Towards Efficient Detection for Small Objects via Attention-Guided Detection Network and Data Augmentation

Xiaobin Wang, Dekang Zhu, Ye Yan2022 · DOI

TL;DR

This paper tackles the persistent challenge of detecting small objects in aerial imagery, a critical task for applications like drone surveillance and autonomous navigation. The authors propose a novel approach combining an enhanced YOLOv5 architecture with a multi-faceted data augmentation strategy, achieving significant improvements in small object detection accuracy.

What Problem Does This Paper Address?

Detecting small objects in images, especially those captured by drones (UAVs), is notoriously difficult. These images often contain objects that occupy very few pixels, making their features weak and easily lost during the downsampling stages of typical deep learning models. Furthermore, small objects tend to appear densely packed, exacerbating the detection challenge. Existing object detection algorithms, primarily designed for larger objects, struggle with these characteristics, leading to significantly lower accuracy for small targets. This paper aims to bridge this gap by developing a more effective method for small object detection in aerial imagery.

Methods

The proposed solution comprises two main components:

  • Data Augmentation Strategy:
    • Image Division: The original images are divided into smaller patches. The division is centered around the ground-truth bounding box of objects. This strategy effectively increases the number of small objects present in the training data and makes them more prominent.
  • Non-Maximum Suppression (NMS) on Divided Images: To mitigate the issue of overlapping patches created by the division process, NMS is applied to reduce redundant detections and computational load.
  • Image Pyramid Mechanism: The divided images are then upsampled to multiple resolutions (640x640, 700x700, 800x800). This "image pyramid" approach helps to enlarge small objects and sparsify dense object clusters, making them easier for the detector to identify.
  • Attention-Guided Detection Network (Multi-Head-Attention-Yolo):
    • Baseline: The method builds upon the YOLOv5 architecture, specifically YOLOv5x, which is known for its strong performance.
  • Additional Detection Head: A detection head is added to a lower-level feature map (P2) in addition to the standard heads on P3, P4, and P5. This allows the network to detect objects at finer resolutions, crucial for small objects.
  • Attention Mechanism: Swin Transformer blocks are integrated into each detection head, replacing some convolutional layers. This allows the network to focus on relevant regions of interest and leverage contextual information, further enhancing feature representation for small objects.

Datasets

The proposed method was evaluated on two widely used datasets for aerial imagery object detection:

  • Visdrone2019: This dataset contains aerial images with a focus on small and dense objects.
  • DOTA: This dataset consists of satellite remote sensing images, also featuring a significant number of small objects across various categories.

Key Insights and Ideas

  • Context is King for Small Objects: The paper strongly emphasizes that small objects, due to their limited pixels and features, heavily rely on contextual information for accurate detection. The attention mechanism and the addition of lower-level feature map detection heads directly address this.
  • Data Augmentation as a Necessity, Not a Luxury: The authors demonstrate that simple data augmentation techniques, like image division and upsampling, can significantly boost the performance of small object detection, making it more feasible to train robust models.
  • Multi-Scale Detection is Crucial: The image pyramid mechanism and the addition of multiple detection heads highlight the importance of processing objects at various scales to capture both small and larger objects effectively within the same framework.
  • Transformer Architectures Enhance Feature Representation: The integration of Swin Transformer blocks shows promise in improving the network's ability to capture global context and fine-grained features, which are vital for distinguishing small objects.

Key Results

The proposed Multi-Head-Attention-Yolo method demonstrated significant improvements on both Visdrone2019 and DOTA datasets:

  • Visdrone2019: The proposed method achieved an overall mAP of 41.8 and an APs (small objects) of 33.7, outperforming the baseline YOLOv5 (mAP 40.3, APs 32.5) and other state-of-the-art methods like Faster R-CNN and Swin Transformer Mask R-CNN.
  • DOTA: On the DOTA dataset, the method achieved an mAP of 50.7 and an APs of 37.8, again showing superiority over baseline YOLOv5 and other comparative algorithms.
  • Ablation Studies: The paper also presented ablation studies showing that the data augmentation strategy (image pyramid and division) improved YOLOv5's mAP by 3.2 and APs by 4.4. The addition of multiple detection heads and the attention mechanism further contributed to performance gains.

Limitations and Future Work

  • Computational Cost: While the proposed method improves accuracy, the paper acknowledges that the attention mechanism and multiple detection heads can increase computational complexity.
  • Future Directions: The authors express their intention to continue researching small object detection, with a focus on improving detection performance while simultaneously reducing computational overhead. This suggests a direction towards more efficient attention mechanisms or network architectures.

Summary updated

BibTeX
@misc{wang2022towards,
  author = {Xiaobin Wang and Dekang Zhu and Ye Yan},
  title = {{Towards Efficient Detection for Small Objects via Attention-Guided Detection Network and Data Augmentation}},
  year = {2022},
  doi = {10.3390/s22197663},
}
Figure