Efficient Object Detection with Deformable Convolution for Optical Remote Sensing Imagery
Haonan Kang, Yundong Liu · 2022 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) · 2022
Deep learning-based object detection in remote sensing images is an important yet challenging task because of the complex background and large variations in size of the targets. Currently, many detectors have made significant progress in improving detection accuracy, but they have not shown good performance in terms of detection speed and model size. To address these issues, a lightweight and efficient object detector in remote sensing images is proposed. Specifically, we utilize MobileNetv3 with Shuffle Attention as the feature extraction backbone network to reduce the parameter of the model. Meanwhile, deformable convolution is introduced to adapt to the deformation of the target and obtain stronger geometric feature expression ability. Extensive experiments on two remote sensing public datasets (RSOD and DIOR) show good performance for the efficiency of our detector. In particular, our strategy achieves 71.3 mAP on the DIOR dataset with 3.05 parameters and a test speed of 12.1 ms, and it also has good performance on the RSOD dataset.