Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DASSF: Dynamic-Attention Scale-Sequence Fusion for Aerial Object Detection

Haodong Li, Haicheng Qu

TL;DR

A dynamic-attention scale-sequence fusion algorithm (DASSF) for small target detection in aerial images and a dynamic scale sequence feature fusion (DSSFF) module that improves the up-sampling mechanism and reduces computational load is proposed.

Abstract

The detection of small objects in aerial images is a fundamental task in the field of computer vision. Moving objects in aerial photography have problems such as different shapes and sizes, dense overlap, occlusion by the background, and object blur, however, the original YOLO algorithm has low overall detection accuracy due to its weak ability to perceive targets of different scales. In order to improve the detection accuracy of densely overlapping small targets and fuzzy targets, this paper proposes a dynamic-attention scale-sequence fusion algorithm (DASSF) for small target detection in aerial images. First, we propose a dynamic scale sequence feature fusion (DSSFF) module that improves the up-sampling mechanism and reduces computational load. Secondly, a x-small object detection head is specially added to enhance the detection capability of small targets. Finally, in order to improve the expressive ability of targets of different types and sizes, we use the dynamic head (DyHead). The model we proposed solves the problem of small target detection in aerial images and can be applied to multiple different versions of the YOLO algorithm, which is universal. Experimental results show that when the DASSF method is applied to YOLOv8, compared to YOLOv8n, on the VisDrone-2019 and DIOR datasets, the model shows an increase of 9.2% and 2.4% in the mean average precision (mAP), respectively, and outperforms the current mainstream methods.

DASSF: Dynamic-Attention Scale-Sequence Fusion for Aerial Object Detection

TL;DR

A dynamic-attention scale-sequence fusion algorithm (DASSF) for small target detection in aerial images and a dynamic scale sequence feature fusion (DSSFF) module that improves the up-sampling mechanism and reduces computational load is proposed.

Abstract

The detection of small objects in aerial images is a fundamental task in the field of computer vision. Moving objects in aerial photography have problems such as different shapes and sizes, dense overlap, occlusion by the background, and object blur, however, the original YOLO algorithm has low overall detection accuracy due to its weak ability to perceive targets of different scales. In order to improve the detection accuracy of densely overlapping small targets and fuzzy targets, this paper proposes a dynamic-attention scale-sequence fusion algorithm (DASSF) for small target detection in aerial images. First, we propose a dynamic scale sequence feature fusion (DSSFF) module that improves the up-sampling mechanism and reduces computational load. Secondly, a x-small object detection head is specially added to enhance the detection capability of small targets. Finally, in order to improve the expressive ability of targets of different types and sizes, we use the dynamic head (DyHead). The model we proposed solves the problem of small target detection in aerial images and can be applied to multiple different versions of the YOLO algorithm, which is universal. Experimental results show that when the DASSF method is applied to YOLOv8, compared to YOLOv8n, on the VisDrone-2019 and DIOR datasets, the model shows an increase of 9.2% and 2.4% in the mean average precision (mAP), respectively, and outperforms the current mainstream methods.
Paper Structure (19 sections, 4 equations, 6 figures, 4 tables, 1 algorithm)

This paper contains 19 sections, 4 equations, 6 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Proposed Method
  3. The Overall Architecture of DASSF-YOLO
  4. Improvements to the Neck
  5. Triple Feature Encoding Module
  6. Dynamic Scale Sequence Feature Fusion Module
  7. Channel and Position Attention Mechanism
  8. Improvements to the Head
  9. X-small Head
  10. Dynamic Head
  11. Experiment and Analysis
  12. Datasets
  13. Implementation Details and Evaluation Metrics
  14. Comparison with State-of-the-art Methods
  15. Comparisons on VisDrone-2019
  16. ...and 4 more sections

Figures (6)

  • Figure 1: The overall architecture of DASSF. DSSFF is dynamic scale sequence feature fusion module, TFE is triple feature encoding module, CPAM is channel and position attention mechanism, and DyHead is dynamic detection head.
  • Figure 2: The structure of the DySample. S represents the upsampling ratio, G represents the grid sampling point coordinates, and O represents the point position offset generated by the dynamic sampling point generator. SH and SW represent respectively sampling height and width. $\text{g}\text{s}^2$ represents the number of channels of the feature map after passing through the linear layer.
  • Figure 3: The structure of the DSSFF. The features are extracted efficiently and accurately through dynamic upsampling, feature map stacking, and 3D convolution normalization activation operations. The detailed process of dynamic upsampling is shown in algorithm \ref{['alg:1']}.
  • Figure 4: (a) The size distribution of objects in VisDrone-2019 dataset; (b) The size distribution of objects in DIOR dataset; (c) The category distribution of objects in VisDrone-2019 dataset; (d) The category distribution of objects in DIOR dataset.
  • Figure 5: Results of DASSF method in different versions of YOLO model.
  • ...and 1 more figures