Detecting the objects of interesting from aerial images captured by UAVs is one of the core modules in the UAV-based applications. However, it is very difficult to detection objects from aerial images. The reason is that the scale of objects in the aerial images captured by UAVs varies greatly and needs to meet certain real-time performance in detection. To deal with these challenges, we proposed a lightweight model named DSYolov3. We made the following improvements to the Yolov3 model: 1) multiple scale-aware decision discrimination network to detect objects in different scales, 2) a multi-scale fusion-based channel attention model to exploit the channel-wise information complementation, 3) a sparsity-based channel pruning to compress the model. Extensive experimental evaluation has demonstrated the effectiveness and efficiency of our approach. By the proposed approach, we could not only achieve better performance than most existing detectors but also ensure the models practicable on the UAVs.

从无人机捕获的空中图像中检测出有趣的物体是基于无人机的应用程序中的核心模块之一。但是，从航空图像中检测物体非常困难。原因是无人机捕获的航拍图像中物体的比例变化很大，需要满足一定的实时检测性能。为了应对这些挑战，我们提出了一个名为DSYolov3的轻量级模型。我们对Yolov3模型进行了以下改进：1）多尺度感知决策鉴别网络，用于检测不同尺度的对象； 2）基于多尺度融合的渠道注意模型，用于利用渠道方面的信息补充； 3）a基于稀疏性的通道修剪以压缩模型。广泛的实验评估已经证明了我们方法的有效性和效率。