# UAV Small Target Detection Model Based on Dual Branches and Adaptive Feature Fusion

## TL;DR

This paper introduces a new drone-based model for detecting small targets in aerial images using dual branches and adaptive feature fusion to improve accuracy and reduce parameters.

## Contribution

The novel approach combines dual branches, semantic-detail injection, and a large separable kernel attention mechanism for enhanced small target detection.

## Key findings

- The model achieved a 20.9% increase in mAP50 compared to the original model.
- It reduced the total number of parameters by 61.3%, making it more suitable for drones.
- The mAP50–95 improved by 23.7%, showing better performance across different target scales.

## Abstract

In order to solve the problem of small and dense targets in drone aerial images, a small target detection model based on dual branches and adaptive feature fusion is proposed. The model first constructs a small target detection framework with dual branches to improve the detection accuracy while reducing the number of parameters. Secondly, the model introduces semantic and detail injection (SDI) in the neck network and embeds bidirectional adaptive feature fusion in the detection head to innovate and optimize the feature fusion mechanism, achieve the full interaction of deep and shallow information, enhance the feature representation of small targets, and overcome the problem of scale inconsistency. Finally, in order to focus on the target area more accurately, we introduce the large separable kernel attention mechanism into the convolutional layer to provide it with a richer and more comprehensive feature representation, which significantly improves the detection accuracy of targets of different scales. The experimental results show that the model algorithm performs well in the VisDrone2019 dataset. Compared with the original model, the mAP50 of this model increases by 20.9%, the mAP50–95 increases by 23.7%, and the total number of parameters decreases by 61.3%, making it more suitable for drones.