面向无人机场景的深度网络模型压缩与加速设计方法

Compression and Acceleration Design Methods of Deep Network Models for UAV Scenes

随着无人机应用日益广泛,无人机应用场景下的目标检测技术已在诸多领域有着重要的应用价值和迫切的应用需求。在无人机边缘设备实时检测的场景中,由于无人机高分辨率图像存在大量的微弱目标实例,将直接降低图像分辨率,从而造成微弱目标丢失,因此输入网络时保持高分辨率信息至关重要。为此,将高分辨率图像裁成多个图像块,将其作为网络输入,以保持网络精度的同时使其能在边缘智能设备上运行。同时,为了加快模型推理速度,使用分组卷积与通道混洗策略对检测算法的主干网络进行轻量化设计,并使用通道注意力机制提升网络精度。实验表明,所提方法相较于YOLOv5,在无人机数据集VisDrone与自建数据集OUC-UAV-DET上的精度分别提升2%、4%,在英伟达硬件(Xavier)上推理速度减少1 ms。在网络剪枝层面,结合具体数据集对检测模型进行通道剪枝,能在维持算法精度的前提下将推理速度减少2 ms。此外,对于无人机单类别检测任务,针对其目标实例尺寸相对固定的特点优化输出部分,能减少30%的模型参数量,使推理速度最多减少2 ms。

With the increasingly widespread application of drones,target detection technology in drone application scenarios has important ap⁃plication value and urgent application needs in many fields.In the scene of real-time detection of drone edge devices,due to the presence of a large number of weak target instances in the high-resolution images of drones,the image resolution will be directly reduced,resulting in the loss of weak targets.Therefore,it is crucial to maintain high-resolution information when inputting into the network.To achieve this,high-res⁃olution images are cut into multiple image blocks and used as network inputs to maintain network accuracy while enabling them to run on edge intelligent devices.At the same time,in order to accelerate the model inference speed,group convolution and channel shuffling strategies are used to lightweight design the backbone network of the detection algorithm,and channel attention mechanism is used to improve network accu⁃racy.The experiment shows that compared to YOLOv5,the proposed method improves accuracy by 2%and 4%on the unmanned aerial vehi⁃cle dataset VisDrone and the self built dataset OUC-UAV-DET,respectively,and reduces inference speed by 1 ms on the Nvidia hardware(Xavier).At the network pruning level,combining specific datasets for channel pruning of detection models can reduce inference speed by 2 ms while maintaining algorithm accuracy.In addition,for the single category detection task of unmanned aerial vehicles,optimizing the output part based on the relatively fixed size of the target instance can reduce the number of model parameters by 30%, resulting in a maximum reduc⁃ tion of 2 ms in inference speed.