Transfer learning techniques for semantic segmentation of machine vision inspection and identification based on label-reserved Softmax algorithms

Guixiong LIU; Jian HUANG

doi:10.37188/OPE.20223001.0117

Optics and Precision Engineering, Volume. 30, Issue 1, 117(2022)

Transfer learning techniques for semantic segmentation of machine vision inspection and identification based on label-reserved Softmax algorithms

Guixiong LIU^* and Jian HUANG

Author Affiliations

School of Mechanical and Automotive Engineering， South China University of Technology， Guangzhou510640， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Fine-tuning transfer learning framework based on label-reserved improved Softmax algorithms

Download full size

View in Article

Fig. 2. Model of label-reserved improved Softmax algorithms

Download full size

View in Article

Fig. 3. Head architecture of label-reserved Mask R-CNN

Download full size

View in Article

Fig. 4. Comparison of different migration learning methods

Download full size

View in Article

Fig. 5. Semantic segmentation results of chassis assemblies

Download full size

View in Article

Fig. 6. Basic functions of MVAQ2 chassis standard parts assembly quality inspection and identification software

Download full size

View in Article

Fig. 7. MVAQ2 chassis standard parts assembly quality inspection and identification devices

Download full size

View in Article

Fig. 8. Chassis learning process of MVAQ2 devices

Download full size

View in Article

Table 1. T_train and L_CE of pre-trained Mask R-CNN model fine-tuning transfer learning initial various weights W_CNN

View table

View in Article

Table 1. T_train and L_CE of pre-trained Mask R-CNN model fine-tuning transfer learning initial various weights W_CNN

源数据集D_source	目标数据集D_target	模型结构调整	模型初始权值W_CNN	L_CE	T_train/min
ImageNet^［7］	iFixit^［18］	输出层	W_main	4.892	258
COCO^［8］	iFixit^［18］	输出层	W_main∪W_seg	1.542	115
iFixit^［18］	iFixit^［18］	—	W_main∪W_seg	1.297	85
iFixit^［18］	iFixit^［18］	—	W_main∪W_seg∪W_cla	0.019	58

Table 2. Parameter comparison of label-reserved and general Softmax algorithms

View table

View in Article

Table 2. Parameter comparison of label-reserved and general Softmax algorithms

参　数	典型Softmax算法^［17］	标签预留改进Softmax算法
输入特征向量	n_source×1维X	（n_source+n_reserved）×1维X'
分类器权值	n_source个n_source×1维 $W_{c l a} = {W_{1}, \dots, W_{n_{s o u r c e}}}$	（n_source+n_reserved）个（n_source+n_reserved）×1维 $W'_{c l a} = {W_{1}, \dots, W_{n_{s o u r c e} + n_{r e s e r v e d}}}$
输出概率向量	n_source×1维P	（n_source+n_reserved）×1维P'
标签集	$C'_{c l a} = C_{s o u r c e}$	$C'_{c l a} = C_{s o u r c e} ⋃ C_{r e v e r s e d}$

Table 3. Training time ofdifferent transfer learning methods

View table

View in Article

Table 3. Training time ofdifferent transfer learning methods

方　法	源数据集D_source	目标数据集D_target	模型结构调整	T_train/min
Fine-tuning^［17］	不可扩展机箱	可扩展机箱	输出层	42.8
特征提取Fine-tuning^［20］	不可扩展机箱	可扩展机箱	输出层	36.5
本文方法	不可扩展机箱	可扩展机箱	—	30.1

Tools

Get Citation

Copy Citation Text

Guixiong LIU, Jian HUANG. Transfer learning techniques for semantic segmentation of machine vision inspection and identification based on label-reserved Softmax algorithms[J]. Optics and Precision Engineering, 2022, 30(1): 117

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites