Lightweight target detection architecture based on multi-scale feature fusion, detection method and application

Abstract

The invention relates to a multi-scale feature fusion lightweight target detection architecture, and relates to the technical field of vision and artificial intelligence application, the multi-scale feature fusion lightweight target detection architecture comprises a backbone network and a neck network connected with the backbone network, the backbone network receives an image and outputs a multi-scale feature map, comprising a plurality of DualConv modules which are sequentially arranged along a data transmission path, a plurality of C2F modules which are arranged in one-to-one correspondence with the plurality of DualConv modules, and an SPPF module which is connected with the C2F module at the tail end of a backbone network, the neck network is connected with the backbone network, fuses the multi-scale feature map and outputs a detection result, and the neck network comprises a Slim-Neck module which receives and processes the multi-scale feature map and comprises a GSConv module, a feature fusion module and a GSBottleneck module, and the feature fusion module and the GSBottleneck module are both integrated with the GSConv module; and the detection head comprises P2-P5 four-stage detection layers which are connected stage by stage, and the resolutions of the P2 detection layer, the P3 detection layer, the P4 detection layer and the P5 detection layer are gradually reduced stage by stage.

Description

Technical Field

[0001] This application relates to the fields of vision and artificial intelligence application technology, and in particular to a lightweight object detection architecture, detection method and application of multi-scale feature fusion. Background Technology

[0002] Surface target detection utilizes computer vision, image processing, and deep learning techniques to identify objects such as ships, buoys, and obstacles from images or videos captured on the water surface. This technology is crucial for navigation safety, especially in autonomous navigation systems, enabling vessels to detect potential collision risks and take appropriate evasive action. However, in complex inland environments, factors such as weather changes, light reflection, waves, and vegetation obstruction pose significant challenges to target detection. Furthermore, the complex distribution, small size, and blurred texture of objects on the water surface make the detection task difficult.

[0003] Object detection is one of the core tasks of computer vision, widely used in fields such as autonomous driving and security monitoring. In recent years, deep learning-based object detection methods using convolutional neural networks (CNNs) have shown significant advantages and have become a mainstream research direction. Contemporary deep learning detection frameworks are divided into two-stage and one-stage architectures. Two-stage detectors, such as the R-CNN series, generate candidate bounding boxes through region proposal networks, and then perform classification and regression, achieving high detection accuracy, but with high computational complexity and insufficient real-time performance. In contrast, one-stage detectors such as YOLO and SSD use an end-to-end architecture, performing object localization and classification simultaneously, making them more suitable for real-time surface applications, such as ship monitoring and unmanned surface vehicle navigation, where low latency is crucial.

[0004] Recent research has advanced water surface target detection through scene-specific optimizations, such as DENS-YOLOv6 and improved YOLOv8 for small debris detection, while dual-stream feature fusion optimization addresses environmental challenges in low-light conditions, including haze and multi-exposure scenarios. Significant progress has been made in optimizing YOLO-based water surface target detection algorithms in recent years. To address challenges such as large-scale target variations and complex background interference in water surface environments, researchers primarily focus on two aspects: network architecture optimization and innovative feature fusion mechanisms.

[0005] For example, embedding the SE-PPM module into YOLOv3 constructs a spatial pyramid pooling structure, enhancing context awareness through multi-directional feature pyramids and introducing FocalLoss to alleviate class imbalance, resulting in a significant improvement in floating debris detection accuracy. Another example is MA-YOLOv4, which prioritizes lightweight models. It reconstructs the backbone network using MobileNetv3. This architecture optimizes anchor box initialization through maximum distance K-means clustering and incorporates an Adaptive Spatial Feature Fusion (ASFF) module for efficient multi-scale feature aggregation, demonstrating a good speed-accuracy tradeoff in ship detection tasks.

[0006] However, despite the aforementioned progress, existing methods still face three main challenges:

[0007] (1) Lightweight design usually sacrifices feature representation capability, which complicates the balance between model efficiency and detection accuracy;

[0008] (2) The multi-scale feature fusion mechanism failed to make full use of shallow detail information, which limited the performance of small target detection;

[0009] (3) Robust optimization strategies for dynamic water surface characteristics such as wave interference and target deformation are still immature. Summary of the Invention

[0010] This application addresses these limitations by constructing a surface target detection framework that integrates lightweight architecture, multi-scale perception, and dynamic adaptability.

[0011] This application proposes a novel object detection architecture (DSH-YOLO) and detection method, achieving efficient detection through three-level structural optimization. At the backbone network level, the main feature extractor is reconstructed by introducing the DualConv lightweight convolutional module. This module utilizes the synergistic effect of group convolution and heterogeneous convolution to effectively maintain feature representation capabilities while significantly reducing computational complexity.

[0012] To address the challenges of multi-scale feature fusion, this application designs the SlimNeck module, which combines with the GSConv module. This module employs depthwise separable convolutions to minimize the number of parameters, while utilizing channel shuffling operations to enhance information flow, thereby improving feature fusion efficiency.

[0013] For small object detection, this application proposes a novel P2 detection head in the prediction layer. By enhancing the response to high-resolution shallow features, the detection architecture achieves superior accuracy in small object detection. Through these lightweight architectural designs and multi-scale feature enhancement techniques, this application successfully achieves a synergistic optimization of accuracy and efficiency.

[0014] Furthermore, the WSOD-12K dataset, a high-quality ship dataset specifically designed for efficient maritime target detection, was constructed based on this foundation. This dataset provides a valuable benchmark for evaluating detection algorithms in the marine environment.

[0015] Firstly, the lightweight target detection architecture based on multi-scale feature fusion provided in this application adopts the following technical solution:

[0016] A lightweight object detection architecture based on multi-scale feature fusion includes:

[0017] The backbone network receives images and outputs multi-scale feature maps, including:

[0018] Multiple DualConv modules are set sequentially along the data transmission path;

[0019] Multiple C2F modules are configured one-to-one with the multiple DualConv modules, and are connected to adjacent modules.

[0020] The DualConv module;

[0021] The SPPF module is connected to the C2F module at the end of the backbone network;

[0022] A neck network, connected to the backbone network, fuses multi-scale feature maps and outputs detection results. The neck network includes:

[0023] The Slim-Neck module receives and processes multi-scale feature maps, including the GSConv module, a feature fusion module that integrates the GSConv module, and the GSBottleneck module.

[0024] The detection head includes four detection layers, P2-P5, connected in a progressive manner, wherein the resolution of the P2, P3, P4 and P5 detection layers decreases progressively.

[0025] Furthermore, the DualConv module includes:

[0026] The channel grouping unit divides the input feature map into two groups along the channel dimension;

[0027] A parallel convolutional unit, connected to the output of the channel grouping unit, includes a 3×3 convolutional kernel and a 1×1 convolutional kernel arranged in parallel;

[0028] A feature recombination unit is connected to the output of the parallel convolution unit to shuffle the two sets of feature maps.

[0029] Furthermore, the GSBottleneck module includes:

[0030] Two GSConv modules are set up sequentially along the data transmission path to process feature maps of different scales.

[0031] The standard convolutional layer, running in parallel with the GSConv module at the output of the Slim-Neck architecture, is used to preserve the original feature information;

[0032] The additive fusion unit receives and fuses the feature maps of the GSConv module and the standard convolutional layer.

[0033] Furthermore, the GSConv module includes:

[0034] A standard convolutional layer receives the input feature map;

[0035] A depth-separable convolutional layer is connected to the output of the standard convolutional layer.

[0036] The channel shuffling unit, located between the two sets of C2 / 2 input channels, rearranges the feature map according to the channel dimensions.

[0037] Furthermore, the feature fusion module is configured with multiple modules connected in a multi-level series, and the feature fusion module performs the following steps:

[0038] Step M1: Obtain the first feature map and the second feature map, wherein the resolution of the first feature map is lower than that of the second feature map;

[0039] Step M2: Upsample the first feature map to match its resolution with that of the second feature map;

[0040] Step M3: Concatenate the upsampled first feature map and the second feature map along the channel dimension to obtain the fused feature map.

[0041] Furthermore, it also includes a data storage module that stores a WSOD-12K dataset that provides visual samples. The objects in the WSOD-12K dataset include ships, obstacles, rocks, platforms, and people.

[0042] Secondly, this application provides a target detection method that adopts the above-mentioned target detection architecture. The target detection method includes the following steps:

[0043] Step S1: Input the image to be detected into the backbone network and extract the initial feature map;

[0044] Step S2: Input the initial feature map into the neck network and perform multi-scale feature fusion on the initial feature map;

[0045] Step S3: Input the fused feature map into the detection head and output the target detection result.

[0046] Thirdly, this application provides an application of the above-mentioned target detection architecture in surface navigation safety monitoring.

[0047] In summary, this application includes at least one of the following beneficial technical effects:

[0048] 1. This application significantly reduces computational cost and parameter count through channel grouping and parallel convolution design in the DualConv module. Compared to traditional convolution, the DualConv module divides channels into two groups and performs 3×3 and 1×1 convolutions respectively, reducing the number of parameters by 40% while maintaining efficient feature extraction capabilities. This lightweight design enables the model to run efficiently on resource-constrained devices (such as mobile devices), reducing hardware requirements and improving the model's deployment flexibility.

[0049] 2. The Slim-Neck module of this application employs a multi-level feature fusion strategy, progressively fusing feature maps of different scales to further optimize the representation of multi-scale features. This multi-scale feature fusion mechanism significantly improves the model's ability to detect small, medium, and large targets, enhancing the model's robustness and adaptability.

[0050] 3. The GSConv module of this application achieves efficient feature representation by combining standard convolution and depthwise separable convolution and introducing a channel shuffling unit. Depthwise separable convolution significantly reduces computation by separating spatial convolution and channel convolution, while the channel shuffling unit enhances cross-channel feature interaction by rearranging channels. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall target detection architecture in this application;

[0052] Figure 2 This is a schematic diagram of the DualConv module in this application;

[0053] Figure 3 This is a schematic diagram of the GSConv module in this application;

[0054] Figure 4 This is a schematic diagram of the GSBottleneck module in this application;

[0055] Figure 5 This is a schematic diagram of the detection head in this application;

[0056] Figure 6 These are representative visual sample images of four typical datasets;

[0057] Figure 7 This is a partial data plot of the WSOD-12K dataset;

[0058] Figure 8 This is a comparison chart of target detection results from different methods;

[0059] Figure 9 This is a visual comparison of the detection results of this application with YOLO11 and YOLO12 on the WSOD-12K dataset;

[0060] Figure 10 This is a visual comparison of the detection results of this application with YOLO11 and YOLO12 on the WSODD dataset. Detailed Implementation

[0061] The following will be combined with the appendix Figure 1-10 The technical solution of this application is clearly and completely described. The following embodiments are exemplary and are only used to explain this application, and should not be construed as limiting this application. In the following description, the same reference numerals are used to denote the same or equivalent elements, and repeated descriptions are omitted.

[0062] In the description of this application, it should be understood that the terms "upper", "lower", "inner", "outer", "left", "right", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this application is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this application and to simplify the description, and are not intended to indicate or imply that the equipment or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0063] Furthermore, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0064] It should also be further understood that the term "and / or" as used in this application specification and the corresponding claims refers to any combination of one or more of the listed items and all possible combinations.

[0065] A lightweight object detection architecture based on multi-scale feature fusion, referring to Figure 1 It includes a backbone network, a neck network, and a detection head. The backbone network receives images and outputs multi-scale feature maps, the neck network receives multi-scale feature maps and performs multi-scale feature fusion, and the detection head outputs target detection results based on the fused feature maps.

[0066] Specifically, the backbone network includes multiple DualConv modules, multiple C2F modules, and an SPPF module. The multiple DualConv modules are set up sequentially along the data transmission path, and the multiple C2F modules are set up one-to-one with the multiple DualConv modules, connecting adjacent DualConv modules. The SPPF module is set at the end of the backbone network and connected to the C2F module located at the end of the backbone network.

[0067] The number of DualConv modules and C2F modules can be set according to the actual situation, such as 2, 3, 4, etc. In this embodiment, the number of DualConv modules and C2F modules is set to 4.

[0068] In this case, for ease of explanation, the DualConv module and the C2F module are numbered, and the corresponding references are provided. Figure 2 The connection logic of the backbone network is as follows: Input image → DualConv1 → C2F1 → DualConv2 → C2F2 → DualConv3 → C2F3 → DualConv4 → C2F4 → SPPF → Output multi-scale feature map.

[0069] In one specific embodiment, the backbone network operates as follows:

[0070] DualConv1 receives the original image and outputs a feature map. C2F1 receives the output of DualConv1 and passes it to DualConv2. After processing by DualConv2, it is passed to the next unit (i.e., DualConv3) via C2F2, and so on until it is transmitted to SPPF. SPPF performs pooling operation on the output of C2F4 and outputs the final multi-scale feature map.

[0071] The DualConv module is used to construct a lightweight feature extraction network.

[0072] Reference Figure 2 The DualConv module includes a channel grouping unit, a parallel convolutional unit, and a feature reorganization unit. The channel grouping unit divides the input feature map into two groups along the channel dimension. The parallel convolutional unit is connected to the output of the channel grouping unit and includes parallel 3×3 convolutional kernels and 1×1 convolutional kernels. The feature reorganization unit is connected to the output of the parallel convolutional unit to shuffle the two groups of feature maps.

[0073] The DualConv module cleverly combines the advantages of group convolutions (GroupConv) and heterogeneous convolutions (HetConv). Its unique design employs parallel 3×3 and 1×1 convolutional kernels, enabling it to simultaneously capture spatial features with large receptive fields and the computational efficiency of small kernels in a single forward pass. DualConv's parallel integration of group and heterogeneous convolutions provides richer feature diversity than depthwise separable convolutions and higher computational efficiency than standard convolutions, balancing feature richness and efficiency in complex tasks to improve performance.

[0074] In this implementation, the input feature channels are first grouped, and then convolutions of different kernel sizes are performed in parallel within each group. Finally, cross-group information exchange is achieved through channel concatenation and recombination. This approach ensures that feature extraction capabilities are maintained while significantly reducing network parameters and computational complexity.

[0075] Furthermore, the C2F module connects adjacent DualConv modules, transferring features and enhancing information flow. By performing a 3×3 convolution (stride = 1) on the input feature map, the output is passed to the next DualConv module while maintaining the same resolution, thus avoiding gradient vanishing and strengthening the connection between shallow and deep features.

[0076] Meanwhile, in object detection tasks, while adding more feature maps can enhance feature representation, it inevitably increases the depth of the neck component, leading to a significant increase in network parameters and computational overhead. To effectively address this challenge, this application proposes an integrated GSConv (Generalized Spatial Convolution) and Slim-Neck architecture on the neck network, aiming to reduce model complexity and computational requirements while maintaining detection accuracy.

[0077] The Slim-Neck module represents an innovative structural design tailored for the "neck" module of a target detector, with the core objective of improving detection accuracy and efficiency. By synergistically combining the GSConv module with a lightweight bottleneck structure (such as GSBottleneck), significant reductions in computational cost and parameter count are achieved while maintaining competitive detection performance.

[0078] The neck network connects to the backbone network, fuses multi-scale feature maps, and outputs detection results. The neck network includes the Slim-Neck module and the detection head. The Slim-Neck module receives and processes multi-scale feature maps. The Slim-Neck includes the GSConv module, a feature fusion module that integrates the GSConv module, and the GSBottleneck module.

[0079] Reference Figure 3The GSConv module includes a standard convolutional layer, a depthwise separable convolutional layer, and a channel shuffling unit. The standard convolutional layer receives the input feature map, the depthwise separable convolutional layer is connected to the output of the standard convolutional layer, and the channel shuffling unit rearranges the channel dimensions of the feature map.

[0080] The internal logic of the GSConv module is as follows: standard convolution → depthwise separable convolution → channel shuffling.

[0081] Furthermore, the feature fusion module, such as Figure 1 As shown in the marked Upsample, the feature fusion module has multiple modules connected in a multi-level series. The feature fusion module performs the following steps:

[0082] Step M1: Obtain the first feature map and the second feature map, wherein the resolution of the first feature map is lower than that of the second feature map;

[0083] Step M2: Upsample the first feature map to match its resolution with that of the second feature map;

[0084] Step M3: Concatenate the upsampled first feature map and the second feature map along the channel dimension to obtain the fused feature map.

[0085] Combination Figure 4 The input feature map is first processed by a standard convolutional (SC) layer with C2 / 2 input channels. Then, depthwise separable convolution (DSC) is applied to independently compute spatial features across the C2 / 2 channels, maintaining the same output channel dimension as the input. The DSC-processed features are concatenated with the original bypass features (also retaining C2 / 2 channels), and a channel shuffling operation is performed. It should be noted that the DSC processing and bypass function are concatenated and the channels are randomly arranged. This three-step design balances computational efficiency with feature representation capability.

[0086] In this embodiment, the resolution root of the feature map can be divided into 20×20, 40×40, 80×80, and 160×160. Taking the first feature fusion module along the Slim-Neck transport path as an example, it upsamples the 20×20 resolution deep feature map by 2 times through a bilinear interpolation unit and connects it to the addition operation unit at the 40×40 feature layer level to obtain the fused 40×40 resolution feature map. The next feature fusion module uses the same principle to further obtain the 80×80 resolution feature map.

[0087] Furthermore, referring to Figure 4 and Figure 5The GSBottleneck module includes two GSConv modules, a standard convolutional layer, and an additive fusion unit. The two GSConv modules are arranged sequentially along the data transmission path to process feature maps of different scales. The standard convolutional layer runs in parallel with the GSConv module at the output of the Slim-Neck architecture to preserve the original feature information. The additive fusion unit receives and fuses the feature maps from the GSConv module and the standard convolutional layer.

[0088] The input feature tensor is simultaneously fed into the GSConv module and a standard convolutional layer, and its output is element-wise summed to produce the final feature representation. Based on this architecture, combined with... Figure 1 The proposed VoVGSCSPC (Vision-Oriented Generalized Spatial Convolution and Shuffle-Connectivity Pyramid Context) employs a one-time aggregation strategy. Specifically, the neck network includes the VoVGSCSPC module, which comprises the GSBottleneck module and three sequentially configured standard convolutional layers. The GSBottleneck module is configured in parallel with the intermediate standard convolutional layers. This design significantly reduces network parameters and computational complexity by fusing multi-scale features in a single pass, while maintaining detection accuracy.

[0089] Furthermore, the detection head includes four detection layers, P2-P5, connected in a progressive manner, wherein the resolution of the P2, P3, P4 and P5 detection layers decreases progressively.

[0090] Different levels of detection layers are used to detect feature maps of different resolutions. In this embodiment, the P2 detection layer is used to detect feature maps of 160×160 resolution, the P3 detection layer is used to detect feature maps of 80×80 resolution, the P4 detection layer is used to detect feature maps of 40×40 resolution, and the P5 detection layer is used to detect feature maps of 20×20 resolution.

[0091] The integration of the P2 detection layer ensures the utilization of complementary features across scales, addressing the weakness of feature degradation in deeper layers of the baseline model. The 160×160 feature map not only inherently preserves rich low-level features but also enriches high-dimensional features with fine-grained details through feature fusion. This dual effect further optimizes the collaboration between object detection and localization (precise localization) tasks.

[0092] The P2 detection layer is specifically optimized to enhance the synergy between these tasks through two mechanisms:

[0093] (1) Fine-grained feature extraction: retain the key low-level patterns that define small targets.

[0094] (2) Efficient architecture design: Simplify information flow to prioritize high-resolution features in the early stages.

[0095] These innovations collectively endow the model with significantly improved accuracy in recognizing smaller objects, especially in scenarios where the target occupies the smallest pixel area (such as distant aquatic organisms). By bridging the semantic gap between low-level and high-level features, the P2 head addresses a key weakness in traditional multi-scale detection frameworks.

[0096] It should also be noted that the performance of surface target detection technology largely depends on high-quality, diverse training datasets. Existing surface target detection datasets vary significantly in terms of scene complexity and target diversity. Taking the four most typical maritime datasets—FloW, SeaShips, McShips, and WSODD—as examples, their key quantitative features are shown in Table 1, and their representative visual samples are as follows: Figure 6 As shown.

[0097] Table 1

[0098]

[0099] The FloW dataset focuses on detecting floating debris in inland rivers from the perspective of unmanned surface vehicles (USVs). It includes a single target type and is suitable for specific application scenarios, but it lacks the value of collaborative training with multiple target types. This limitation reduces the model's generalization ability in complex water environments.

[0100] The SeaShips dataset, by collecting close-range ship images from on-site video surveillance systems, establishes an annotation framework covering six major ship types. However, its high target-to-image ratio may cause the model to overfit large-scale target features during training, thereby reducing the detection sensitivity for small and distant targets and affecting the overall balance of detection performance.

[0101] The McShips dataset integrates annotated data from six classes of warships and seven classes of civilian vessels, establishing a ship detection resource pool with moderate class diversity. While this dataset achieves fine-grained classification of ship targets, the high similarity of targets within a class may limit the model's ability to capture subtle feature variations, thereby reducing inter-class distinguishability.

[0102] The Water Surface Object Detection Dataset (WSODD) constructs a comprehensive detection benchmark encompassing 14 object categories by covering diverse water surface environments (ocean, lakes, rivers) and varying lighting conditions (sunny days, foggy days, nighttime). While its environmental diversity provides strong support for model generalization, the sparsity of samples for certain categories (such as trees and animals) may lead to insufficient detection accuracy for these categories, thus affecting the stability of overall detection performance. The differences in scene complexity, object diversity, and annotation granularity among these datasets provide a multi-dimensional experimental benchmark for water surface object detection research.

[0103] Based on this, the object detection architecture also includes a data storage module, which stores the WSOD-12K dataset providing visual samples. The WSOD-12K dataset contains objects including ships, obstacles, rocks, platforms, and people. By strategically selecting or fusing datasets, researchers can effectively enhance the adaptability and reliability of the model in specific application scenarios, driving the development of surface object detection technology towards greater robustness and accuracy.

[0104] Furthermore, this application provides a target detection method that adopts the aforementioned target detection architecture. The target detection method includes the following steps:

[0105] Step S1: Input the image to be detected into the backbone network and extract the initial feature map;

[0106] Step S2: Input the initial feature map into the neck network and perform multi-scale feature fusion on the initial feature map;

[0107] Step S3: Input the fused feature map into the detection head and output the target detection result.

[0108] The following experiments will be conducted to verify the superiority of this scheme.

[0109] Related experiments:

[0110] 1) Experimental equipment:

[0111] The device is equipped with an AMD Ryzen 9 7945HX processor and an NVIDIA GeForce RTX 4060 graphics card, and can run a 64-bit Windows 11 operating system. The software environment is implemented using PyTorch 1.13.1 and Python 3.9. The YOLOv8n model was chosen as the baseline architecture.

[0112] 2) Preliminary preparations:

[0113] To ensure experimental consistency, training parameters were standardized, input images were adjusted to 640×64 pixels, the initial learning rate was 0.001, the batch size was 16, and the training epochs were 300. This configuration remained consistent across all experiments, isolating the impact of algorithm modifications while reflecting the limitations of real-world deployments. Hardware-software integration was validated, supporting efficient model training by leveraging the CUDA acceleration of the RTX 4060 and the multi-threaded processing of the Ryzen 9 CPU. This setup strikes a balance between practical computational needs and rigorous experimental control, adhering to the standards of reproducible research.

[0114] Correspondingly, the detailed data of the five object categories in the WSOD-12K dataset—ships, obstacles, rocks, platforms, and people—as well as the division of the validation set, test set, and training set are shown in Table 2.

[0115] Table 2

[0116]

[0117] Reference Figure 7 The proposed WSOD-12K dataset contains 11,973 images labeled with five classes: boats, obstacles, rocks, platforms, and people, totaling 32,730 instances. The dataset is divided into a training set (8,381 images), a validation set (1,796 images), and a test subset (1,796 images) in a 7:1.5:1.5 ratio. The class distribution is detailed in Table 2. The training set emphasizes major object types (e.g., boats account for 70.3% of the annotations). The validation and test sets maintain the same class ratio to rigorously evaluate the model's generalization ability. The dataset construction prioritizes scene diversity—covering diverse aquatic environments and meteorological conditions—and incorporates target scale distribution characteristics to reflect real-world detection challenges. This design ensures a comprehensive evaluation of the detector in terms of environmental variability and object size differences.

[0118] 3) Evaluation Indicators

[0119] This study establishes a multi-dimensional evaluation framework covering accuracy, complexity, and real-time performance. Each metric is defined as follows:

[0120] (1) Mean Precision (mAP):

[0121] As a primary precision metric in object detection, mAP quantifies the overall performance of a model by averaging the mean precision (AP) across all classes. The calculation is based on the harmonic relationship between precision and recall.

[0122] Precision: P = TP / (TP + FP), where TP represents true positives and FP represents false positives.

[0123] Recall: R = TP / (TP + FN), where FN represents false negatives.

[0124] By iteratively applying confidence thresholds, a precision-recall (PR) curve is constructed, and accuracy (AP) is defined as the area under this curve. For multi-class detection:

[0125] APi: Represents the AP of the i-th target category.

[0126] mAP:=(1 / N)ΣAPi, where N is the total number of categories.

[0127] Two specific mAP protocols were used:

[0128] mAP@0.5: mAP calculated when the Intersection over Union (IoU) threshold is fixed at 0.5.

[0129] mAP@0.5:0.95: The average mAP value calculated over an IoU threshold ranging from 0.5 to 0.95 in increments of 0.05.

[0130] (2) Model parameters

[0131] Measured in millions (M), this metric reflects the complexity of the architecture, directly impacting memory consumption during training and storage requirements during inference. A larger number of parameters enhances the model's expressiveness but increases deployment costs.

[0132] (3) GFLOPs (gigaflops per second)

[0133] This metric measures the computational complexity of the forward propagation process. Lower GFLOPs indicate higher algorithm efficiency, which is particularly beneficial for resource-constrained embedded devices.

[0134] (4) Reasoning speed (FPS)

[0135] Frames per second (FPS) directly reflects real-time processing capability. A higher FPS value indicates faster inference speed, which is crucial for real-time detection systems.

[0136] 4) Ablation test

[0137] To verify the proposed synergistic effect of improvements, we conducted a systematic ablation study on the YOLOv8n baseline model, focusing on the DualConv module, SlimNeck architecture, and small object detection head. All experiments were conducted under strictly controlled conditions, with identical hyperparameters, 300 training epochs, and consistent hardware configuration. Evaluation metrics included mAP@0.5, mAP@0.5:0.95, parameters, computational complexity (GFLOPs), and inference speed (FPS), detailed in Table 3.

[0138] Table 3

[0139]

[0140] The integration of the DualConv and SlimNeck modules maintained comparable mAP metrics while achieving a 17.3% reduction in parameters and a 21.6% reduction in GFLOPs. This demonstrates the effectiveness of dual-path convolution and channel optimization in enhancing feature representation efficiency. The P2-based detection head improved mAP@0.5 and mAP@0.5:0.95 by 1.4% and 2.1%, respectively, validating its ability to capture fine-grained features by expanding the receptive field. Although the FPS decreased from 93.5 to 75.4, the system still met the requirements for real-time operation. The integrated DSH-YOLO model improved mAP@0.5 and mAP@0.5:0.95 by 2.0% and 1.0% over the baseline, respectively, while maintaining parameter efficiency. These results highlight the necessity of multi-component collaborative optimization, providing a solution that balances accuracy and efficiency for marine target detection in complex aquatic environments.

[0141] 5) Comparative Experiment

[0142] To verify the effectiveness of the proposed model, we conducted comparative experiments with various object detection methods, including YOLOv3-tiny, YOLOv5n, YOLOv8n, YOLOv9t, YOLOv10n, YOLOv11n, and YOLOv12n. These models represent different development stages of the YOLO family, from early architectures to the latest versions. All experiments were conducted under identical conditions to ensure the fairness of the results. The experimental results are shown in Table 4.

[0143] Table 4

[0144]

[0145] The analysis in Table 4 shows that the overall performance of YOLOv3-tiny and YOLOv10n is inferior to that of YOLOv8. Although YOLOv5n has a higher inference speed, its detection accuracy is significantly lower. While YOLOv9t is comparable to YOLOv8 in terms of mAP and GFLOPs, its lower FPS makes it unsuitable for real-time applications. In contrast, our improved DSH-YOLO model outperforms YOLOv8n by 2% and 1% in mAP@0.5 and mAP@0.5:0.95, respectively, while reducing the number of parameters by 17.3%.

[0146] Furthermore, the detection performance metrics of the YOLOv8n model, which serves as the baseline architecture, were compared, and the experimental results are shown in Table 5.

[0147] Table 5

[0148]

[0149] Table 5 further demonstrates that DSH-YOLO outperforms the baseline model in all assessment categories (ship, obstacle, rock, platform, and person), significantly improving detection accuracy. Furthermore, DSH-YOLO (this scheme) also surpasses the latest YOLOv11n and YOLOv12n in detection accuracy.

[0150] Furthermore, to verify the versatility of this scheme, the dataset was replaced with the WSODD dataset, and comparative experiments were conducted on various object detection methods, including YOLOv3-tiny, YOLOv5n, YOLOv8n, YOLOv9t, YOLOv10n, YOLOv11n, and YOLOv12n. The experimental results are shown in Table 6.

[0151] Table 6

[0152]

[0153]

[0154] Table 6 shows that the proposed method also has good detection performance on other datasets, verifying its versatility.

[0155] Combination Figure 8 The results show the training curves of different models over 300 epochs, demonstrating that DSH-YOLO exhibits superior average accuracy throughout the training process. These results collectively prove that DSH-YOLO not only improves detection accuracy but also maintains high inference efficiency, outperforming existing mainstream detection frameworks in overall performance.

[0156] To comprehensively evaluate the performance of DSH-YOLO, experiments were conducted in five different scenarios (sunny day, foggy day, night, single target, and multi-target environment).

[0157] like Figure 9 As shown, DSH-YOLO demonstrated superior detection accuracy and efficiency under all test conditions, exhibiting strong environmental adaptability, especially in adverse weather conditions such as fog and nighttime. In both single-target and multi-target scenarios, DSH-YOLO achieved not only high accuracy but also outstanding computational efficiency.

[0158] Figure 10 The visualizations showcase the results of multiple detectors on the WSODD dataset, particularly excelling in small object detection. Overall, DSH-YOLO demonstrates enhanced robustness and superior performance across various complex environments, consistently maintaining a high mAP@0.5 score across different categories and scenarios.

[0159] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A lightweight target detection architecture based on multi-scale feature fusion, characterized in that, include: The backbone network receives images and outputs multi-scale feature maps, including: Multiple DualConv modules are set sequentially along the data transmission path; Multiple C2F modules are configured to correspond one-to-one with the multiple DualConv modules and are connected to adjacent DualConv modules; The SPPF module is connected to the C2F module at the end of the backbone network; A neck network, connected to the backbone network, fuses multi-scale feature maps and outputs detection results. The neck network includes: The Slim-Neck module receives and processes multi-scale feature maps, including the GSConv module, a feature fusion module that integrates the GSConv module, and the GSBottleneck module. The detection head includes four detection layers, P2-P5, connected in a progressive manner, wherein the resolution of the P2, P3, P4 and P5 detection layers decreases progressively.

2. The lightweight target detection architecture for enhanced multi-scale feature fusion according to claim 1, characterized in that, The DualConv module includes: The channel grouping unit divides the input feature map into two groups along the channel dimension; A parallel convolutional unit, connected to the output of the channel grouping unit, includes a 3×3 convolutional kernel and a 1×1 convolutional kernel arranged in parallel; A feature recombination unit is connected to the output of the parallel convolution unit to shuffle the two sets of feature maps.

3. The lightweight target detection architecture for enhanced multi-scale feature fusion according to claim 1, characterized in that, The GSBottleneck module includes: Two GSConv modules are set up sequentially along the data transmission path to process feature maps of different scales. The standard convolutional layer, running in parallel with the GSConv module at the output of the Slim-Neck architecture, is used to preserve the original feature information; The additive fusion unit receives and fuses the feature maps of the GSConv module and the standard convolutional layer.

4. A lightweight target detection architecture for enhanced multi-scale feature fusion according to claim 2 or 3, characterized in that, The GSConv module includes: A standard convolutional layer receives the input feature map; A depth-separable convolutional layer is connected to the output of the standard convolutional layer. The channel shuffling unit, located between the two sets of C2 / 2 input channels, rearranges the feature map according to the channel dimensions.

5. The lightweight target detection architecture for enhanced multi-scale feature fusion according to claim 1, characterized in that, The feature fusion module is configured with multiple modules connected in a multi-level series. The feature fusion module performs the following steps: Step M1: Obtain the first feature map and the second feature map, wherein the resolution of the first feature map is lower than that of the second feature map; Step M2: Upsample the first feature map to match its resolution with that of the second feature map; Step M3: Concatenate the upsampled first feature map and the second feature map along the channel dimension to obtain the fused feature map.

6. The lightweight target detection architecture for enhanced multi-scale feature fusion according to claim 1, characterized in that, The P2 detection layer detects feature maps with a resolution of 160×160.

7. The lightweight target detection architecture for enhanced multi-scale feature fusion according to claim 1, characterized in that, It also includes a data storage module that stores a WSOD-12K dataset that provides visual samples. The objects in the WSOD-12K dataset include ships, obstacles, rocks, platforms, and people.

8. A target detection method, characterized in that, Employing the target detection architecture described in any one of claims 1-7, the target detection method includes the following steps: Step S1: Input the image to be detected into the backbone network and extract the initial feature map; Step S2: Input the initial feature map into the neck network and perform multi-scale feature fusion on the initial feature map; Step S3: Input the fused feature map into the detection head and output the target detection result.

9. The application of the target detection architecture according to any one of claims 1-7 in surface navigation safety monitoring.

Images