机电工程技术2025，Vol.54Issue(8)：28-32,69,6.DOI:10.3969/j.issn.1009-9492.2024.00086
基于深度可分离与重参数化的轻量化目标检测模型
A Lightweight Object Detection Model Based on Depth Separability and Reparameterization
摘要
Abstract
In order to solve the problem of lightweight deployment of the railroad track fastener detection model,taking the YOLOX-s model as a benchmark,a feature pyramid based on depth-separable convolution is proposed,the lightweight feature fusion of the neck network structure is realized,so as to improve the feature extraction capability of the model after lightweighting.Aiming at the greater number of parameters and computational overhead generated by the decoupling head in the YOLOX architecture,the depth separable convolution is introduced to simplify the number of parameters of the decoupling head and further using the heavily parameterized convolution to reach the accuracy optimization of the detection head,so as to achieve the balance of accuracy loss caused by the lightweight of the model parameters.Experiments on the detection of railroad track fasteners show that the method reduces the number of parameters by 22.26%and the floating-point computation by 31.3%compared with the baseline model,and the accuracy is maintained at the original level.The method reduces the dependence of the inspection model on the storage space of the inspection equipment.The lightweight target detection model based on depth separability and reparameterization effectively reduces the model parameters and provides a feasible method for the deployment of the lightweight model on lightweight equipment for on-line railroad inspection.关键词
模型轻量化/铁路轨道扣件检测/深度可分离卷积/模型重参数化Key words
model lightweighting/railway track fastener detection/deep-wise convolution/model re-parameterization分类
计算机与自动化引用本文复制引用
林初欣,陈姜男,鄢化彪,邓亚峰,肖文祥..基于深度可分离与重参数化的轻量化目标检测模型[J].机电工程技术,2025,54(8):28-32,69,6.基金项目
江西省研究生创新专项资金项目(YC2022-S692) （YC2022-S692）
