Tiny Target Detection with Adaptive Weighted Feature Fusion in UAV Images

With the development of drones, imaging lenses and other equipment, effective and accurate detection of tiny targets in drone remote sensing images has become a hotspot. However, the features of tiny targets are weak and difficult to extract. Most existing methods enhance the target feature expression by directly superimposing and fusing multi-scale feature maps. Only the features with stronger responses in adjacent feature maps can be enhanced, while the targets with weaker responses may be diluted. For this reason, an adaptive weighted fusion small target detection algorithm is designed. To enhance the proportion of small targets in feature maps of different resolutions, a bidirectional feature pyramid network is designed. The adaptive weights are set according to the feature density in the feature map, and they will be used to perform bidirectional fusion on feature maps to reduce the interference of background and large targets on the features of small targets. Besides, we design multi-scale dilated convolution, which is injected into the feature fusion network. In this way, we successfully enhance the expression of small target feature. Finally, we test the algorithm via the public dataset VisDrone, the results show that our methods significantly improve the accuracy of small target detection.