Attentional single-shot network with multi-scale feature fusion for object detection in aerial images

In recent years, UAV (Unmanned Aerial Vehicles) are widely used in military reconnaissance and traffic control scenarios. Due to the rapid development of UAV technology, the object detection technology for aerial images has drawn the most attention in computer vision. However, detection in aerial images is a non-trivial task due to low target resolution, large scale variation, and occlusion and illumination variation. Our method is derived from the SSD (Single Shot Detector) which is well- balanced between detection speed and accuracy. However, in SSD network, not all feature map channels contain useful information for detection. Moreover, the SSD network use feature map of different scales for predicting objects with different sizes which is not reasonable. Aiming at overcoming these two shortcomings, we introduced the mechanism of channel-wise attention which can help choose channels that contain the most discriminative information. We also propose the multi-scale feature fusion mechanism in which high and low- level feature maps are fused which boost the dis-criminative power of the feature maps for prediction typically for small objects in aerial image. Experimental results shows that our method has remarkably improved the performance in object detection in aerial images.