LPGNet：基于多尺度特征的轻量化像素级抓取检测网络

LPGNet: A Lightweight Pixel-Level Grasp Detection Network Based on Multi-Scale Features

摘要: 机器人抓取技术在仓储物流、工业分拣及服务机器人等领域具有广泛的应用前景。现有像素级抓取检测网络大多依赖参数量较大的主干网络，往往面临计算复杂度高、推理速度慢、难以部署于实时平台等问题。为此本文提出一种轻量化像素级抓取检测网络(Lightweight Pixel-level Grasp Network, LPGNet)。该网络以MobileNet为主干结构，引入轻量化多尺度特征增强模块(MDM-Lite)，通过编码器–解码器架构实现抓取区域、角度及宽度的像素级联合预测。在公开抓取数据集上对所提出方法进行了系统的实验验证，结果表明该方法在保证抓取检测精度的同时，具有更优的实时性能和部署潜力。

Abstract: Robotic grasping has broad application prospects in warehousing and logistics, industrial sorting, and service robotics. Most existing pixel-level grasp detection networks rely on backbones with a large number of parameters, and thus often suffer from high computational complexity, slow inference speed, and limited suitability for deployment on real-time platforms. To address these issues, this paper proposes a lightweight pixel-level grasp detection network, termed the Lightweight Pixel-level Grasp Network (LPGNet). LPGNet adopts MobileNet as the backbone and introduces a lightweight multi-scale feature enhancement module (MDM-Lite). With an encoder–decoder architecture, the network performs pixel-wise joint prediction of grasp regions, grasp angles, and grasp widths. Extensive experiments on public grasping datasets demonstrate that the proposed method maintains competitive grasp detection accuracy while achieving better real-time performance and deployment potential.