光学精密工程2025，Vol.33Issue(19)：3093-3105,13.DOI:10.37188/OPE.20253319.3093
基于SDFSN-HiFuse网络的减速器工件分类
Reducer workpiece classification based on SDFSN-HiFuse network
摘要
Abstract
Accurate classification of visually similar reducer parts is essential for precise assembly.Exist-ing visual classification methods struggle with highly similar parts due to limited discriminative features and low robustness to complex background interference,which can introduce errors in assembly.To address these challenges,a HiFuse-based Spatial Dual-Focus Synergy Network(SDFSN-HiFuse)is proposed for classification of reducer workpieces,targeting scenarios with large intra-class variance and small inter-class variance.A multi-branch spatially adaptive dilation-rate selection mechanism is introduced to enable auto-matic determination of appropriate receptive fields for deformed regions of workpieces.A two-stage geo-metric-local collaborative attention mechanism provides stepwise fine-grained guidance to features from each dilation branch,dynamically reweighting features and enhancing discrimination of salient regions via a coarse-to-fine refinement process.A deformable geometric graph is employed to model geometric topolo-gy flexibly,overcoming the constraints of traditional fixed grids.Following deformable convolution,a cur-vature gating mechanism preserves adaptive geometric deformation features,substantially improving re-sponsiveness and representation accuracy on complex curved surfaces.On a custom dataset,SDFSN-HiFuse achieves a 3.57%absolute improvement in accuracy and a 2.99%increase in precision over the baseline,while meeting real-time requirements with a processing rate of 300.39 frame/s.关键词
减速器工件分类/深度学习/注意力机制/多尺度膨胀卷积Key words
reducer workpiece classification/deep learning/attention mechanism/multi-scale dilated convolution分类
计算机与自动化引用本文复制引用
于智龙,张雪寒,齐丽华,杨佳欣,于广滨,李忠刚..基于SDFSN-HiFuse网络的减速器工件分类[J].光学精密工程,2025,33(19):3093-3105,13.基金项目
黑龙江省重点研发计划资助项目(No.2023ZX01A03) （No.2023ZX01A03）
