面向智能视觉货柜的轻量级商品目标检测方法

Lightweight commodity target detection method orienting to intelligent visual containers

针对目标检测方法参数量大、计算复杂度高以及对复杂目标出现误检和漏检等问题,本文提出了一种适用于嵌入式设备的轻量级商品目标检测方法。通过重构高效混洗轻量网络,大幅度降低了网络的参数量和计算复杂度;融合多重感知注意力,将通道和空间域混合并行考虑来突出重点特征,弥补网络可能造成的特征损失,提高对复杂目标的检测率;构建双级快速特征加权金字塔网络用于特征学习,结合Hard-swish可高效快速的进行多尺度特征融合,提升网络的表征能力。该方法在商品目标检测任务中的均值平均精度达98.6%,且参数量降低了约41.2%,与先进的检测方法相比有更高检测精度并且更轻量,能够实现高质量的商品实时检测。

In order to solve the problems of large amount of parameters,high computation complexity,false detection and missed detection of complex targets in the target detection methods,we propose a lightweight commodity target detection method suitable for embedded devices.By reconstructing a high-efficiency shuffle lightweight network,the parameter quantity and computational complexity of the network are greatly reduced.The multiple perceptual attention is integrated,and the channel and spatial domains are mixed and considered in parallel to highlight key features and make up for the feature loss that may be caused by the network,improving the detection rate of complex targets.Building a double-stage fast feature-weighted pyramid network for feature learning.The combination with Hard-swish can quickly perform multi-scale feature fusion with high efficiency,improving the representation ability of the network.The mean average precision of this method in commodity object detection task reaches 98.6%,and the parameter amount is reduced by about 41.2%.Compared with advanced detection methods,it has higher detection accuracy and lighter weight,and can achieve high-quality detection of commodities in real time.