一种面向移动端应用的实时目标检测算法

A real-time object detection algorithm for mobile application on a mobile terminal

针对目标检测算法部署在移动端存在内存消耗大、精度低等问题,在NanoDet模型的基础上提出一种引入改进注意力机制的轻量级目标检测网络。首先,设计通道双池化及空间双向拆分的注意力模块,在尽可能不增加计算消耗的同时加强网络对感兴趣区域的关注能力;其次,引入空洞卷积及Mish函数增加网络的感受野及特征判别能力,并缩减冗余的降采样单元结构以加快网络的实时性;最后,在MS COCO2017数据集及安卓设备上的实验验证可知,本文算法在少量模型参数下提高了检测准确率,并保证30帧/秒的移动端检测速度,效果优于YOLO系列等轻量级网络。实验结果表明,本文算法参数量较YOLO系列模型参数量更低,更适合移动端和嵌入式设备的实时目标检测场景。

Aiming at the problems of large memory consumption and low precision of object detection algorithm deployed on a mobile terminal,a lightweight object detection network with improved attention mechanism is proposed based on NanoDet model.Firstly,the attention module is designed for double pools on the channel and double splits in space,so as to enhance the network's ability to focus on the region of interest without increasing the computing consumption as much as possible.Secondly,dilated convolution and Mish function are introduced to increase the receptive field and feature discrimination ability of the network,and reduce redundant down-sampling unit structures to speed up the real-time performance of the network.Finally,experimental verification on MS COCO2017 data set and Android devices shows that the proposed algorithm can improve the detection accuracy under a few model parameters,and ensure the detection speed of 30 frames per second on mobile terminals.The effect is better than that of lightweight networks such as YOLO series,and it is more suitable for real-time target detection scenarios on mobile terminals and embedded devices.