CoChat
Sign Up
Cited by 415Advanced Neural Network ApplicationsDomain Adaptation and Few-Shot LearningVisual Attention and Saliency Detection

A Modified YOLOv8 Detection Network for UAV Aerial Image Recognition

Open access
Li, Yiting, Fan, Qingsong, Huang, Haisong, Han, Zhenggong, Gu, Qiang·2023·Drones
DOI 10.3390/drones7050304
Open this paper in CoChat — free accountOpen PDF

Abstract

UAV multitarget detection plays a pivotal role in civil and military fields. Although deep learning methods provide a more effective solution to this task, changes in target size, shape change, occlusion, and lighting conditions from the perspective of drones still bring great challenges to research in this field. Based on the above problems, this paper proposes an aerial image detection model with excellent performance and strong robustness. First, in view of the common problem that small targets in aerial images are prone to misdetection and missed detection, the idea of Bi-PAN-FPN is introduced to improve the neck part in YOLOv8-s. By fully considering and reusing multiscale features, a more advanced and complete feature fusion process is achieved while maintaining the parameter cost as much as possible. Second, the GhostblockV2 structure is used in the backbone of the benchmark model to replace part of the C2f module, which suppresses information loss during long-distance feature transmission while significantly reducing the number of model parameters; finally, WiseIoU loss is used as bounding box regression loss, combined with a dynamic nonmonotonic focusing mechanism, and the quality of anchor boxes is evaluated by using “outlier” so that the detector takes into account different quality anchor boxes to improve the overall performance of the detection task. The algorithm’s performance is compared and evaluated on the VisDrone2019 dataset, which is widely used worldwide, and a detailed ablation experiment, contrast experiment, interpretability experiment, and self-built dataset experiment are designed to verify the effectiveness and feasibility of the proposed model. The results show that the proposed aerial image detection model has achieved obvious results and advantages in various experiments, which provides a new idea for the deployment of deep learning in the field of UAV multitarget detection.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper proposes a modified YOLOv8-s detector for UAV aerial images, adding Bi-PAN-FPN in the neck, GhostblockV2 in the backbone, and Wise-IoU v3 as the bounding-box regression loss. On VisDrone2019, the authors report improved mAP and fewer parameters versus the YOLOv8-s baseline, with additional comparison, interpretability, and self-built dataset experiments.

Key findings

  • Compared with the original YOLOv8-s benchmark, the proposed model improved mAP on the VisDrone2019 test set by 9.06% and reduced the number of parameters by 13.21%, as reported in the abstract and conclusions.
  • In the overall ablation table, the final model (A + B + C + D) reached test mAP 0.337 versus 0.309 for the baseline A, with parameters reduced from 11.129 million to 9.659 million and model size from 21.972 MB to 19.157 MB.
  • The improved model reached 167 f/s in the ablation experiment, while the baseline reached 182 f/s; the authors state this still supports real-time deployment requirements.
  • On the self-built dataset, the proposed model achieved the best detection accuracy, with the conclusion reporting mAP 91.7% and FPS up to 293/f.s-1; Table 6 lists mAP 0.917 and FPS 294 for the proposed model.

Methods

  • The model modifies YOLOv8-s in three ways: replacing PAN-FPN with Bi-PAN-FPN and adding upsampling to fuse B2 features; replacing some backbone C2f modules with GhostblockV2; and replacing CIoU with Wise-IoU v3 using a dynamic nonmonotonic focusing mechanism.
  • Experiments use VisDrone2019, divided into 6471 training, 548 validation, and 1610 test samples, with images normalized to 640 × 640; a self-built DJI Mavic 3 dataset of 906 images with four categories is split into 725 training and 181 testing samples.
  • The paper reports ablation experiments, comparison experiments against YOLOv4-s, YOLOv5-s, YOLOv5-m, YOLOX-s, YOLOv7-tiny, and MobileNetv2-SSD, Grad-CAM interpretability analysis, and self-built dataset validation.

Limitations

  • The paper states that the proposed model does not outperform other structures in all small categories: tricycle and bus performance is not as good as A + B + C, and van performance is not as good as A + B.
  • The paper also notes missed detections and false detections in small-target and dense-target tasks, including truck/bus confusion and car/pedestrian targets being treated as background.
  • Evident from the text: the self-built dataset is relatively small (906 images, four categories), and all models were trained without pretrained weights, which may limit generality and comparability with pretrained deployment settings.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/drones7050304

Cites

39 of 39 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 9 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Compared with the original benchmark network, the number of parameters of the model is reduced by 13.21% on the test set.

    “the number of parameters is reduced by 13.21% (test set)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The VisDrone2019 dataset is widely used worldwide and is a suitable benchmark for evaluating UAV aerial image detection.
  • Findingown data

    Compared with the original benchmark network, the MAP performance of the model on VisDrone2019 is improved by 9.06% on the test set.

    “Compared with the original benchmark network, the MAP performance of the model on the international open-source dataset VisDrone2019 is improved by 9.06% (test set)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The VisDrone2019 dataset is widely used worldwide and is a suitable benchmark for evaluating UAV aerial image detection.
  • Findingown data

    MobilNetv2-SSD is not suitable for complex tasks such as drone multi-target detection.

    “There was verification in Section 4.3 that MobilNetv2-SSD is not suitable for complex tasks such as drone multi target detection.”
    Quote in the paperSubstantial claim
  • Findingnone given

    YOLOX-s had the worst FPS performance.

    “However, the model had the worst FPS performance.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The FPS of the improved model reached 167 f/s.

    “The FPS index of the improved model reached 167/f.s-1,which can also ensure real-time requirements in actual deployment.”
    Quote in the paperSubstantial claim
  • May combine two claims. 51
  • Classified as probably not a substantial claim. 6
  • The quote only partly states this claim. 2

  • May combine two claims.

    Findingown data

    The proposed model led the original YOLOv8-s by nearly two percentage points in test-set mAP, and led the worst-performing YOLOv7-tiny by 18.4 percentage points.

    “The model in this article led the original YOLOv8-s by nearly two percentage points in the map of the test set, and led the worst performing YOLOv7-tiny by 18.4 percentage points.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    From the three indicators of FPS, parameters, and model size, the proposed model's performance was in the first echelon, showing that the model had the best comprehensive performance.

    “From the three indicators of FPS, parameters, and model size, the model performance was in the first echelon, which shows that the model had the best comprehensive performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    YOLOv4-s's performance was relatively mediocre, only better than MobilNetv2-SSD in the detection task, but its R value was greatly improved compared with the previous two models, and the missed detection rate decreased.

    “The performance of YOLOv4-s was relatively mediocre, and it was only better than MobilNetv2-SSD in the detection task, but the R value of this model was greatly improved compared with the previous two models, and the missed detection rate decreased.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In FPS, YOLOv5-s and YOLOv8-s both exceeded 300 f/s, while YOLOX-s did not reach 100 f/s.

    “In terms of FPS, YOLOv5-s and YOLOv8-s both exceeded 300/f.s-1,but YOLOX-s did not reach 100 in this indicator”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the self-built dataset, YOLOv4-s and YOLOv7-tiny performed similarly, both obtaining relatively low mAP values on the test set.

    “YOLOv4 s and YOLOv7-tiny performed similarly on the self-built dataset, both obtaining relatively low map values on the test set.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The performance of tricycle and bus is not as good as that of A + B + C, and the performance of van is not as good as that of A + B.

    “For example, the performance of tricycle and bus is not as good as that of A + B + C, and the performance of van is not as good as that of A + B.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    YOLOX-s had a relatively low R value on this dataset and a high missed detection rate.

    “The R value of the model on this dataset was relatively low, and the missed detection rate was high.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the self-built dataset the proposed model achieved the optimal detection accuracy (mAP 91.7%) with an FPS of up to 293 f/s.

    “the model also achieved the optimal detection accuracy (MAP: 91.7%) on the self-built dataset, with an FPS of up to 293/f.s-1.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Compared with the non-lightweight model YOLOv5-m, the P, R, and MAP metrics of the proposed model all performed comparably or better than YOLOv5-m.

    “Compared with the non-lightweight model YOLOv5-m, the P, R, and MAP metrics all performed comparably or better than YOLOv5-m.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The performance of YOLOv5-s and YOLOv7-tiny in the test set was second only to the lightweight model proposed in this paper, and they were also suitable for target detection tasks in UAV aerial images.

    “The performance of YOLOv5-s and YOLOv7-tiny in the test set was second only to the lightweight model proposed in this paper, and they were also suitable for target detection tasks in UAV aerial images.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    YOLOX-s's p value achieved the best results in both the validation set and the test set, making it achieve better results than YOLOv4-s.

    “The p value achieved the best results in both the validation set and the test set, which made the YOLOX-s model achieve better results (better than those of YOLOv4-s).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    YOLOv5-s, YOLOX-s, and YOLOv8-s all achieved approximately the same level of detection accuracy, with YOLOv8-s having the best accuracy.

    “YOLOv5-s, YOLOX-s, and YOLOv8-s all achieved excellent results, approximately at the same level of detection accuracy, but YOLOv8-s had the best accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The benchmark model (A) performed poorly, with the lowest accuracy indicators but the highest FPS, reaching 182 f/s.

    “The A model (i.e., the benchmark model) performed poorly.Its accuracy indicators were in the lowest position, but the FPS was in the first place, reaching 182/fs-1.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    YOLOv7-tiny was the smallest model.

    “YOLOv7-tiny was the smallest model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The A model (benchmark YOLOv8-s) performed poorly, with accuracy indicators in the lowest position, but its FPS was in the first place, reaching 182/f.s-1.

    “The A model (i.e., the benchmark model) performed poorly.Its accuracy indicators were in the lowest position, but the FPS was in the first place, reaching 182/f.s-1.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    In some cases the proposed model produces redundant detection boxes, and in a few cases similar backgrounds are mistaken for targets.

    “However, in some cases, redundant detection boxes may appear, and in a few cases, similar backgrounds may be mistaken for targets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The Grad-CAM map of the model in this paper is the best, and the dark red parts (focus areas) are the same type of targets.

    “the Grad-CAM map of the model in this paper is the best, and the dark red parts (focus areas) are the same type of targets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the whole, the sequential improvement of the three structures made the model in the VisDrone dataset improve the accuracy continuously, the number of parameters gradually decreased, and the final model size was only 19.157 MB.

    “On the whole, the sequential improvement of the three structures made the model in the VisDrone dataset improve the accuracy continuously, the number of parameters gradually decreased, and the final model size was only 19.157 MB.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The self-built dataset contains over 10,000 manually annotated bounding boxes, with four detection categories (people, motorcycles, cars, and bicycles), totaling 906 sheets.

    “The dataset contains a total of over 10,000 manually annotated bounding boxes, with four detection categories (people, motorcycles, cars, and bicycles), totaling 906 sheets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    After incorporating the D structure, the model's indicator data in some categories were not as good as before; in some cases, A + B + C was better than A + B + C + D, but this did not affect the overall trend.

    “However, after incorporating the D structure, the model's indicator data in some categories were not as good as before.”
    • Also stated in: § 4.2. Ablation Experiment · plain
      “That is, in some cases, A + B + C was better than A + B + C + D, but this did not affect the overall trend.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The two lightweight models, YOLOv5-s and YOLOv7-tiny, both achieved excellent performance in the test set.

    “The two lightweight models, YOLOv5-s and YOLOv7-tiny, both achieved excellent performance in the test set.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The P and R indicators of YOLOv5-s and YOLOv7-tiny were in a relatively balanced state, and the detection rate and the correct rate were relatively coordinated.

    “The P and R indicators of the two models were in a relatively balanced state, and the detection rate and the correct rate were relatively coordinated.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The recognition of cars and buses was relatively good; the difficult-to-recognize categories included bicycles and people; trunks, tricycles, and awning-tricycles had the worst recognition effect.

    “The detailed analysis shows that the recognition of cars and buses was relatively good; the difficult-to-recognize categories included bicycles and people; trunks, tricycles, and awning-tricycles had the worst recognition effect.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    MobilNetv2-SSD's R index was the lowest in both the validation set and the test set, indicating a large number of missed detections.

    “In the target detection task, its R index was the lowest in both the validation set and the test set, which means that the model had a large number of missed detections.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    MobilNetv2-SSD's p value showed a high performance, indicating that the objects detected by the model were easier to identify, except for the missed objects.

    “However, its p value showed a high performance, and it can be seen that the objects detected by the model were easier to identify, except for the missed objects.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Compared with the baseline model, the proposed method's mAP on the test set improved by 9.06% and its number of parameters was reduced by 13.21%.

    “Compared with the baseline model, the MAP performance of our method on the test set is improved by 9.06%, and the number of parameters is reduced by 13.21%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Both YOLOv4-s and YOLOv7-tiny had more than 150 f/s FPS.

    “Both had more than 150/f.s-1 in FPS, and are capable of being deployed in Edge device”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    YOLOv7-tiny had a certain area of concern for similar targets but could not cover a large number of targets, especially when there were occlusions and tiny types of targets.

    “YOLOv7-tiny, which had the worst performance among the three models, had a certain area of concern for similar targets but could not cover a large number of targets.This was especially true when there were occlusions and tiny types of targets;”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    YOLOv5-s had a greater improvement than YOLOv7-tiny; the focus area in YOLOv5-s was significantly improved, which played a crucial role in comprehensively improving the detection accuracy of various categories.

    “YOLOv5-s had a greater improvement than YOLOv7-tiny.The focus area in YOLOv5-s was significantly improved.This played a crucial role in comprehensively improving the detection accuracy of various categories;”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    The GhostblockV2 structure suppresses information loss during long-distance feature transmission while significantly reducing the number of model parameters.

    “the GhostblockV2 structure is used in the backbone of the benchmark model to replace part of the C2f module, which suppresses information loss during longdistance feature transmission while significantly reducing the number of model parameters”
    • Also stated in: § 3. Improved Aerial Image Detection Model · plain
      “use the GhostblockV2 structure to replace some C2f modules in backbone, which suppresses information loss during long-distance feature transmission while significantly reducing the number of model parameters.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Ghostblock follows the essence of GhostNet and ensures feature diversity while efficiently extracting features.

    “On the one hand, Ghostblock follows the essence of GhostNet.It first uses conventional convolution to generate the original feature map and then combines various linear transformation operations to enhance the feature map's information.This ensures feature diversity while efficiently extracting features.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [32] Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., and Xu, C. (2020, January 13–19). Ghostnet: More Features from Cheap Operations. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA. source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    Ghostblock combines cheap operation with DFC, which greatly reduces the complexity of the model while taking into account the global information of features.

    “Ghostblock combines cheap operation with DFC, which greatly reduces the complexity of the model while taking into account the global information of features.”
    • Also stated in: § 3.2. Improvement of the Backbone · plain
      “Ghostblock combines cheap operation with DFC, which greatly reduces the complexity of the model while taking into account the global information of features.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Complex object detection models are difficult to deploy in edge devices, and lightweight object detectors have difficulty improving accuracy.

    “Complex object detection models are difficult to deploy in edge devices, and lightweight object detectors have difficulty improving accuracy [13][14][15].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [13][14][15]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    UAV aerial photography images have characteristics of large scenes, small targets, multiscale, complex backgrounds, and overlapping occlusions.

    “the aerial photography images of UAVs are different from ground photography and have the characteristics of large scenes, small targets, being multiscale, having complex backgrounds [9], and overlapping occlusions [10][11][12].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [9] Cai, 2020, Remote Sensing Image Classification Based on a Cross-Attention Mechanism and Graph Convolution, IEEE Geosci. Remote Sens. Lett.

      Not checked in the free run.

    • cites [10][11][12]

      Not checked in the free run.

  • The quote only partly states this claim.

    Interpretationreasoning

    WiseIoU loss combined with a dynamic nonmonotonic focusing mechanism improves the overall performance of the detection task.

    “WiseIoU loss is used as bounding box regression loss, combined with a dynamic nonmonotonic focusing mechanism, and the quality of anchor boxes is evaluated by using”
    • Also stated in: § 3. Improved Aerial Image Detection Model · plain
      “It incorporates a dynamic nonmonotonic focusing mechanism to evaluate anchor box quality by using”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Interpretationown data

    Introducing Bi-PAN-FPN improves the detection ability of the model for small targets in aerial images.

    “the idea of Bi-PAN-FPN is introduced to improve the neck part in YOLOv8-s.”
    • Also stated in: § 1. Introduction · plain
      “this work improves the detection ability of the model for small targets”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Small targets in aerial images are prone to misdetection and missed detection.

  • May combine two claims.

    Premisereasoning

    The decoupling of the two directions greatly simplifies the process of extracting global information of features, and the complexity of DFC is greatly reduced.

    “The decoupling of the two directions greatly simplifies the process of extracting global information of features.At the same time, due to the use of deep separable structures such as 1 * K H and K W * 1, the complexity of the DFC is greatly reduced (full connection: O(H 2 W + HW 2 ); DFC: O(K H HW + K W HW)).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The proposed model improves detection ability for small targets by introducing Bi-PAN-FPN and increasing multiscale feature fusion.

    “From the perspective of paying attention to large-size feature maps and introducing the idea of Bi-PAN-FPN, this work improves the detection ability of the model for small targets, and at the same time increases the probability and time of multiscale feature fusion to obtain better feature engineering.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Small targets in aerial images are prone to misdetection and missed detection.

  • May combine two claims.

    Premisereasoning

    Pointwise convolution loses the interaction process with other pixels in space, weakening the representation of spatial information and affecting the detection accuracy of the model.

    “However, the drawbacks of doing so are also obvious: pointwise convolution loses the interaction process with other pixels in space, which results in only the feature map obtained using depth-wise convolution capturing spatial information.The representation of spatial information will be significantly weakened, thus affecting the detection accuracy of the model.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The cheap operation significantly reduces parameter cost and computational cost by reusing features and discards redundant information that may exist in conventional convolutions.

    “This processing method significantly reduces the parameter cost and computational cost by reusing features and discards redundant information that may exist in conventional convolutions.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [32] Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., and Xu, C. (2020, January 13–19). Ghostnet: More Features from Cheap Operations. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA. source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The model suppresses the number of parameters while improving accuracy.

    “Suppresses the number of parameters of the model while improving the accuracy of the model.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The DFC attention mechanism avoids the limitations of traditional attention algorithms in computational complexity and captures feature information over long distances.

    “Through its particularity, this mechanism avoids the limitations of traditional attention algorithms in terms of computational complexity and captures feature information over long distances.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] Tang, Y., Han, K., Guo, J., Xu, C., Xu, C., and Wang, Y. (2022). GhostNetV2: Enhance Cheap Operation with Long-Range Attention. arXiv.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Bi-PAN-FPN achieves more advanced and complete feature fusion while maintaining parameter cost as much as possible.

    “By fully considering and reusing multiscale features, a more advanced and complete feature fusion process is achieved while maintaining the parameter cost as much as possible.”
    • Also stated in: § 3. Improved Aerial Image Detection Model · plain
      “By fully considering and reusing multi-scale features, we achieve more advanced and comprehensive feature fusion while maintaining parameter costs as much as possible.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The proposed YOLOv8-s-based aerial image detection model can accurately detect aerial image targets in real time while satisfying edge-device deployment.

    “This paper proposes an aerial image detection model based on YOLOv8-s, which can accurately detect aerial image targets in real time under the premise of satisfying the deployment of edge devices.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Due to the high similarity between the truck class and the bus class in the high-altitude perspective, there were many misdetected targets in the detection process; if the car class and pedestrian class are too small, they may be considered by the model as backgrounds and missed detections.

    “For example, due to the high similarity between the truck class and the bus class in the high-altitude perspective, there were many misdetected targets in the detection process; if the car class and pedestrian class are too small, they may be considered by the model as backgrounds and missed detections.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    YOLOv4-s and YOLOv7-tiny are capable of being deployed in edge devices.

    “Both had more than 150/f.s-1 in FPS, and are capable of being deployed in Edge device”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The proposed method is suitable for deployment in complex working conditions and has considerable universality and robustness.

    “In general, the proposed method is suitable for deployment in complex working conditions, and also has considerable universality and robustness.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The scenarios and datasets tested (VisDrone and the self-built dataset) adequately represent the range of real aerial detection conditions, so that best performance on them implies strong universality and robustness.

  • May combine two claims.

    Interpretationown data

    The proposed model achieved the best detection performance across various scenarios and datasets and has strong universality.

    “This also indicates that the model in this paper achieved the best detection performance in various scenarios and datasets, and has strong universality.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The scenarios and datasets tested (VisDrone and the self-built dataset) adequately represent the range of real aerial detection conditions, so that best performance on them implies strong universality and robustness.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    All ablation experiments and various model training processes in the comparison experiments did not use any pretraining weights to ensure fairness and comparability of the model effects.

    “To ensure the fairness and comparability of the model effects, all ablation experiments and various model training processes in the comparison experiments did not use any pretraining weights.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    To ensure fairness, all models in the comparison experiment were not trained with pretrained weights.

    “To ensure fairness, all models were not trained with pretrained weights.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    After integrating the B, C, and D structures, the model improved performance in several aspects: focusing on small target features, multiplexing multiscale features, suppressing information loss during long-range feature transmission, and taking into account anchor boxes of different qualities, feature engineering was significantly strengthened.

    “After integrating the B, C, and D structures, the model improved performance in several aspect-focusing on small target features, multiplexing multiscale features, suppressing information loss during long-range feature transmission, and taking into account anchor boxes of different qualities, feature engineering was significantly strengthened.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Regardless of the scenario, the constructed model had strong detection ability, and the robustness met the actual engineering needs.

    “Regardless of the scenario, the constructed model had strong detection ability, and the robustness met the actual engineering needs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Using the GhostblockV2 structure in the backbone to replace part of the C2f module suppresses information loss during long-distance feature transmission while significantly reducing the number of model parameters.

    “the GhostblockV2 structure is used in the backbone of the benchmark model to replace part of the C2f module, which suppresses information loss during longdistance feature transmission while significantly reducing the number of model parameters”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    In the UAV aerial image target detection task involved in this paper, the proposed model met the needs of actual production scenarios in terms of detection accuracy and deployment difficulty and had considerable robustness and practicability.

    “In the UAV aerial image target detection task involved in this paper, the proposed model met the needs of actual production scenarios in terms of detection accuracy and deployment difficulty and had considerable robustness and practicability.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Processing aerial images in real time is more in line with practical engineering scenarios.

  • May combine two claims.

    Interpretationown data

    The improvement to the baseline model is feasible and effective, taking into account the accuracy and speed of edge device detection scenarios.

    “This shows that the improvement to the baseline model is feasible and effective, taking into account the accuracy and speed of edge device detection scenarios.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    The performance of the proposed method has strong interpretability.

    “The performance of this method has strong interpretability.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Even though the improved model reduced the number of parameters to 9.659 million, it still increased the number of network layers and some inference time.

    “This indicates that even if the number of model parameters was reduced (only 9.659 million) in the improved model, it would still increase the number of network layers and some inference time.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    In the comparison experiment, all models were trained without pretraining weights.

    “The comparison indicators are consistent with those in Table 3 of the ablation experiment, and all models do not use pretraining weights for training.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The sample size normalization to 640 × 640 can make the model truly deployable to edge devices without destroying too much of the useful information of the image.

    “Such a size can make the model truly deployable to edge devices without destroying too much of the useful information of the image.”
    Quote in the paperNot a substantial claim
59 of 59

  • The paper relies on this without support.

    Small targets in aerial images are prone to misdetection and missed detection.

    “in view of the common problem that small targets in aerial images are prone to misdetection and missed detection”
    “aiming at the common problem that small targets in aerial images are prone to misdetection and missed detection”

  • The paper relies on this without support.

    The VisDrone2019 dataset is widely used worldwide and is a suitable benchmark for evaluating UAV aerial image detection.

    “The algorithm's performance is compared and evaluated on the VisDrone2019 dataset, which is widely used worldwide”

  • The paper relies on this without support.

    Intelligent understanding of UAV visual data is more difficult than conventional computer vision tasks due to shooting angle, light, background, and other factors.

    “Due to the comprehensive influence of shooting angle, light, background, and other factors, intelligent understanding of UAV visual data is more difficult than conven-tional computer vision tasks.”

  • The paper relies on this without support.

    Processing aerial images in real time is more in line with practical engineering scenarios.

    “Processing aerial images in real time is more in line with practical engineering scenarios.”

  • The paper relies on this without support.

    MobilNetv2-SSD often has high applicability in simpler tasks.

    “MobilNetv2-SSD often has high applicability in simpler tasks, but it is not suitable for this type of complex task.”

  • The paper relies on this without support.

    Deep learning models lack interpretability, which has limited their progress in some high-tech fields.

    “Although deep learning models are widely used in various types of engineering fields, due to the lack of interpretability of algorithms, deep learning has not made great progress in some high-tech fields.”

  • The paper relies on this without support.

    The scenarios and datasets tested (VisDrone and the self-built dataset) adequately represent the range of real aerial detection conditions, so that best performance on them implies strong universality and robustness.

    “(a) Dense + small target ( b) Cloudy (c) Sunny ( d) Complex background (e) Very small target + sparse ( f) Small target + shaded”

  • The paper relies on this without support.

    An FPS of about 167 f/s is sufficient to meet real-time requirements in actual edge deployment.

    “The FPS index of the improved model reached 167/f.s-1,which can also ensure real-time requirements in actual deployment.”
8 of 8

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Autoencoders and GANs have notably expanded unsupervised learning's applicability, enabling more sophisticated data generation and representation learning.

    “This method has seen significant advancements with the introduction of autoencoders [171], [172] and Generative Adversarial Networks (GANs) [173], [174], [175], which have notably expanded unsupervised learning's applicability, enabling more sophisticated data generation and representation learning capabilities.”

    From From Google Gemini to OpenAI Q* (Q-Star): A Survey on Reshaping the Generative Artificial Intelligence (AI) Research Landscape (2025) · cites this paper as [171]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Self-built dataset FPS is reported as both 293 and 294

    Minor

    The conclusion reports an FPS of 293 for the proposed model on the self-built dataset, while Table 6 reports 294.

  • Parameter count does not decrease at every improvement step

    Minor

    The paper says the number of parameters gradually decreased as structures were added, but Table 3 shows an increase from 11.129 million for A to 11.409 million for A + B. It then remains unchanged between A + B + C and A + B + C + D.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. Ge, Z., Liu, S., Wang, F., Li, Z., and Sun, J. (2021). Yolox: Exceeding Yolo Series in 2021. arXiv.
  2. Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., and Vasudevan, V. (November, January 27). Searching for mobilenetv3. Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea.

Cited by

Sorted by citations

Related

Authors
Li, Y., Fan, Q., Huang, H., Han, Z., & Gu, Q.
Affiliations
  • Guizhou University of Finance and Economics (CN)
  • Guizhou University (CN)
Corresponding author
Fan, Qingsong
ORCID
Venue
Drones
Publisher
MDPI AG
Published
2023-05-05
Volume
7
Issue
5
Pages
304
ISSN
2504-446X (ISSN-L)
DOI
10.3390/drones7050304
OpenAlex
W4372279253
Cited by
415
References
39
Citations by year
2026: 51 · 2025: 164 · 2024: 171 · 2023: 29
FWCI
37.51
Type
journal article
Language
en
Open access
Licence
Locations
  • Drones (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Youth Science and Technology Talent Growth Project of Guizhou Provincial Department of Education (KY [2022] 199, 2021KYYB08, 52165063, ZK [2021]337, QianJiaoJi [2022]436, YJSCXJH [2021] 068, ZK[2023]029)
  • Research Fund of Guizhou University of Finance and Economics (KY [2022] 199, 2021KYYB08, 52165063, ZK [2021]337, QianJiaoJi [2022]436, YJSCXJH [2021] 068, ZK[2023]029)
  • National Natural Science Foundation of China (KY [2022] 199, 2021KYYB08, 52165063, ZK [2021]337, QianJiaoJi [2022]436, YJSCXJH [2021] 068, ZK[2023]029)
  • Guizhou Provincial Science and Technology Plan Project (KY [2022] 199, 2021KYYB08, 52165063, ZK [2021]337, QianJiaoJi [2022]436, YJSCXJH [2021] 068, ZK[2023]029)
  • the Open Fund Project supported by the Key Laboratory of Advanced Manufacturing Technology Ministry of Education (KY [2022] 199, 2021KYYB08, 52165063, ZK [2021]337, QianJiaoJi [2022]436, YJSCXJH [2021] 068, ZK[2023]029)
  • the Guizhou Province Graduate Research Fund (KY [2022] 199, 2021KYYB08, 52165063, ZK [2021]337, QianJiaoJi [2022]436, YJSCXJH [2021] 068, ZK[2023]029)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Visual Attention and Saliency Detection
Keywords
aerial image object detection, YOLOv8, small object detection, multi-scale feature fusion, occlusion, VisDrone2019

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review