Lightweight oriented object detection with Dynamic Smooth Feature Fusion Network

Abstract

In remote sensing visual tasks, detection of small objects remains a challenge due to low resolution, complexity and scale variations. In addition, many of these detection tasks need to be deployed to front-end low-resource devices on Unmanned Aerial Vehicle (UAV), which requires that detection method should be lightweight. This paper presents a lightweight oriented object detection method for small objects in remote sensing images, which is named Dynamic Smooth Feature Fusion Network (DSFF-Net). DynamicConv (DC) module is designed to enhance the depth of the network by stacking and continuously fusing small modules to capture contextual information of small objects at lower computational cost. Smooth Attention (SA) module is developed to incorporate attention mechanisms along spatial height and width directions of feature maps. The SA module enhances spatial feature extraction by generating attention maps that emphasize the selection of critical object features while suppressing background noise. It is worth mentioning that the proposed DC and SA modules can be integrated into many classical object detection frameworks and enhance the detection performance of remote sensing small objects consistently. Extensive experiments verify the effectiveness of the proposed DSFF-Net.