A low-altitude visual angle multi-target detection method based on YOLOv12s

Abstract

The application provides a low-altitude visual angle multi-target detection method based on YOLOv12s, adopts an improved C3k2 feature extraction module to replace a native C3k2 module in a native YOLOv12s backbone network, and extracts multi-scale features of P2 to P5 levels; the improved C3k2 feature extraction module is internally provided with a feature extraction unit of double parallel branches; features of P3 feature layers, P4 feature layers and P5 feature layers output by the backbone network are input into a multi-scale feature depth interaction module, after input features are aligned in terms of channel number and spatial scale, a three-dimensional feature is constructed by adding a scale dimension, spatial and scale joint information interaction of different scale features is realized through three-dimensional convolution, and multi-scale fusion features are output; the multi-scale fusion features are fused with native P3 features output by the backbone network, enhanced P3 features are generated, and the enhanced P3 features are fused with native P2 features output by the backbone network, enhanced P2 features are generated; multi-scale joint prediction is adopted, and target detection results are output.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and deep learning technology, specifically relating to a low-altitude multi-target detection method based on YOLOv12s. Background Technology

[0002] Target detection at low altitudes presents a persistent technical challenge in computer vision due to its unique imaging conditions. In this scenario, sensors are typically positioned at a pitch angle, leading to significant differences in target scale (objects appear smaller in the distance and larger in the foreground), spatial density (objects occluding or overlapping each other), tilted viewpoints (objects with abnormal aspect ratios, such as long, narrow shapes like ships and vehicles), and complex backgrounds (indistinguishable from environmental textures like water, roads, and buildings). These inherent challenges result in general target detection models facing difficulties when directly applied to this scenario, including low recall rates for small and densely packed targets, inaccurate localization of targets with abnormal aspect ratios, and increased false positive rates under complex background interference.

[0003] To address these challenges, existing technical solutions explore several pathways. At the data level, common methods employ general data augmentation techniques such as random cropping, scaling, and color dithering to increase sample diversity. However, these methods often lack specific design for the unique morphology and defects of different target categories from a low-altitude perspective. For example, they fail to specifically enhance the pose diversity of elongated targets or alleviate the blurring and occlusion of small targets, resulting in limited improvement in the model's generalization ability. At the model structure level, a mainstream approach is to build deeper networks or introduce complex attention mechanisms to enhance feature representation capabilities. However, this usually involves a significant increase in the number of model parameters and computational complexity, which is unfavorable for deployment on low-altitude platforms with stringent real-time requirements, such as drones and edge computing devices. Another common approach is to improve multi-scale feature fusion, such as fusing features at different levels through Feature Pyramid Network (FPN) or Path Aggregation Network (PAN) structures. However, many existing fusion operations (such as simple concatenation or element-wise addition) are insufficient in terms of interaction depth, failing to fully realize the deep, bidirectional complementarity between the spatial details of shallow high-resolution features and deep high-level semantic features. This makes the detailed information crucial to small targets easily diluted or lost during transmission. Furthermore, although some studies have attempted to introduce specific designs for small targets, such as adding detection heads to high-resolution feature maps, how to effectively utilize the semantic information of deep networks to guide and enhance these shallow features while avoiding the introduction of excessive noise remains an unresolved issue.

[0004] The YOLO series of models is widely regarded as a benchmark framework for real-time detection scenarios due to its good balance between speed and accuracy. Among them, the YOLOv12 model has made progress in feature extraction and multi-scale fusion through architectural optimization. However, when it is directly applied to the extreme scenario of low-altitude perspective, its inherent limitations still become apparent. The standard convolutional units in its backbone network are not efficient enough in capturing long-range contextual dependencies of long, slanted targets; the feature fusion method of its neck network still lacks sufficient specificity and interaction depth to solve the problem of extremely large scale differences and semantic balance of details in low-altitude scenes; at the same time, the model's sensitivity to extremely small targets is limited by the representation ability of shallow features and the information loss of existing fusion paths, and there is still considerable room for improvement. Summary of the Invention

[0005] To address the shortcomings and deficiencies of existing technologies, this invention provides a low-altitude multi-target detection method based on an improved YOLOv12s. This invention focuses on core issues in low-altitude vision, such as the large target scale span, dense occlusion, tilted viewpoint, and complex background leading to easy attenuation of small target features, insufficient contextual modeling for elongated targets, and loss of shallow detail features during fusion. A systematic optimization scheme is proposed: a dual-path parallel feature extraction structure is embedded in the backbone network. One path aggregates global long-range contextual information in the horizontal and vertical directions through strip pooling to enhance the representation of slender targets such as ships, while the other path preserves local detail features through standard convolution. The outputs of the two paths are fused and output through residual connections, improving feature representation capabilities while maintaining a lightweight design. A multi-scale network is constructed in the neck network. The feature deep interaction mechanism aligns the P3, P4, and P5 level features output from the backbone network with channel and spatial scale, then stacks them along the scale dimension to form a three-dimensional feature tensor. Three-dimensional convolution is used to achieve joint information interaction between the spatial and scale dimensions. The generated fused features are first fused with the original P3 features to enhance the mid-level semantics, and then upsampled and fused with the original P2 features to generate enhanced high-resolution features, effectively mitigating information attenuation in feature transmission for extremely small targets (such as distant pedestrians). In the detection stage, a four-layer detection head is used to jointly predict the enhanced P2, P3, P4, and P5 features, achieving full coverage detection from extremely large to extremely small targets. Simultaneously, considering the differences in morphology and scene characteristics of the four target categories—ships, pedestrians, vehicles, and non-motorized vehicles—customized data augmentation strategies are implemented before model training to specifically expand sample diversity. The entire solution is implemented through modular incremental replacement, without significantly increasing the number of model parameters and computational burden. It significantly improves the accuracy and robustness of multi-target detection in low-altitude perspective scenarios, and is particularly effective in reducing the false negative rate of small targets and elongated targets, providing efficient and reliable detection capabilities for edge deployment scenarios such as UAV inspection and traffic monitoring.

[0006] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0007] A low-altitude multi-target detection method based on YOLOv12s is proposed. The method is implemented using a three-segment architecture of the YOLOv12s model: backbone network, neck network, and detection head. The steps include preprocessing the low-altitude image to be detected, followed by feature extraction, multi-scale feature fusion, and multi-scale target prediction.

[0008] In the feature extraction step, an improved C3k2 feature extraction module is used to replace the original C3k2 module in the original YOLOv12s backbone network to extract multi-scale features from P2 to P5 levels. The improved C3k2 feature extraction module has a built-in feature extraction unit with dual parallel branches. The dual parallel branches include a strip pooling branch for capturing long-range contextual information of long strip-shaped tilted targets from a low-altitude perspective, and a standard convolution branch for extracting local detail features of small targets from a low-altitude perspective. The output features of the two branches are fused and then output.

[0009] In the multi-scale feature fusion step, the features of the P3, P4, and P5 feature layers output by the backbone network are input into the multi-scale feature depth interaction module. After aligning the number of channels and spatial scale of the input features, a new scale dimension is added to construct three-dimensional features. Through three-dimensional convolution, the spatial and scale joint information interaction of features at different scales is realized, and multi-scale fused features are output. The multi-scale fused features are fused with the original P3 features output by the backbone network to generate enhanced P3 features. Then, the enhanced P3 features are fused with the original P2 features output by the backbone network to generate enhanced P2 features.

[0010] In the multi-scale target prediction step, a four-level detection head is used to perform multi-scale joint prediction on the enhanced P2 features, enhanced P3 features, and the features of the P4 and P5 feature layers output by the backbone network, and output the target detection results.

[0011] Furthermore, the improved C3k2 feature extraction module is a C3k2_CDB module. The output features of its two branches are processed by channel concatenation and convolution, and then output after identity mapping residual connection with the original input features of the module. The C3k2_CDB module is embedded into the native YOLOv12s architecture in an incremental replacement manner.

[0012] Furthermore, the C3k2_CDB module is embedded in the P2 feature layer, P3 feature layer, and P5 feature layer of the backbone network, replacing the native C3k2 module at the corresponding positions.

[0013] Furthermore, the feature extraction unit with dual parallel branches is an MPM multi-branch pooling fusion unit; the strip pooling branch includes adaptive average pooling sub-modules in the horizontal and vertical directions, which aggregate global long-range features in the corresponding directions respectively; the standard convolution branch includes at least three parallel convolution units with different receptive fields, which extract local detail features at different scales respectively.

[0014] Furthermore, the multi-scale feature deep interaction module is a ScalSeq module, whose output multi-scale fusion features are reduced in dimensionality by 1×1 convolution and then fused with the original P3 features element by element to generate the enhanced P3 features.

[0015] Furthermore, the generation steps of the enhanced P2 feature are as follows: the enhanced P3 feature is upsampled and amplified by 2 times, then concatenated with the original P2 feature output by the backbone network, and generated after processing by convolution and A2C2f modules.

[0016] Furthermore, the input features of the four-layer detection head are enhanced P2 features, enhanced P3 features, and features from the P4 and P5 feature layers output by the backbone network.

[0017] Furthermore, the improved YOLOv12s model is trained through the following steps: collecting image data of four types of targets—ships, pedestrians, vehicles, and non-motorized vehicles—from a low-altitude perspective, completing annotation and dataset partitioning, performing differentiated data augmentation on the characteristic defects of different target categories, and expanding the training samples; the specific strategies for differentiated data augmentation are as follows: for ship targets, rotation, random scaling, Gaussian blur, and mesh deformation are used for augmentation strategies; for pedestrian and vehicle targets, small-angle rotation, ISO noise simulation, sharpening, and contrast adjustment are used for augmentation strategies; for non-motorized vehicle targets, a comprehensive augmentation strategy adapting to their diverse shapes and dense occlusion features is used. The comprehensive augmentation strategy includes augmentation methods for ship, pedestrian, and vehicle targets, and the range of augmentation parameter values ​​is wider than that for other target categories.

[0018] Furthermore, a low-altitude multi-target detection device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0019] Furthermore, the device is mounted on a drone low-altitude inspection device, a road traffic aerial photography device, or a security monitoring device, and the processor includes at least one of an embedded edge processor and a cloud server processor.

[0020] Compared to existing technologies, this invention and its preferred solution systematically improve the overall performance of multi-target detection from a low-altitude perspective through collaborative improvements from the data source to the model architecture. Specifically, differentiated data augmentation strategies for different target categories effectively simulate the diversity of real-world scenes, fundamentally enhancing the model's feature learning and generalization capabilities for targets with different morphological defects. In the core model structure, the improved C3k2_CDB feature extraction module, through a dual-branch parallel design, simultaneously strengthens the model's ability to capture long-range contextual information and local detail features of elongated, tilted targets, thereby improving the recognition accuracy for such targets. The introduced ScalSeq multi-scale feature deep interaction module utilizes 3D convolution to achieve cross-scale deep feature fusion, promoting efficient complementarity between shallow details and deep semantic information, mitigating feature loss during transmission, which is crucial for improving the recall rate of dense and small target detection. Furthermore, by using deep fusion features to guide the generation of enhanced shallow features and constructing a four-layer detection head, the perception and localization capabilities for extremely small targets in images are significantly enhanced. All of the above improvements are integrated in a lightweight, pluggable, incremental manner, which enables the improved model to significantly improve detection accuracy while maintaining the original model complexity and inference efficiency. This balances detection performance and real-time performance, ensuring the feasibility of deployment on various platforms, from cloud servers to embedded devices. Attached Figure Description

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0022] Figure 1 This is a diagram of the improved YOLOv12s network architecture according to an embodiment of the present invention.

[0023] Figure 2 This is a diagram showing the overall hierarchical structure of the C3k2_CDB module in an embodiment of the present invention.

[0024] Figure 3 This is a structural diagram of the core MPM unit and sub-modules of the C3k2_CDB module in an embodiment of the present invention;

[0025] Figure 4 This is a structural diagram of the ScalSeq multi-scale feature depth interaction module in an embodiment of the present invention;

[0026] Figure 5 The detection effect of the embodiment of the present invention Figure 1 ;

[0027] Figure 6 The detection effect of the embodiment of the present invention Figure 2 ;

[0028] Figure 7 The detection effect of the embodiment of the present invention Figure 3 . Detailed Implementation

[0029] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0030] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] This invention addresses the shortcomings of existing technologies by combining the pain points of low-altitude multi-target detection scenarios with the inherent limitations of the YOLOv12s algorithm in such scenarios, and provides a YOLOv12s-based low-altitude multi-target detection method. Through a collaborative design involving differentiated data augmentation, an improved feature extraction module, multi-scale feature deep fusion, and optimized detection head architecture, this invention significantly improves the model's detection accuracy for small, dense, and tilted targets at low altitudes without significantly increasing the number of model parameters and while fully preserving the real-time detection advantages of YOLOv12s. It effectively reduces false negatives and false positives, while also enhancing the model's resistance to interference from complex backgrounds and its scene generalization ability, providing efficient and reliable algorithmic support for the engineering application of low-altitude multi-target detection.

[0033] The technical solution of this invention consists of two main parts: first, an improved YOLOv12s model construction method for low-altitude multi-target detection; and second, a low-altitude multi-target detection method based on the improved model. The construction of the improved YOLOv12s model is the core of achieving high-precision detection, and its complete network architecture is as follows: Figure 1 As shown, the overall architecture adopts the classic three-part YOLO architecture of backbone network - neck network - head. All improved modules are embedded into the native architecture by incremental replacement or addition, without the need for major modifications to the YOLOv12s core framework.

[0034] 1. Construction of a multi-target dataset from a low-altitude perspective and differential data augmentation

[0035] This step is used to construct a high-quality training dataset adapted to low-altitude viewpoint detection scenarios, addressing the problem that existing enhancement strategies are not designed differently according to target types and have limited enhancement effects. It provides sufficient and diverse sample support for model training. The specific implementation method is as follows:

[0036] Dataset foundation construction: Image data of four types of targets—ships, people, cars, and motor vehicles—were collected from scenarios such as drone inspections and low-altitude aerial photography. After data cleaning, the image resolution was uniformly set to 640×640 pixels. The LabelImg annotation tool was used to annotate the bounding boxes of the targets in the images in YOLO format, generating annotation files containing target category, bounding box center coordinates, and width and height information. The annotated dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio to ensure uniform distribution of target categories and scales within each dataset.

[0037] Category-specific data augmentation: Addressing the core feature deficiencies of different target categories from a low-altitude perspective, we designed and implemented dedicated augmentation strategies to expand the training sample size while specifically enhancing the model's feature learning ability for different targets.

[0038] 1. For ship targets, considering their characteristics of simple attitude, complex water background, and large scale differences, we adopt enhancement strategies such as rotation, random scaling, Gaussian blur and grid deformation to enhance the sample diversity of targets in the variable water environment.

[0039] 2. For pedestrians, vehicles, and non-motorized vehicles, and addressing their characteristics such as blurred features, susceptibility to background interference, dense occlusion, and poor image quality at long distances, an enhancement strategy is adopted, which includes small-angle rotation, ISO noise simulation, sharpening, and contrast adjustment, to specifically solve the blurring problem of small targets at long distances and improve the feature recognition of small targets.

[0040] As a preferred embodiment of the present invention, the above-mentioned differential data augmentation is implemented based on Albumentations and the OpenCV library. Differential augmentation parameters are set for different targets, and the number of training set samples is finally expanded to 24,000, providing high-quality sample support for model training.

[0041] 2. Construction of an improved backbone network integrating Strip Pooling

[0042] This step is used to optimize the feature extraction capabilities of YOLOv12s, addressing the shortcomings of the native model in capturing contextual information of long, small targets from a low-altitude perspective, insufficient feature extraction of small targets, and easy loss of shallow feature transmission. The core improvement is to design and embed the C3k2_CDB module, replacing the C3k2 module in the native YOLOv12s backbone network.

[0043] C3k2_CDB Module Overall Design

[0044] The overall hierarchical structure of the C3k2_CDB module is as follows: Figure 2 As shown, the module is based on the native C3k2 module of YOLOv12s, with C3k_CDB as the basic feature extraction unit and CDBottleneck as the smallest stacking unit. Deep feature extraction is achieved through the concatenated stacking of n CDBottlenecks (the preferred value of n is 1-3). The input features of the module are first subjected to 1×1 convolution for channel dimensionality reduction, and then split along the channel dimension by a Split operation. After splitting, they are input into n parallel C3k_CDB units for feature processing. The processed features are then concatenated by a Concat operation to complete channel concatenation, and then subjected to 1×1 convolution for channel dimensionality increase to restore the number of channels of the input features. Finally, the dimensionality-increased features are added to the original input features by an identity mapping to achieve residual connection, avoid gradient vanishing, and improve the model training stability.

[0045] CDBottleneck and MPM core unit design

[0046] CDBottleneck is the smallest feature extraction unit of the C3k2_CDB module. Its core is the MPM multi-branch pooling fusion module, and its complete structure is as follows: Figure 3 As shown. The MPM module adopts a dual-parallel branch design, incorporating strip pooling into the native C3k2 module of YOLOv12s. To address the potential representation bias issues caused by strip pooling, a parallel standard convolution branch is introduced, enabling the model to adaptively learn the weight balance between the two paths, thus balancing long-range contextual information capture with local detail feature extraction. The specific processing procedure is as follows:

[0047] (1) After the input features are reduced in dimensionality by 1×1 convolution, they are divided into two parallel branches for feature extraction.

[0048] (2) Strip pooling branch: It consists of two sub-modules, X_SPB and Y_SPB, which aggregate global long-distance features through adaptive average pooling in the horizontal and vertical directions, respectively. After convolution and upsampling to restore the feature scale, they are added and fused, and then processed by ReLU activation and 3×3 convolution to output long-distance context features, thus solving the defect of the original model in capturing context information of long strip small targets.

[0049] (3) Standard convolution branch: It consists of three parallel units: 3×3 convolution, 5_PB, and 3_PB (5_PB and 3_PB are the pooling branches of the 5×5 and 3×3 pooling kernels, respectively, and PB is an abbreviation for Pooling Branch). They extract local detail features from different receptive fields. After feature fusion, they are activated by ReLU and processed by 3×3 convolution to output local detail features, thus preserving the detailed information of small targets.

[0050] (4) The dual-branch output features are concatenated by Concat and processed by 1×1 convolution. They are then added to the original input features by residual connection and activated by ReLU, thus completing the single-step feature extraction.

[0051] Module embedding deployment: The above C3k2_CDB module is embedded into the P2, P3, and P5 feature layers of the YOLOv12s backbone network, replacing the original C3k2 module, completing the optimization and reconstruction of the backbone network, and enhancing the backbone network's ability to extract features of various targets from a low-altitude perspective.

[0052] 3. Design of ScalSeq multi-scale feature deep interaction module and fusion process

[0053] This step aims to achieve deep interaction and efficient fusion of features at different scales, addressing the lack of specificity in the original model's multi-scale fusion method and the inefficient combination of shallow detail features and deep semantic features. The core improvement is the design of the ScalSeq multi-scale feature deep interaction module, with the complete structure as follows: Figure 4 As shown.

[0054] ScalSeq module design

[0055] The ScalSeq module is specifically designed to process feature maps from different pyramid levels (P3, P4, P5) of the backbone network. It achieves deep fusion of full-scale contextual information through dynamic sampling to align feature scales, adding a new scale dimension to construct 3D features, and 3D convolution to realize joint spatial-scale interaction. The specific processing logic is as follows:

[0056] (1) Feature preprocessing: The input P3, P4 and P5 feature maps are processed separately. The P4 and P5 feature maps are first uniformized by 1×1 convolution, and then spatial scale alignment is achieved by DySample dynamic sampling (DySample is a well-known dynamic upsampling algorithm in the field of object detection. Its core function is to adaptively complete upsampling according to the content of the input features, so as to realize the spatial size alignment of feature maps of different scales. Compared with the traditional fixed interpolation algorithm, it can better preserve the detailed features of small targets; bilinear interpolation is also used as a simplified implementation), so that the three have the same size and number of channels.

[0057] (2) Dimensional transformation: Perform Unsqueeze operation on the three preprocessed feature maps respectively to add a scale dimension, and convert the two-dimensional feature map in the format of [B, C, H, W] into a three-dimensional feature map in the format of [B, C, S, H, W], providing a basis for information interaction in the scale dimension;

[0058] (3) Scale interaction: The features after dimensional transformation are spliced ​​along the scale dimension to construct 3D features containing spatial and scale dimensions. The 3D features are processed by 1×1×1 3D convolution to realize the deep information interaction between features of different scales. Then, the feature optimization is completed by BN layer, ReLU activation function and MaxPool pooling layer.

[0059] (4) Feature output: Perform a Squeeze operation on the features that have completed the interaction to remove redundant scale dimensions, and finally output the enhanced multi-scale fusion feature P3'.

[0060] Multi-scale feature fusion process

[0061] Based on the ScalSeq module and the newly added Add module ( Figure 1 The module corresponding to WtAdd (i.e., Weighted Add) enables efficient fusion of multi-scale features. The specific process is as follows: extract the P3, P4, and P5 features output from the backbone network, input them into the ScalSeq module to generate multi-scale fused features; after channel dimensionality reduction through 1×1 convolution, the fused features are input into the Add module for weighted addition with the original P3 features to obtain enhanced P3 features, thereby achieving efficient fusion of deep semantic features and shallow detail features and reducing transmission loss of small target features.

[0062] As a further preferred implementation, the Add module in this embodiment adopts a channel-dimensional learnable weighted element-wise addition mechanism to assign trainable channel weights to the multi-scale fusion features and the original P3 features respectively. The weight parameters are optimized synchronously with model training, and adaptively balance the contribution of deep semantic features and shallow detail features without significantly increasing the number of model parameters. As the simplest and most lightweight implementation, the Add module can be directly implemented by equal-weighted element-wise addition without introducing additional training parameters.

[0063] 4. Construction of a multi-scale fusion head architecture including a P2 detection head

[0064] This step is used to optimize the multi-scale prediction capability of the detection head and solve the problem that the original model is insufficient in capturing extremely small targets. The specific implementation method is as follows:

[0065] A head architecture with a P2 detection head and multi-scale fusion is constructed, and the enhanced P3 features generated in the above steps are used to guide the learning of P2 features: the enhanced P3 features are upsampled by 2 times and concatenated with the original P2 features to fuse the high-resolution detail information of the original P2 features and the deep semantic information of the enhanced P3 features; the concatenated features are processed by convolution and the A2C2f module (YOLOv12s' native attention enhancement module) to generate enhanced P2 features.

[0066] Finally, a detection module was introduced to perform multi-scale joint prediction of features at four levels: enhanced P2, enhanced P3, P4, and P5, forming a four-level detection head architecture. This significantly improves the model's ability to capture extremely small targets from a low-altitude perspective and further reduces the false negative rate.

[0067] 5. A Low-Altitude Viewpoint Multi-Target Detection Method Based on an Improved YOLOv12s Model

[0068] The improved YOLOv12s model, built and trained based on the above steps, is then used to perform low-altitude multi-target detection. The specific steps are as follows:

[0069] (1) Acquire the low-altitude view image to be detected, preprocess the image, and unify the image resolution and format;

[0070] (2) Input the preprocessed image into the trained improved YOLOv12s model, and extract P2-P5 multi-scale features through the optimized backbone network;

[0071] (3) The ScalSeq module completes the deep interaction of P3-P5 features to generate enhanced P3 features and enhanced P2 features;

[0072] (4) Multi-scale prediction of features at four levels (P2, P3, P4, P5) is performed using a four-level detection head, and the target category information and bounding box information are output.

[0073] (5) Perform non-maximum suppression processing on the prediction results to remove redundant detection boxes, and finally output the multi-target detection results of the image to be detected.

[0074] The core improvements of the solutions provided by the present invention include:

[0075] To address the characteristic defects of different types of targets from a low-altitude perspective, a unique and differentiated augmentation strategy is designed. Unlike existing general data augmentation solutions, this strategy can specifically improve the sample diversity of ship targets, solve the ambiguity problem of small targets such as pedestrians, vehicles, and non-motorized vehicles, and provide high-quality sample support for model training.

[0076] Based on the native C3k2 module of YOLOv12s, a strip pooling design is incorporated to construct a C3k2_CDB module with dual parallel branches. This module captures both long-range contextual information and local detail features, addressing the inherent shortcomings of the native model in capturing contextual information of long strip-shaped small targets and insufficient extraction of small target features. Furthermore, the module is lightweight and does not significantly increase the number of model parameters.

[0077] The ScalSeq module is designed based on 3D convolution to achieve spatial-scale joint interaction of features at different scales. Unlike existing simple feature splicing and fusion schemes, it can achieve deep fusion of full-scale contextual information and reduce the transmission loss of shallow small target features. At the same time, it guides P2 feature learning with enhanced P3 features and adds a P2 detection head to build a four-level multi-scale prediction architecture, which greatly improves the model's ability to capture extremely small targets.

[0078] Compared to existing technologies, this invention, through the synergistic optimization of three core improvements, fully covers four major pain points in low-altitude perspectives: large differences in target scale, dense occlusion, tilted viewpoint, and complex background, resulting in a significant improvement in detection accuracy.

[0079] All improved modules in this invention adopt a lightweight design, achieving performance improvements only through module replacement and incremental addition, without significantly increasing the number of model parameters or computational load, thus fully preserving the lightweight and real-time detection advantages of YOLOv12s. The improved model has essentially the same number of parameters as the original YOLOv12s, maintaining an inference frame rate of over 30 FPS on NVIDIA RTX 4090 GPUs, roughly on par with the original YOLOv12s; and an inference frame rate of over 15 FPS on edge devices (such as Jetson Nano), fully meeting the engineering requirements for real-time target detection from a low-altitude perspective. Furthermore, the improved model supports multiple lightweight deployment formats such as ONNX and TensorRT, and can be widely adapted to devices with different computing power, including drones, embedded devices, edge servers, and cloud computing servers, enabling full-scenario deployment across edge and cloud environments, demonstrating extremely high engineering application value.

[0080] This invention, through differentiated data augmentation, fully simulates various complex environments under low-altitude perspectives (such as imaging noise, different shooting angles, different shooting distances, and changes in illumination). It enhances the model's resistance to background noise through multi-scale feature deep fusion, avoiding interference from complex backgrounds in target detection. The improved model exhibits excellent detection performance in various low-altitude scenarios, including maritime inspection, road traffic aerial photography, and security monitoring. Its adaptability to different lighting, weather, and background environments is significantly improved, demonstrating strong scene generalization ability. The corresponding detection results are as follows: Figure 5 , Figure 6 , Figure 7 As shown.

[0081] All improved modules in this invention adopt a decoupled design, with modules operating independently, facilitating individual optimization and replacement. As a preferred embodiment of this invention, this solution is based on the Ultralytics open-source framework, requiring no significant modifications to the original core code. Network reconstruction can be completed simply by adding modules, importing and parsing data, and modifying configuration files, facilitating subsequent algorithm iteration and optimization. Furthermore, the improvement strategy of this invention is not limited to the YOLOv12s model and can be flexibly migrated to other versions of YOLOv12 (such as YOLOv12n and YOLOv12m), and can also be adapted to other low-altitude target detection scenarios and target types, possessing strong engineering scalability and portability.

[0082] This invention draws on the multi-scale feature fusion and dense connection techniques in the field of infrared small target segmentation. Combining the task characteristics of low-altitude target detection with the architectural features of YOLOv12s, it has carried out targeted technical integration and customized improvements. It has independently designed core innovative modules such as C3k2_CDB and ScalSeq, realizing effective cross-domain integration of technologies and solving the inherent defects of the native YOLOv12s in low-altitude small target detection scenarios. The algorithm has clear innovation points and technological advancements.

[0083] The technical solution of the present invention will be described in more detail and rigorously below with reference to specific embodiments, accompanying drawings and experimental data. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0084] The implementation environment of this embodiment is as follows: the software environment is built based on the open-source frameworks Python 3.8, PyTorch 1.13, and Ultralytics 8.0; the hardware environment consists of a Linux system, a 2.50GHz CPU, 11GB RAM, and an NVIDIA RTX 4090 GPU; the detection targets are four categories of targets from a low-altitude perspective: ships, people, cars, and motor vehicles. This embodiment is a complete end-to-end implementation process, forming a closed loop from dataset construction, model improvement and reconstruction, model training to performance verification, fully replicating the technical solution and effects of this invention.

[0085] 1. Construction of a multi-target dataset from a low-altitude perspective and implementation of differential data augmentation

[0086] This step provides high-quality samples to support model training. The complete implementation process is as follows:

[0087] 1.1 Construction of the basic dataset

[0088] A total of 10,000 images were collected from three typical low-altitude scenarios: maritime inspection by drones, aerial photography of urban road traffic, and low-altitude security monitoring. Among them, there were 2,137 images of ships, 1,852 images of people, 3,049 images of cars, and 2,962 images of motors. The resolution of all images was unified to 640×640 pixels, and invalid data was cleaned and formatted.

[0089] The LabelImg annotation tool was used to annotate the bounding boxes of targets in the images. The annotation format adopted the YOLO standard format, generating annotation files containing target category, bounding box center coordinates, and bounding box width and height information. The annotated dataset was randomly divided into a training set of 8000 images, a validation set of 1000 images, and a test set of 1000 images in an 8:1:1 ratio. The partitioning process ensured that the target category and target scale distribution were uniform within each dataset, avoiding model overfitting due to uneven distribution.

[0090] 1.2 Implementation of Category Differentiation Data Augmentation Strategy

[0091] To address the core characteristic defects of four types of low-altitude targets, a differentiated data augmentation strategy was designed and implemented based on Albumentations and the OpenCV library. The augmentation methods and parameter settings for each type of target are shown in Table 1 below:

[0092] Table 1. Differentiated Data Augmentation Methods and Parameter Settings

[0093]

[0094] The enhancement strategy calls the corresponding enhancement function based on the core target category of a single image, and performs enhancement operations on the image and the bounding box simultaneously to ensure that the enhanced bounding box and the target position are completely matched. Through the above strategy, the number of training set samples was expanded to 24,000, the sample diversity of various targets was significantly improved, and scene features such as target tilt, dense occlusion, and scale difference under low-altitude perspective were fully simulated, and the feature recognition of small targets was significantly enhanced.

[0095] 2 Implementation and Embedding of the C3k2_CDB Improved Feature Extraction Module

[0096] This step involves a core improvement to the YOLOv12s backbone network, with the overall module structure corresponding to... Figure 2 The core unit structure corresponds to Figure 3 The complete implementation process is as follows:

[0097] 2.1 Core Design and Implementation of C3k2_CDB Module

[0098] The C3k2_CDB module is based on the native C3k2 module of YOLOv12s, and integrates the strip pooling mechanism with identity mapping residual connections. The implementation process is as follows:

[0099] 1. Input features are first subjected to 1×1 convolution to reduce channel dimensionality, thereby reducing the computational load of the module;

[0100] 2. The dimensionality-reduced features are split into two parallel branches along the channel dimension, which perform long-range context feature extraction and local detail feature extraction respectively:

[0101] Strip pooling branch: The Strip Pooling layer is used to aggregate the global long-distance context information of the feature map through adaptive average pooling in both horizontal and vertical directions. After convolution and upsampling to restore the feature scale, the fusion is completed, which solves the defect of the original model in capturing the context information of long strip small targets.

[0102] Standard convolution branch: 3×3 standard convolution is used to extract local detail features of the target, fully preserve the detailed information of small targets, and make up for the local feature loss caused by strip pooling;

[0103] 3. After the output features of the two branches are concatenated, they are subjected to 1×1 convolution to increase the channel dimensionality and restore the number of channels of the input features;

[0104] 4. The upgraded features are added to the original input features using an identity mapping to form residual connections, which avoids the gradient vanishing problem during training and improves the stability of model training.

[0105] The main implementation code for the above modules is shown below:

[0106] Python

[0107] import torch

[0108] import torch.nn as nn

[0109] import torch.nn.functional as F

[0110] # Strip pooling core layer

[0111] class StripPooling(nn.Module):

[0112] def __init__(self, in_channels, out_channels):

[0113] super().__init__()

[0114] # Adaptive average pooling in horizontal / vertical directions

[0115] self.horizontal_pool = nn.AdaptiveAvgPool2d((None, 1))

[0116] self.vertical_pool = nn.AdaptiveAvgPool2d((1, None))

[0117] # Channel Transformation Convolution and Activation Functions

[0118] self.conv1 = nn.Conv2d(in_channels, out_channels, 1, 1, 0)

[0119] self.conv2 = nn.Conv2d(out_channels*2, out_channels, 1, 1, 0)

[0120] self.relu = nn.ReLU(inplace=True)

[0121] def forward(self, x):

[0122] # Horizontal Feature Extraction and Upsampling Restoration

[0123] x_h = self.horizontal_pool(x)

[0124] x_h = self.conv1(x_h)

[0125] x_h = F.interpolate(x_h, size=x.shape[2:], mode="bilinear",align_corners=False)

[0126] # Vertical feature extraction and upsampling restoration

[0127] x_v = self.vertical_pool(x)

[0128] x_v = self.conv1(x_v)

[0129] x_v = F.interpolate(x_v, size=x.shape[2:], mode="bilinear",align_corners=False)

[0130] # Bidirectional Feature Fusion

[0131] return self.relu(self.conv2(torch.cat([x_h, x_v], dim=1)))

[0132] # C3k2_CDB complete module

[0133] class C3k2_CDB(nn.Module):

[0134] def __init__(self, c1, c2, n=1, shortcut=True, e=0.5):

[0135] super().__init__()

[0136] c_ = int(c2 * e) # Channel scaling factor

[0137] self.cv1 = nn.Conv2d(c1, c_*2, 1, 1, 0)

[0138] self.cv2 = nn.Conv2d(c_*2, c2, 1, 1, 0)

[0139] self.strip_pool_branch = StripPooling(c_, c_)

[0140] self.conv_branch = nn.Conv2d(c_, c_, 3, 1, 1)

[0141] self.relu = nn.ReLU(inplace=True)

[0142] self.shortcut = shortcut and c1 == c2

[0143] def forward(self, x):

[0144] residual = x if self.shortcut else 0

[0145] x = self.cv1(x)

[0146] # Dual Parallel Branch Feature Extraction

[0147] x1, x2 = torch.chunk(x, 2, dim=1)

[0148] x1 = self.strip_pool_branch(x1)

[0149] x2 = self.relu(self.conv_branch(x2))

[0150] # Feature fusion and residual connection

[0151] x_out = self.cv2(torch.cat([x1, x2], dim=1))

[0152] return x_out + residual

[0153] 2.2 Module Embedded Deployment

[0154] The C3k2_CDB module was embedded into the YOLOv12s backbone network, replacing the native C3k2 module. Specifically, the embedding locations were the 2nd layer (P2 feature layer), the 4th layer (P3 feature layer), and the 20th layer (P5 feature layer) of the backbone network. After the embedding was completed, the backbone network's ability to extract long-range contextual information and local detail features of targets from a low-altitude perspective was significantly improved.

[0155] 3. Implementation of ScalSeq Multi-Scale Feature Fusion Module and Fusion Process

[0156] This step represents a core improvement to the YOLOv12s neck network, with the module's complete structure and processing flow corresponding to this. Figure 4 The complete implementation process is as follows:

[0157] 3.1 Core Design and Implementation of the ScalSeq Module

[0158] The ScalSeq module is specifically designed to process the feature maps at three different levels (P3, P4, and P5) output by the backbone network. Its core functionality involves using 3D convolution to achieve joint information interaction between spatial and scale dimensions, fusing full-scale contextual information. This addresses the issue of low efficiency in fusing shallow detail features and deep semantic features in the native model. The implementation process is shown below:

[0159] 1. Channel and scale unification: Perform 1×1 convolution on the input P3, P4, and P5 feature maps respectively to unify the number of channels in each feature map; upsample the P4 and P5 feature maps to the spatial scale of the P3 feature to complete scale alignment;

[0160] 2. Dimensional Expansion: A scale dimension is added to each aligned feature map, converting the two-dimensional feature map into a three-dimensional feature format, providing a foundation for information exchange across the scale dimension;

[0161] 3. Scale-dimensional interaction: Multiple sets of features are concatenated along the scale dimension to construct 3D features that include both spatial and scale dimensions. Cross-scale information interaction is performed through 3D convolution, and full-scale contextual information is fused.

[0162] 4. Feature Output: After the interactive features are flattened, convolutional and activation functions are applied, the final multi-scale fused features are output.

[0163] The main implementation code for the above modules is shown below:

[0164] Python

[0165] class ScalSeq(nn.Module):

[0166] def __init__(self, c1_list, c2, kernel_size=3, stride=1, padding=1):

[0167] super().__init__()

[0168] # Input Channel List and Output Channel Settings

[0169] self.c1_list = c1_list

[0170] self.c2 = c2

[0171] # Channel Unified Convolutional Layer

[0172] self.channel_conv = nn.ModuleList([nn.Conv2d(c1, c2 / / 3, 1, 1,0) for c1 in c1_list])

[0173] # 3D Convolutional Scale Interaction Layer

[0174] self.conv3d = nn.Conv3d(c2 / / 3, c2 / / 3, kernel_size, stride,padding)

[0175] # Output fused convolutional layer

[0176] self.out_conv = nn.Conv2d(c2, c2, 1, 1, 0)

[0177] self.relu = nn.ReLU(inplace=True)

[0178] def forward(self, x_list):

[0179] # Input format for x_list: [P3 features, P4 features, P5 features]

[0180] feat_list = []

[0181] # Channel unification and scale alignment

[0182] for x, conv in zip(x_list, self.channel_conv):

[0183] feat = conv(x)

[0184] feat = F.interpolate(feat, size=x_list[0].shape[2:], mode="bilinear", align_corners=False)

[0185] feat_list.append(feat.unsqueeze(2)) # Add scale dimension

[0186] # Scale-Dimensional Stitching and 3D Convolution Interaction

[0187] feat_3d = torch.cat(feat_list, dim=2)

[0188] feat_3d = self.conv3d(feat_3d)

[0189] # Dimensional Flattening and Feature Output

[0190] feat_flat = feat_3d.view(feat_3d.shape[0], -1, feat_3d.shape[3], feat_3d.shape[4])

[0191] return self.relu(self.out_conv(feat_flat))

[0192] # Element-by-element addition fusion module

[0193] class Add(nn.Module):

[0194] def __init__(self):

[0195] super().__init__()

[0196] def forward(self, x1, x2):

[0197] # x1 represents the original P3 feature, and x2 represents the dimensionality-reduced fusion feature output by ScalSeq.

[0198] return x1 + x2

[0199] 3.2 Implementation of Multi-Scale Feature Fusion and P2 Feature Enhancement Process

[0200] Based on the ScalSeq and Add modules, the complete implementation process of multi-scale feature fusion and P2 feature enhancement is as follows:

[0201] 1. Extract the P3 (layer 4), P4 (layer 6), and P5 (layer 8) features from the YOLOv12s backbone network output, input them into the ScalSeq module, and generate 256-channel multi-scale fusion features;

[0202] 2. The fused features are reduced to 128 channels by 1×1 convolution, and then added element-wise with the original P3 features (128 channels) to the Add module to obtain enhanced P3 features, thus achieving efficient fusion of deep semantic features and shallow detail features;

[0203] 3. The enhanced P3 feature is upsampled by 2 times and then concatenated with the original P2 feature (128 channels) to obtain a concatenated feature with 256 channels;

[0204] 4. The spliced ​​features are processed by 3×3 convolution and A2C2f module to obtain 128-channel enhanced P2 features, thus completing the enhancement and optimization of shallow high-resolution features;

[0205] 5. Input the enhanced P2 feature, enhanced P3 feature, P4 feature and P5 feature into the detection head to construct a four-level multi-scale joint detection architecture.

[0206] 4. Improve the overall architecture reconstruction and training implementation of the YOLOv12s model

[0207] 4.1 Implementation of Overall Model Architecture Restructuring

[0208] The architecture of the improved YOLOv12s model was restructured based on the Ultralytics open-source framework. The overall architecture corresponds to... Figure 1 The core implementation steps are as follows:

[0209] 1. Create a custom module directory under the nn directory of the ultralytics framework, save the implementation code of the C3k2_CDB, ScalSeq, and Add modules into the corresponding files, and complete the package interface definition and framework import configuration.

[0210] 2. Add parsing logic for a custom module to the model parsing function of the Ultralytics framework, so that the framework can recognize and load the improved module of this invention;

[0211] 3. Write the YAML configuration file for the improved model. The configuration file should specify the replacement position of the C3k2_CDB module in the backbone network, the new paths of the ScalSeq and Add modules in the neck network, the P2 feature enhancement process, and the output settings of the four-layer detection head, thus fully defining the network architecture of the improved model.

[0212] 4.2 Model Training Parameter Settings

[0213] The enhanced low-altitude view dataset from Section 1 was used for model training, and the core training parameters are set as shown in Table 2 below:

[0214] Table 2. Core parameter settings for model training

[0215]

[0216] 4.3 Model Training Execution and Convergence

[0217] The training function based on the Ultralytics framework initiates model training, and validation and weight saving are performed simultaneously during training. During training, the model enters a convergence plateau after about 60 epochs. Finally, the validation set mAP@0.5 stabilizes at 0.773-0.774, the training loss continues to decrease steadily, the validation loss shows no obvious oscillations, and the model does not exhibit overfitting.

[0218] 5. Performance testing and verification of the improved model

[0219] The improved model was tested using a test set. mAP@0.5 (mean precision), a common metric in object detection, was used as the core evaluation indicator, while Precision and Recall were used as auxiliary evaluation indicators.

[0220] The core performance comparison between the improved model of this invention and the original YOLOv12s model after testing is shown in Table 3 below:

[0221] Table 3. Performance Comparison of Improved Model and Original Model at mAP@0.5

[0222]

[0223] As shown in the table above, the overall detection accuracy of the improved model in this invention has increased from 0.735 in the original YOLOv12s to 0.774, with improvements of 6.7% and 5.5% in the detection accuracy of small targets such as people and motors, respectively. Experimental verification shows that the improved model reduces the false negative rate by more than 12% and the false positive rate by more than 8% in low-altitude scenarios, and significantly enhances the recognition capabilities for tilted and partially occluded targets. The inference frame rate on an NVIDIA RTX 4090 GPU remains above 30 FPS, and on a Jetson Nano edge device, it reaches above 15 FPS, essentially matching the inference speed of the original YOLOv12s model, achieving a balance between detection accuracy and real-time performance. The model demonstrates stable and excellent detection performance in three low-altitude scenarios: maritime inspection, road traffic aerial photography, and security monitoring, exhibiting strong scene generalization ability and adaptability to complex environments. The corresponding detection results are as follows: Figure 5 , Figure 6 , Figure 7 As shown.

[0224] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0225] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive other forms of low-altitude multi-target detection methods based on YOLOv12s. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A low-altitude multi-target detection method based on YOLOv12s, the method being implemented based on a three-segment architecture of the YOLOv12s model: backbone network, neck network, and detection head, comprising the steps of preprocessing the low-altitude image to be detected, performing feature extraction, multi-scale feature fusion, and multi-scale target prediction, characterized in that: In the feature extraction step, an improved C3k2 feature extraction module is used to replace the original C3k2 module in the original YOLOv12s backbone network to extract multi-scale features from P2 to P5 levels. The improved C3k2 feature extraction module has a built-in feature extraction unit with dual parallel branches. The dual parallel branches include a strip pooling branch for capturing long-range contextual information of long strip-shaped tilted targets from a low-altitude perspective, and a standard convolution branch for extracting local detail features of small targets from a low-altitude perspective. The output features of the two branches are fused and then output. In the multi-scale feature fusion step, the features of the P3, P4, and P5 feature layers output by the backbone network are input into the multi-scale feature depth interaction module. After aligning the number of channels and spatial scale of the input features, a new scale dimension is added to construct three-dimensional features. Through three-dimensional convolution, the spatial and scale joint information interaction of features at different scales is realized, and multi-scale fused features are output. The multi-scale fused features are fused with the original P3 features output by the backbone network to generate enhanced P3 features. Then, the enhanced P3 features are fused with the original P2 features output by the backbone network to generate enhanced P2 features. In the multi-scale target prediction step, a four-level detection head is used to perform multi-scale joint prediction on the enhanced P2 features, enhanced P3 features, and the features of the P4 and P5 feature layers output by the backbone network, and output the target detection results.

2. The low-altitude multi-target detection method based on YOLOv12s according to claim 1, characterized in that: The improved C3k2 feature extraction module is the C3k2_CDB module. The output features of its two branches are processed by channel concatenation and convolution, and then connected with the original input features of the module by identity mapping residual connection before output. The C3k2_CDB module is embedded into the native YOLOv12s architecture in an incremental replacement manner.

3. The low-altitude multi-target detection method based on YOLOv12s according to claim 2, characterized in that: The C3k2_CDB module is embedded in the P2, P3, and P5 feature layers of the backbone network, replacing the native C3k2 module at the corresponding positions.

4. The low-altitude multi-target detection method based on YOLOv12s according to claim 1, characterized in that: The feature extraction unit with dual parallel branches is an MPM multi-branch pooling fusion unit; the strip pooling branch includes adaptive average pooling sub-modules in the horizontal and vertical directions, which aggregate global long-range features in the corresponding directions respectively; the standard convolution branch includes at least three parallel convolution units with different receptive fields, which extract local detail features at different scales respectively.

5. The low-altitude multi-target detection method based on YOLOv12s according to claim 1, characterized in that: The multi-scale feature deep interaction module is the ScalSeq module. Its output multi-scale fusion features are reduced in dimensionality by 1×1 convolution and then fused with the original P3 features element by element to generate the enhanced P3 features.

6. The low-altitude multi-target detection method based on YOLOv12s according to claim 1, characterized in that: The steps for generating the enhanced P2 feature are as follows: the enhanced P3 feature is upsampled and amplified by 2 times, then concatenated with the original P2 feature output by the backbone network, and generated after processing by convolution and the A2C2f module.

7. The low-altitude multi-target detection method based on YOLOv12s according to claim 1, characterized in that: The input features of the four-layer detection head are enhanced P2 features, enhanced P3 features, and features from the P4 and P5 feature layers output by the backbone network.

8. The low-altitude multi-target detection method based on YOLOv12s according to claim 1, characterized in that: The improved YOLOv12s model is trained through the following steps: collecting image data of four types of targets—ships, pedestrians, vehicles, and non-motorized vehicles—from a low-altitude perspective, completing annotation and dataset partitioning, performing differentiated data augmentation on the characteristic defects of different target categories, and expanding the training samples; the specific strategies for differentiated data augmentation are as follows: for ship targets, rotation, random scaling, Gaussian blur, and mesh deformation enhancement strategies are adopted; for pedestrian and vehicle targets, small-angle rotation, ISO noise simulation, sharpening, and contrast adjustment enhancement strategies are adopted. For non-motorized vehicle targets, a comprehensive enhancement strategy is adopted that adapts to their diverse shapes and dense occlusion features. This comprehensive enhancement strategy includes enhancement methods for ship, pedestrian and vehicle targets, and the range of enhancement parameter values ​​is wider than that for other target categories.

9. A low-altitude, multi-target detection device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 7.

10. A low-altitude multi-target detection device according to claim 9, characterized in that, The device is mounted on a drone low-altitude inspection device, a road traffic aerial photography device, or a security monitoring device, and the processor includes at least one of an embedded edge processor and a cloud server processor.

Images