Multi-Scale Feature Fusion Algorithm for Small Object Detection in Aerial Images

Aiming at the problems of blurred multi-scale features and dense distribution of targets in aerial images, which lead to low detection accuracy of small targets, an improved lightweight aerial detection algorithm that fuses multi-scale features is proposed. By constructing information interaction among features at different levels, the utilization efficiency of shallow features is improved. Based on the feature enhancement module of partial convolution, the multi-scale cross-fusion path is optimized, which helps to retain more original features, and the low-frequency enhancement filter improves the ability of the deep network in the model to capture low-frequency features. The experimental results show that compared with the baseline model, the improved model has increases of 3.9 and $5.7 \mathrm{in} \mathrm{mAP0.5} / \%$ and $\mathrm{mAP} 0.5: 0.95 / \%$ respectively, and the number of parameters and the amount of computation are reduced by $\mathbf{2 8. 3 4 \%}$ and $\mathbf{1 7. 0 7 \%}$ respectively.