base map to obtain transform coefficients include: Set the basic range for each block, and use an edge detection algorithm to traverse the edge pixels in the image; The number of edge pixels within each base range is counted, and the base range is adjusted based on the count, wherein the adjustment includes expanding and shrinking.

4. The method for detecting flying animals according to claim 1, characterized in that, The steps of calculating the intersection-union ratio, obtaining the intersection and union, and inserting the intersection into the flying animal instance include: Calculate the crossover ratio (CRR) of each image, select a first set value and a second set value, define images with an CRR less than the first set value as negative samples, and define images with an CRR greater than the second set value as positive samples; Positive and negative samples are integrated to generate a training set, which is then used to train the global object detection model and the local object detection model.

5. The method for detecting flying animals according to claim 1, characterized in that, The steps of issuing a verification task and obtaining the verification result when a detection request is received include: Set up several distribution channels for the verification task, and calculate the confidence score of each verification box based on the verification results; The verification results with confidence scores greater than the threshold are marked on the image.

6. A flying animal detection system, characterized in that, The system includes: The receiving module is used to receive detection requests uploaded by users, extract the images to be detected, trace back to the source device, find the edge nodes, deploy the pre-built global object detection model to the edge nodes, input the images into the global object detection model, and output several flying animal instances and candidate boxes. The magnification module is used to calculate the brightness value of each pixel using the brightness formula, draw a two-dimensional frequency base map composed of basis functions, divide the image into several blocks, quantize the blocks using the two-dimensional frequency base map to obtain the transformation coefficients, and use the preset brightness matrix and chromaticity matrix to traverse the high-frequency signal region and perform interpolation magnification. The insertion module is used to create a local target detection model and deploy it to a cloud server. The high-frequency signal region is input into the local target detection model, and several flying animal instances and verification boxes are output. The candidate boxes and verification boxes are compared, the intersection-union ratio is calculated, the intersection and union are obtained, and the intersection is inserted into the flying animal instances. The sending module is used to integrate edge nodes and cloud servers to build a distributed processing architecture. It uploads the verification box to the cloud server, extracts the corresponding fragment from the image, generates a verification task, sends the verification task when a detection request is received, obtains the verification result, marks the flying animal instance and the verification result in the image, and sends it to the preset terminal. The receiving module includes: The acquisition unit is used to build a scheduling platform for edge nodes and acquire evaluation metrics for each edge node, wherein the evaluation metrics include at least: latency and available bandwidth; The update unit is used to configure the weight value of each evaluation indicator, collect the real-time value of the evaluation indicator, calculate the evaluation score, upload it to the scheduling platform, and perform dynamic updates.

7. The flying animal detection system according to claim 6, characterized in that, The amplification module includes: The traversal unit is used to set the basic range of each block and uses an edge detection algorithm to traverse the edge pixels in the image; A statistics unit is used to count the number of edge pixels within each base range, and to adjust the base range based on the count, wherein the adjustment includes expanding and shrinking.

8. The flying animal detection system according to claim 6, characterized in that, The insertion module includes: The calculation unit is used to calculate the crossover ratio (CVR) of each image, select a first set value and a second set value, define images with a CVR less than the first set value as negative samples, and define images with a CVR greater than the second set value as positive samples. The training unit is used to integrate positive and negative samples to generate a training set for training the global object detection model and the local object detection model.

Images