面向移动端的单阶段端到端目标检测压缩算法

近30年间，深度学习异军突起。它在各项计算机视觉任务中都取得了令人瞩目的进步，加之大量高质多样化数据的出现，使得各种依赖数据的目标检测方法重现曙光。然而，这些深度网络算法通常需要大量数据来支持数百亿参数的计算，其运行效率较低并且对存储空间的要求越来越高，使得在小型设备或移动端中无法嵌入大型神经网络。因此，本文提出优化目标检测算法以适应移动端环境，利用CNN卷积核多样性和可分离的原理，应用深度可分离卷积（Depthwise Separable Convolution）结构的理论，提出单阶段-端到端目标检测压缩网络DW-YOLOv3。最后，在带有详细标注的地面观测实况大规模基准数据集VisDrone2018数据集上的结果表明，本文提出的改进单阶段-可分离卷积目标检测压缩网络算法可以将网络参数压缩8-9倍，由于其增加了整体网络的深度，在对网络整体性能影响较小的同时提升了对无人机视角图像中小目标物体的识别性能。 

One-Stage End-to-End Object Detection Compression Algorithm For Mobile Terminal

Over the past 30 years, deep learning has sprung up. It has made remarkable progress in various computer vision tasks, coupled with the emergence of a large number of high-quality and diversified data, making a variety of data-dependent target detection methods reappear in the dawn. However, these deep network algorithms usually require a large amount of data to participate in the calculation of tens of billions of parameters, which is inefficient and requires higher and higher storage space, making it impossible to embed such a large network in small devices or mobile terminals. Therefore, this paper proposes an optimized target detection algorithm to adapt to the mobile environment. Based on the diversity and separability of CNN convolution kernels and the theory of Depthwise Separable Convolution structure, a One-Stage End-to-End object detection compression network DW-YOLOv3 is proposed. Finally, the results on VisDrone 2018 data set with detailed annotation show that the improved one-stage separable convolutional target detection compression network algorithm proposed in this paper can compress the network parameters 8-9 times, and improve the recognition performance of small target objects while having little impact on the overall performance of the network.