Multiscale Feature Adaptive Fusion for Object Detection in Optical Remote Sensing Images
Hao Lv, Weixing Qian, Tianxiao Chen, Han Yang, Xuecheng Zhou
Abstract
So far, the accuracy of object detection in natural images has been continuously improved, and at the same time, object detection has received more and more attention in the field of remote sensing. However, different from natural images, remote sensing images have a large number of multi-scale objects and large, complex backgrounds. As the size of remote sensing images is generally large, small objects are easy to lose information after down-sampling. To deal with these problems, we propose a multi-scale feature adaptive fusion (MFAF) method. Specifically, we first use the multi-scale feature integration (FI) module and the spatial attention weight (SAW) module to construct the feature fusion module to achieve the adaptive fusion of multi-scale features. Then we add a detail enhancement (DE) module at the back of the backbone to enhance the quality of features in each scale. Next, we obtain the relationship between features of different channels by embedding a squeeze and excitation (SE) block to highlight the useful features. Finally, a cross stage partial (CSP) block is adopted to replace continuous convolutions to reduce the number of parameters and the loss of features. We implement our work in YOLOv4. In experiments, we evaluate our method on large-scale remote sensing image data sets HRRSD and DIOR, and respectively achieve improvements of 2.7% and 1.9% in AP50 compared with YOLOv4, successfully improving the multi-scale object detection performance.