Small target detection method and system

Abstract

The invention discloses a small target detection method and system, the method is specially designed for server-side processing of image acquisition of an unmanned aerial vehicle, the unmanned aerial vehicle is only responsible for image acquisition and transmission, and all image processing and target detection are completed on a server. According to the method, intrusion identification of people and vehicles and early detection of petroleum leakage are divided into two independent modules: for intrusion identification of people and vehicles, an improved YOLOv8 network is adopted; for early detection of petroleum leakage, a ResNeXt network is selected for processing. According to the technical scheme, the small target detection effect in the unmanned aerial vehicle image can be improved, the calculation resources of the model cannot be obviously increased by introducing the module, and the detection performance and the calculation resources can be better balanced; moreover, an effective solution is provided for the current situation that the petroleum leakage detection technology is deficient in the desert environment, the technical blank in the field is filled, and the accuracy of desert petroleum leakage detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a small target detection method and system based on an improved YOLOv8 algorithm and a ResNeXt network. Background Art

[0002] Unmanned aerial vehicle (UAV) target detection plays a vital role in civil, commercial and military fields. The proportion of small targets in the images acquired by UAVs is high, while the platform resources are limited, resulting in the low accuracy of most detection models embedded in existing UAVs, making it difficult to achieve a good balance between detection performance and resource consumption. In addition, as an important environmental protection task, the monitoring and detection of oil spills has made great progress in the fields of satellite remote sensing and ship monitoring, but the application of UAVs in this field is relatively small.

[0003] In the field of target detection, especially in the task of detecting small targets in drone aerial photography, traditional target detection algorithms face a series of unique challenges. Due to the changes in the height and angle of drone photography, small targets in images usually have low resolution, blurred outlines, and complex background interference. These factors make it particularly difficult to accurately identify small targets.

[0004] To address the above challenges, researchers have proposed a variety of methods to improve the detection performance of small objects, including:

[0005] 1) Multi-scale feature fusion: By combining feature maps extracted by networks at different levels, the response to small targets is enhanced.

[0006] 2) High-resolution input: Use higher-resolution images as input to improve the visibility and recognizability of small objects.

[0007] 3) Attention mechanism: Introduce the attention module to highlight important areas and help the model better focus on small targets.

[0008] 4) Data enhancement technology: such as random cropping and scaling, to increase the diversity of training data and improve the robustness of the model.

[0009] 5) Cascade structure: A multi-stage detection framework is adopted to gradually refine the prediction results and improve positioning accuracy.

[0010] Although the above methods have improved the effect of small target detection to a certain extent, they are often accompanied by increased computational complexity or require more training resources. For example, Wise-IoU (WIoU) v3 was introduced into the YOLOv8 model as a new bounding box regression loss, and the backbone network was optimized with an attention mechanism called BiFormer to form the UAV-YOLOv8 model. This model is specially designed to handle small target detection problems in drone aerial photography scenarios. Although it significantly improves the overall performance, it still has shortcomings when dealing with extremely small objects, and its relatively complex structure also increases the computational cost.

[0011] Some people also introduced the idea of ​​Bi-PAN-FPN to improve the neck part of YOLOv8-s. The GhostblockV2 structure was used in the backbone of the benchmark model to replace part of the C2f module, which significantly reduced the number of model parameters while suppressing information loss during long-distance feature transmission. However, although its overall performance has been improved, it has not achieved better results than other structures in all small categories.

[0012] In addition, in the application of oil leak detection technology, marine oil leaks cause serious pollution to the environment, so in the marine environment, oil leak detection and monitoring technology has become relatively mature. The main detection methods include:

[0013] 1) Remote sensing technology: Satellite remote sensing and aerial remote sensing are used to detect oil films on the ocean surface. Synthetic aperture radar (SAR) is widely used in marine oil pollution monitoring because it is not affected by weather and light.

[0014] 2) Optical imaging: High-resolution optical satellite and aircraft imaging can identify oil-contaminated areas on the sea surface.

[0015] 3) Hyperspectral imaging: Identify the spectral characteristics of oil pollution by analyzing the spectral information of different bands.

[0016] 4) Drone monitoring: Use drones to carry sensors to conduct close-range, high-precision monitoring of oil pollution on the ocean surface.

[0017] These techniques rely on the characteristics of the marine environment, such as the significant difference in dielectric constant between the oil film and seawater, the smooth sea surface background, and the specific response of the oil film to radar and optical signals.

[0018] In contrast, oil leak detection technology in desert environments is relatively scarce, mainly due to the following reasons:

[0019] 1) Difficulty in data acquisition: Desert areas are vast and sparsely populated, with harsh environments, making it difficult to obtain real data on oil spills and conduct on-site monitoring.

[0020] 2) Complex background: The desert terrain is changeable, with various landform features such as sand dunes, rocks, and vegetation, which increases the difficulty of image recognition.

[0021] 3) Lack of dedicated algorithms: Most of the existing oil leak detection algorithms are designed for marine environments and are not very effective when directly applied to desert environments. There is a lack of dedicated detection algorithms for desert characteristics.

[0022] Therefore, the differences between ocean and desert environment images are mainly reflected in:

[0023] 1) Background texture difference. That is, the ocean surface is relatively flat and has uniform texture, so the reflection characteristics of the oil film on the sea surface are obvious. However, the desert surface has complex texture, consisting of various terrains and materials such as sand dunes, rocks, and vegetation, which makes the difference between the oil pollution and the background unclear, increasing the difficulty of identification.

[0024] 2) Differences in optical properties. In other words, strong solar radiation and frequent dust storms in desert areas lead to strong noise and illumination changes in images. Strong illumination can produce overexposure or shadow areas, while dust can cause image blur, increasing the complexity of detection.

[0025] 3) Differences in oil pollution forms:

[0026] 3.1) Marine oil pollution form: Marine oil pollution usually appears as a continuous oil film covering the flat sea surface. The shape is relatively uniform and easy to identify through optical or radar characteristics.

[0027] 3.2) Desert oil spill patterns: Due to the influence of terrain, such as undulating sand dunes and surface cracks, oil spills in the desert often form unique branching flow traces. When oil flows on the desert surface, it will be affected by factors such as terrain slope, sand grain size and wind force, resulting in a bifurcated and branching flow pattern. These branching structures are changeable and difficult to be identified by traditional detection algorithms.

[0028] 4) Importance of bifurcation features: The bifurcation features of desert oil spills are the key morphological features that distinguish them from marine oil spills and are also the key to effective detection. Identifying these branching flow traces requires the development of detection algorithms that take into account the unique behavior patterns of oil spills in desert environments.

[0029] In summary, the detection efficiency of existing drones for small targets including people, vehicles, oil leaks, etc. is relatively low, and it is difficult to strike a balance between detection performance and computing resources. Summary of the invention

[0030] In order to overcome the shortcomings of the prior art, the present invention provides a small target detection method, which can improve the small target detection efficiency in drone images, efficiently detect human and vehicle intrusion and oil leakage from real-time drone videos, and the introduction of the module will not significantly increase the computing resources of the model, so that a better balance can be achieved between detection performance and computing resources; and, in view of the current lack of oil leakage detection technology in desert environments, an effective solution is provided, which fills the technical gap in this field and improves the accuracy of desert oil leakage detection.

[0031] A small target detection method, comprising:

[0032] S1, obtaining the image to be detected;

[0033] S2, using the improved YOLOv8 model or the ResNeXt model to process the image to be detected, or using the improved YOLOv8 model and the ResNeXt model in sequence to process the image to be detected, or using the ResNeXt model and the improved YOLOv8 model in sequence to process the image to be detected; wherein,

[0034] Using the improved YOLOv8 model to process the image to be detected includes:

[0035] S201, performing multi-scale feature extraction on the image to be detected through a DWRSeg network constructed by a DWR module and a SIR module;

[0036] S202, regional residualization: generating a plurality of regional feature maps of different sizes through a convolutional layer of a preset size;

[0037] S203, semantic residualization: applying a multi-rate dilated depth convolution to the regional feature map to perform morphological filtering to capture semantic information of different scales;

[0038] S204, sampling features from the deformed position on the regional feature map using bilinear interpolation;

[0039] Using the ResNeXt model to process the image to be detected includes:

[0040] S211, generating oil bifurcation simulation data;

[0041] S212, performing feature extraction and bifurcation feature capture on the petroleum bifurcation simulation data and the image to be detected through a ResNeXt network;

[0042] S3, integrating the detection results of the improved YOLOv8 model and / or the ResNeXt model to generate a detection report.

[0043] A small target detection system includes a drone end and a server end; wherein,

[0044] The drone end is used to obtain the image to be detected and transmit the image to the server end via a wireless network;

[0045] The server side is used to process the image to be detected using an improved YOLOv8 model or a ResNeXt model, or to process the image to be detected using the improved YOLOv8 model and the ResNeXt model in sequence, or to process the image to be detected using the ResNeXt model and the improved YOLOv8 model in sequence; wherein;

[0046] Using the improved YOLOv8 model to process the image to be detected includes:

[0047] S201, performing multi-scale feature extraction on the image to be detected through a DWRSeg network constructed by a DWR module and a SIR module;

[0048] S202, regional residualization: generating a plurality of regional feature maps of different sizes through a convolutional layer of a preset size;

[0049] S203, semantic residualization: applying a multi-rate dilated depth convolution to the regional feature map to perform morphological filtering to capture semantic information of different scales;

[0050] S204, sampling features from the deformed position on the regional feature map using bilinear interpolation;

[0051] Using the ResNeXt model to process the image to be detected includes:

[0052] S211, generating oil bifurcation simulation data;

[0053] S212, performing feature extraction and bifurcation feature capture on the petroleum bifurcation simulation data and the image to be detected through a ResNeXt network;

[0054] The improved YOLOv8 model and / or the ResNeXt model detection results are integrated to generate a detection report.

[0055] The small target detection method and system provided by the present invention are designed for server-side processing of images collected by drones. The drones are only responsible for image collection and transmission, and all image processing and target detection are completed on the server. The method divides the intrusion identification of people and vehicles and the detection of oil leaks into two independent modules, which are processed by improved YOLOv8 and ResNeXt networks respectively to improve the detection accuracy and efficiency, and improve the efficiency of feature extraction, especially for the detection of small targets (such as people, vehicles, oil leaks) collected by drones. More accurate and efficient.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] 1) Accurate detection: Use specialized models for different targets to improve detection accuracy.

[0058] 2) Efficient computing: Server-side processing reduces the load on drones and makes full use of computing resources.

[0059] 3) Modular design: The two detection models run independently, which is easy to maintain and upgrade.

[0060] 4) Strong adaptability: The model design takes into account the particularity of the desert environment and enhances its adaptability to complex backgrounds.

[0061] At the same time, in terms of application scenarios and expansion, it also has the following characteristics:

[0062] 1) Environmental monitoring: timely detect oil leaks in the desert and protect the ecological environment.

[0063] 2) Security protection: Detect illegal intruders and vehicles to ensure the safety of facilities.

[0064] 3) Scalability: The model structure can be extended to detect other targets, such as wildlife monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Flow chart of a small target detection method in an embodiment of the present invention;

[0066] Figure 2 This is a training flow chart of a desert oilfield leakage detection model in an embodiment of the present invention;

[0067] Figure 3 This is an image to be detected of a desert oil field leak in an embodiment of the present invention (not identified by initial model reasoning);

[0068] Figure 4 It is the inference result of the pre-trained ocean leakage YOLOv8 model in the embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0070] A small target detection method and system are provided. The method is specially designed for server-side processing of images collected by drones, that is, after obtaining the image to be detected collected and transmitted by the drone, the target in the image is detected.

[0071] Specifically, the overall system architecture includes the drone side and the server side. The drone side is responsible for collecting high-resolution images and transmitting them to the server via a wireless network. The server side contains two independent target detection modules: an improved YOLOv8 model for intrusion detection of people and vehicles, and a ResNeXt-based model for oil spill detection.

[0072] During the detection process, the server performs multi-task processing and result fusion, including:

[0073] Parallel processing: The server can run two detection models at the same time to detect human and vehicle intrusion and oil leakage respectively on the same image, or only use one of them to process the image.

[0074] Result fusion: The detection results of the two models are integrated to generate a complete detection report. If people, vehicles and oil leaks are detected in the same area, the alarm level can be raised and emergency measures can be taken in time.

[0075] In one embodiment, a drone is used to take aerial photos in a desert oil field area to obtain high-resolution photos of the oil field. These images contain the terrain in the actual desert environment and possible traces of oil leakage. The flowchart of the small target detection method and system is as follows: Figure 1 As shown, the main steps include:

[0076] 1) UAV image acquisition: The UAV cruises over the target area and collects high-resolution images.

[0077] 2) Image transmission: transmit the image to the server in real time.

[0078] 3) Image preprocessing: The server grayscales, normalizes, and performs data enhancement on the image.

[0079] 4) Object Detection:

[0080] Human and vehicle intrusion detection: Use the improved YOLOv8 model to detect the intrusion of people and vehicles.

[0081] Oil Spill Detection: Detect oil spills in the desert using the ResNeXt101 model.

[0082] 5) Result fusion and analysis: Integrate the test results and generate reports.

[0083] 6) Emergency response: Take appropriate measures based on the test results.

[0084] Among them, image preprocessing is performed on the server side to preprocess the received images to meet the needs of the two detection models. Specifically, it includes: Grayscale processing: converting color images into grayscale images to simplify calculations. Data enhancement: using random cropping, rotation, scaling and other methods to enhance the robustness of the model. Normalization processing: normalizing the image to unify the pixel value range.

[0085] In addition, in oil leak detection, image preprocessing also includes: Image correction: perform distortion correction on the collected images to eliminate image deformation caused by the lens and posture. Image stitching: stitch multiple aerial images to generate a large-scale panoramic view. Image enhancement: Contrast enhancement: improve the contrast of the image to make the traces of oil leaks more obvious. Noise reduction: use filtering and other technologies to reduce the impact of noise on image quality. Grayscale processing: convert color images into grayscale images, focus on brightness information, and reduce the impact of background noise.

[0086] Use the improved YOLOv8 model for intrusion detection of people and vehicles, including:

[0087] S201, introduces the Dilation-wise Residual (DWR) module and the Simple Inverted Residual (SIR) module, and constructs the DWRSeg network for multi-scale feature extraction to improve feature extraction capabilities. The DWR module is responsible for extracting high-level semantic features and using dilated convolution to capture large-scale information; the SIR module is responsible for extracting low-level detail features and improving computational efficiency.

[0088] S202, region residualization, for example, using 3×3 convolution, batch normalization (BN) and ReLU activation function to generate regional feature maps of different sizes.

[0089] S203, Semantic Residualization, applies multi-rate dilated depth convolution to the regional feature map, performs morphological filtering, and captures semantic information at different scales.

[0090] S204, through the deformable self-attention module, uses bilinear interpolation to sample features from deformed positions on the feature map. Its structure is similar to multi-head self-attention, but adds the flexibility of deformable sampling, which can enhance the model's detection ability for small targets such as people and vehicles, as well as suppress background interference and improve detection accuracy.

[0091] The above S202 to S204 are equivalent to providing an efficient multi-scale feature extraction method, which simplifies the role of multi-rate depth-separable dilated convolution by decomposing the single-step process into two parts: regional residualization and semantic residualization, improves the feature extraction efficiency, and combines with the variable self-attention module to efficiently detect small targets such as people, vehicles and oil field leaks.

[0092] Furthermore, after S204, the method further includes:

[0093] S205, head detection and training: predicting target information through a preset head network suitable for human and vehicle detection, and performing training optimization based on the target information. Head detection is to design a head network suitable for human and vehicle detection to predict the category and position of the target. Training includes adopting a loss function and an optimization method; that is, adopting an improved loss function can improve the accuracy of bounding box regression and classification; using optimization methods such as cosine annealing learning rate and regularization can prevent overfitting.

[0094] The image to be detected is processed using the ResNeXt model, including:

[0095] S211, generating oil bifurcation simulation data. Specifically, it includes two stages: fluid simulation and data generation.

[0096] In the fluid simulation stage, fluid mechanics simulation tools such as OpenFOAM and COMSOL are used to simulate the flow process of liquid on the desert terrain model. Among them, parameter settings include physical parameter adjustment and terrain feature setting. Physical parameter adjustment adjusts the physical parameters of the liquid such as viscosity and density to make the simulated flow pattern close to the leakage behavior of oil in the desert. Terrain feature setting is used to build a real desert terrain model, including landform features such as sand dunes and cracks. The simulation process includes at least two types: branch flow mode and dynamic simulation: the branch flow mode refers to simulating the bifurcation and branch-like flow traces of liquid on the desert surface to reproduce the morphological characteristics of oil leakage. Dynamic simulation refers to obtaining the dynamic process of liquid flow to capture the leakage morphology at different time points.

[0097] The data generation stage includes at least: key frame extraction and data enhancement.

[0098] Keyframe extraction refers to extracting keyframes from fluid simulation, generating a large number of high-resolution images containing bifurcation traces, and constructing a simulated oil spill dataset.

[0099] Data augmentation includes image transformation and illumination simulation; image transformation refers to rotating, scaling, flipping, and other operations on the simulated image to increase data diversity. Illumination simulation refers to adjusting the brightness and contrast of the image to simulate different illumination conditions and improve the model's robustness to illumination changes.

[0100] S212, extracting features and capturing bifurcation features of the oil bifurcation simulation data and the image to be detected through the ResNeXt network.

[0101] That is, using ResNeXt101 as the feature extraction network and utilizing its multi-branch parallel convolution and Cardinality characteristics can enhance the model's feature extraction capabilities. Among them, multi-branch parallel convolution divides the convolution into multiple independent branches, each branch processes a set of feature maps, increasing the width of the network without increasing the computational complexity. As a result, parallel branches can capture more diverse features, improve the model's ability to recognize complex forms, and improve performance. Cardinality represents the number of parallel branches. By increasing Cardinality, the model's expressiveness is improved, and the number of parameters and computational costs are kept relatively stable.

[0102] Feature extraction and bifurcation feature capture specifically include:

[0103] Grayscale image processing: Input preprocessed grayscale images, focus on brightness information, reduce color interference, and improve sensitivity to texture and shape features. Bifurcation feature extraction: Feature layer optimization: Focus on optimizing the feature representation of the middle layer to enhance the ability to capture bifurcations and branching morphologies. Custom convolution kernel: Design a convolution kernel suitable for extracting bifurcations to capture the unique texture and morphological features of oil spills. Introduction of attention mechanism: Channel attention: Emphasize the response to important feature channels and increase the model's attention to key features. Spatial attention: Focus on areas where there may be traces of leakage to suppress background interference.

[0104] Further, after S212, if Figure 2 As shown, it also includes: S213, training and optimizing the results of feature extraction and bifurcation feature capture. Training optimization includes three parts: pre-training and transfer learning, loss function and optimization strategy, and data mixing and enhancement. In this embodiment, the schematic diagram of the desert oil field leakage is as follows Figure 3 As shown, the inference results of the pre-trained ocean leakage yolov8 model are as follows Figure 4 shown.

[0105] Among them, pre-training and transfer learning include:

[0106] S231, obtains a large number of marine oil pollution images from the public marine oil spill dataset to provide data support for model pre-training. The ResNeXt101 model is initially trained through the collected preset marine oil spill image dataset, and the basic texture, morphology and edge features of oil spills are learned.

[0107] S232, migrate the initially trained model parameters to the desert oil spill scenario and retain the shared feature representation. Specifically, it includes:

[0108] Fine-tune the model: Use real aerial data and simulated data in desert environments to fine-tune the model;

[0109] Adjust the learning rate: During fine-tuning, use a lower learning rate to prevent destroying the effective features of the pre-trained model;

[0110] Layer-level fine-tuning: Focus on fine-tuning high-level feature layers to adapt to the bifurcation characteristics of the desert background and oil spill.

[0111] Improve generalization ability: The pre-trained model has better initial parameters and can better adapt to the desert environment after fine-tuning.

[0112] Accelerate training convergence: Transfer learning shortens training time and speeds up model convergence.

[0113] Feature sharing: Ocean and desert oil spills have some commonalities in texture and morphology, and pre-training helps the model capture these common features.

[0114] The loss function and optimization strategy include:

[0115] S233, improves the accuracy of bounding box positioning and the ability to distinguish oil spill categories through a preset loss function;

[0116] S234, optimizing the training results by any one or more of learning rate scheduling, regularization technology, and small batch training.

[0117] Specifically, the loss function includes bounding box regression loss and classification loss; Bounding box regression loss: using GIoU (Generalized IoU) or DIoU (Distance IoU) loss function to improve the accuracy of bounding box positioning, especially for irregularly shaped targets. Classification loss: using weighted cross entropy loss to address the problem of class imbalance and improve the model's ability to distinguish oil spill categories.

[0118] Optimization strategies include learning rate scheduling, regularization techniques, and small batch training: Learning rate scheduling: Use cosine annealing or learning rate decay strategies to optimize the training process. Regularization techniques: Use weight decay, Dropout, and other methods to prevent model overfitting. Small batch training: Appropriately adjust the batch size to stabilize gradient updates and improve training results.

[0119] Data blending and augmentation include:

[0120] S235, combines ocean pre-training, desert simulation data and a small amount of real desert data and allocates different data source ratios;

[0121] S236, optimizing the training result by any one or more of random cropping, random scaling, adding noise, imaging blur, brightness and contrast adjustment.

[0122] Specifically, data hybrid training includes: Combining real and simulated data: combining ocean pre-training, desert simulation data and a small amount of real desert data to enrich training data and improve the generalization ability of the model. Data balance: reasonably allocating the proportion of different data sources in the training set to prevent the model from being biased towards a certain type of data.

[0123] Data enhancement includes: Random cropping and scaling: simulating different camera perspectives and target scales to enhance the robustness of the model. Adding noise and blurring: simulating sandstorms and imaging blur to improve the model's adaptability to harsh environments. Brightness and contrast adjustment: simulating different lighting conditions to enhance the model's robustness to lighting changes.

[0124] It can be understood that the technical solution provided in the present invention has the following outstanding features:

[0125] 1. Innovation in multi-scale feature extraction based on improved YOLOv8: The traditional YOLO model has limitations in detecting small targets. By decomposing the feature extraction process into two steps: regional residualization and semantic residualization, the multi-rate deep convolution process is simplified. The efficiency of feature extraction is improved, especially for the detection of small targets (such as people, vehicles, and oil spills) collected by drones, which is more accurate and efficient.

[0126] 2. Collaborative design of division of labor between drones and servers: The present invention specially designs an architecture for drones to collect images and process them on the server side. After the drone images are transmitted, the server side efficiently runs the improved YOLOv8 and ResNeXt101 models. This division of labor and collaborative architecture design greatly optimizes the utilization of limited drone resources, while achieving complex detection tasks through the powerful computing power of the server.

[0127] 3. Detection technology for bifurcation morphological characteristics of desert oil spills. Morphological feature analysis: In-depth study of bifurcation and branching flow characteristics of desert oil spills, development of targeted algorithms, and improvement of the accuracy of detection of such complex morphologies; Algorithm optimization: In model training and result interpretation, focus on the identification and labeling of bifurcation structures, and optimize the model's adaptability to the unique morphology of desert spills.

[0128] 4. Detection method combining simulation data with deep learning: Using simulated data for model training breaks through the bottleneck of obtaining real data on desert oil leaks; a complete solution from fluid simulation, data generation, model training to practical application is proposed, providing new ideas for desert oil leak detection.

[0129] 5. Integration and innovation of ocean and desert oil leak detection technologies: Based on the analysis of the differences between ocean and desert oil leak images, the detection method was innovatively adjusted to meet the particularities of the desert environment. In view of the complex background and lighting conditions of the desert environment, the image preprocessing and model training methods were improved to improve the robustness of detection.

[0130] 6. Flexibility and reliability of multi-mode detection: Traditional petroleum detection equipment is mostly based on a single mode. Once a certain mode fails, the overall detection effect will be significantly affected. This project has designed a multi-mode detection solution, integrating 5G real-time transmission, memory card offline data upload, edge computing local detection and other means to ensure continuous monitoring capabilities under different network environments. This multi-mode, full-coverage detection method provides a strong guarantee for responding to various emergencies.

[0131] 6. Efficient collaborative detection based on 5G network: Although 5G technology has been applied in many fields, its combination in oil leak detection is still relatively rare. The ultra-high speed and low latency characteristics of 5G network enable drones to transmit high-definition images in real time in the desert and conduct large-scale data analysis and model training synchronously with the back-end system. This collaborative working mode breaks through the bottleneck of existing oil detection methods in data collection and analysis, making the detection process more efficient.

[0132] The above is an explanation of the small target detection method and system of the present invention, which is used to help understand the present invention; however, the implementation mode of the present invention is not limited to the above-mentioned embodiments, and any changes, modifications, substitutions, combinations, and simplifications made without departing from the principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A small target detection method, characterized in that: include: S1, obtaining the image to be detected; S2, using the improved YOLOv8 model or the ResNeXt model to process the image to be detected, or using the improved YOLOv8 model and the ResNeXt model in sequence to process the image to be detected, or using the ResNeXt model and the improved YOLOv8 model in sequence to process the image to be detected; wherein, Using the improved YOLOv8 model to process the image to be detected includes: S201, performing multi-scale feature extraction on the image to be detected through a DWRSeg network constructed by a DWR module and a SIR module; S202, regional residualization: generating a plurality of regional feature maps of different sizes through a convolutional layer of a preset size; S203, semantic residualization: applying a multi-rate dilated depth convolution to the regional feature map to perform morphological filtering to capture semantic information of different scales; S204, sampling features from the deformed position on the regional feature map using bilinear interpolation; Using the ResNeXt model to process the image to be detected includes: S211, generating oil bifurcation simulation data; S212, performing feature extraction and bifurcation feature capture on the petroleum bifurcation simulation data and the image to be detected through a ResNeXt network; S3, integrating the detection results of the improved YOLOv8 model and / or the ResNeXt model to generate a detection report.

2. The small target detection method according to claim 1, characterized in that: Before using the improved YOLOv8 model to process the image to be detected, or before using the ResNeXt model to process the image to be detected, the method further includes: Preprocessing the image to be detected, wherein the preprocessing includes: S01, grayscale processing: converting the color image to be detected into a grayscale image; S02, data enhancement: processing the image to be detected by any one or more methods of random cropping, rotation, and scaling; S03, normalization processing: unifying the pixel value range of the image to be detected.

3. The small target detection method according to claim 1, characterized in that: The regional residualization includes: using 3×3 convolution, batch normalization and ReLU activation function to generate regional feature maps of different sizes.

4. The small target detection method according to claim 1, characterized in that: Before using the ResNeXt model to process the image to be detected, it also includes: performing any one or more of image correction, image stitching, contrast enhancement, and noise reduction processing on the image to be detected.

5. The small target detection method according to claim 1, characterized in that: The generating of oil bifurcation simulation data comprises: Fluid simulation: The flow process of liquid on desert terrain is simulated by fluid dynamics simulation tools; the flow process includes branching flow patterns and dynamic simulation; Data Generation: Extract key frames from the fluid simulation and construct a simulated oil spill dataset.

6. The small target detection method according to claim 1, characterized in that: After S04, the method further includes: S205, head detection and training: predicting target information through a preset head network suitable for human and vehicle detection, and performing training optimization according to the target information; And after S212, the method further includes: S213, training and optimizing the results of feature extraction and bifurcation feature capture.

7. The small target detection method according to claim 6, characterized in that: The training and optimization of the results of feature extraction and bifurcation feature capture include: Pre-training and transfer learning: S231, using the preset marine oil spill image dataset, the ResNeXt101 model is initially trained to learn the basic texture, morphology and edge features of oil spills; S232, migrate the initially trained model parameters to the desert oil spill scenario and retain the shared feature representation.

8. The small target detection method according to claim 7, characterized in that: After the pre-training and transfer learning, it also includes: Loss function and optimization strategy: S233, improves the accuracy of bounding box positioning and the ability to distinguish oil spill categories through a preset loss function; S234, optimizing the training results by any one or more of learning rate scheduling, regularization technology, and small batch training.

9. The small target detection method according to claim 8, characterized in that: After the loss function and optimization strategy, it also includes: Data blending and augmentation: S235, combines ocean pre-training, desert simulation data and a small amount of real desert data and allocates different data source ratios; S236, optimizing the training result by any one or more of random cropping, random scaling, adding noise, imaging blur, brightness and contrast adjustment.

10. A small target detection system, characterized in that: Including drone side and server side; among them, The drone end is used to obtain the image to be detected and transmit the image to the server end via a wireless network; The server side is used to process the image to be detected using an improved YOLOv8 model or a ResNeXt model, or to process the image to be detected using the improved YOLOv8 model and the ResNeXt model in sequence, or to process the image to be detected using the ResNeXt model and the improved YOLOv8 model in sequence; wherein; Using the improved YOLOv8 model to process the image to be detected includes: S201, performing multi-scale feature extraction on the image to be detected through a DWRSeg network constructed by a DWR module and a SIR module; S202, regional residualization: generating a plurality of regional feature maps of different sizes through a convolutional layer of a preset size; S203, semantic residualization: applying a multi-rate dilated depth convolution to the regional feature map to perform morphological filtering to capture semantic information of different scales; S204, sampling features from the deformed position on the regional feature map using bilinear interpolation; Using the ResNeXt model to process the image to be detected includes: S211, generating oil bifurcation simulation data; S212, performing feature extraction and bifurcation feature capture on the petroleum bifurcation simulation data and the image to be detected through a ResNeXt network; The improved YOLOv8 model and / or the ResNeXt model detection results are integrated to generate a detection report.

Images