Scholarlink

<p>Real-time small-object detection in aerial images and videos is a critical challenge due to scale variation, limited spatial resolution, and strict onboard computational constraints. We propose LASD-YOLO (lightweight area-attention-based small-object detector), a YOLOv11-based detector specifically tailored for real-time small-object detection in aerial images. The proposed architecture introduces a high-resolution detection head with a stride of 4, enabling the detection of very small objects that occupy 4 to 8 pixels in an image. To balance the increased computational cost, we replace the standard convolutional layers with depthwise convolution operations. Selected backbone stages are also augmented with area attention blocks to enhance contextual feature representation efficiently. Extensive experiments on the VisDrone2019-DET dataset demonstrate that the proposed model significantly outperforms baseline YOLOv11 variants and several state-of-the-art aerial object detectors. Our most accurate variant (LASD-YOLO-l), based on YOLOv11-l, achieves 55.9% mAP@0.5 and 35.8% mAP@0.5:0.95 with only 23.99M parameters. Even our smallest variant achieves superior accuracy and performance compared with heavier baseline models. To statistically quantify small-object detection performance, we propose a metric called minimum detectable object area (MDOA), which measures the mean and median area of the “n%” smallest correctly detected objects with respect to the input image area. Our model (“l” variant) successfully identifies objects that occupy only 0.0061% of the entire image on average, which is almost 36% smaller than the baseline model performance. We further validate the effectiveness of the proposed architectural components through additional ablation studies and Grad-CAM visualizations. To confirm generalization capability, we additionally evaluate our model on the UAVDT dataset without any training or transfer learning. The proposed architecture offers a balance between accuracy and computational efficiency, making it suitable for resource-constrained aerial object detection scenarios. The code is available at <a href="https://github.com/Riddhick/LASDYOLO">https://github.com/Riddhick/LASDYOLO</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LASD-YOLO: light-weight area attention based small object detector for UAV/aerial platforms

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Riddhick Dalal,
  • M. Swapna,
  • Debasish Pradhan,
  • J. V. Satyanarayana

摘要

Real-time small-object detection in aerial images and videos is a critical challenge due to scale variation, limited spatial resolution, and strict onboard computational constraints. We propose LASD-YOLO (lightweight area-attention-based small-object detector), a YOLOv11-based detector specifically tailored for real-time small-object detection in aerial images. The proposed architecture introduces a high-resolution detection head with a stride of 4, enabling the detection of very small objects that occupy 4 to 8 pixels in an image. To balance the increased computational cost, we replace the standard convolutional layers with depthwise convolution operations. Selected backbone stages are also augmented with area attention blocks to enhance contextual feature representation efficiently. Extensive experiments on the VisDrone2019-DET dataset demonstrate that the proposed model significantly outperforms baseline YOLOv11 variants and several state-of-the-art aerial object detectors. Our most accurate variant (LASD-YOLO-l), based on YOLOv11-l, achieves 55.9% mAP@0.5 and 35.8% mAP@0.5:0.95 with only 23.99M parameters. Even our smallest variant achieves superior accuracy and performance compared with heavier baseline models. To statistically quantify small-object detection performance, we propose a metric called minimum detectable object area (MDOA), which measures the mean and median area of the “n%” smallest correctly detected objects with respect to the input image area. Our model (“l” variant) successfully identifies objects that occupy only 0.0061% of the entire image on average, which is almost 36% smaller than the baseline model performance. We further validate the effectiveness of the proposed architectural components through additional ablation studies and Grad-CAM visualizations. To confirm generalization capability, we additionally evaluate our model on the UAVDT dataset without any training or transfer learning. The proposed architecture offers a balance between accuracy and computational efficiency, making it suitable for resource-constrained aerial object detection scenarios. The code is available at https://github.com/Riddhick/LASDYOLO.