Segmentation Method of Broken Ore Image Based on Improved HED Network Model

Qinghua Gu; Fawen Wei; Mengli Guo; Song Jiang; Shunling Ruan

doi:10.3788/LOP202259.0210020

Laser & Optoelectronics Progress, Volume. 59, Issue 2, 0210020(2022)

Segmentation Method of Broken Ore Image Based on Improved HED Network Model

Qinghua Gu^1,2、*, Fawen Wei^1,2, Mengli Guo^1,2, Song Jiang^1,2, and Shunling Ruan^1,2

Author Affiliations

¹School of Resource Engineering, Xi'an University of Architecture and Technology, Xi'an , Shaanxi 710055, China

²Xi'an Key Laboratory of Smart Industry Perception Computing and Decision Making, Xi'an , Shaanxi 710055, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Structure of VGG network

Download full size

Fig. 2. Structure of deformable convolution module

Download full size

Fig. 3. Schematic of convolution sampling points. (a) Standard convolution; (b) deformable convolution

Download full size

Fig. 4. Expansion results of convolution kernel of 3×3 size under different cavity rates. (a) l=1; (b) l=2; (c) l=3

Download full size

Fig. 5. Framework structure of proposed network

Download full size

Fig. 6. Images before and after bilateral filtering. (a) Original image of ore; (b) image filtered by 3×3 filtering window

Download full size

Fig. 7. Relationship among loss, accuracy, and iterations of improved HED network model. (a) Loss; (b) accuracy

Download full size

Fig. 8. Segmentation results of different network models. (a) Original images; (b) Canny operator; (c) original HED network; (d) improved HED network

Download full size

Table 1. Training parameters of model
View table
Table 1. Training parameters of model
Parameter Value
Number of calculations 300
Iterations 100
Number of images per iteration 4
Number of training samples 1000

Table 2. Performance indicators of different networks

View table

Table 2. Performance indicators of different networks

Image	Accuracy			Recall			Precision
Image	Canny	HED	Improved HED	Canny	HED	Improved HED	Canny	HED	Improved HED
1	0.7283	0.8172	0.9183	0.8162	0.8438	0.8275	0.8754	0.9219	0.9428
2	0.7882	0.8635	0.9117	0.8225	0.8035	0.8173	0.8829	0.9178	0.9537
3	0.7937	0.8879	0.9225	0.8156	0.8167	0.8158	0.8478	0.9378	0.9389
4	0.7014	0.8918	0.9308	0.8068	0.8124	0.8067	0.8513	0.9247	0.9409
5	0.7691	0.8815	0.9126	0.7995	0.8037	0.8117	0.8218	0.9345	0.9128
6	0.7437	0.8857	0.9215	0.8093	0.7969	0.8027	0.8145	0.9362	0.9390
7	0.7659	0.8972	0.9118	0.8109	0.8089	0.8164	0.8047	0.9419	0.9268
8	0.7573	0.8027	0.9169	0.8038	0.8129	0.8058	0.8439	0.9370	0.9503
9	0.7654	0.8896	0.9027	0.7995	0.8097	0.8187	0.8057	0.9105	0.9218
10	0.7715	0.8889	0.9047	0.8092	0.8126	0.8213	0.8120	0.9279	0.9348
Average	0.7585	0.8706	0.9154	0.8093	0.8121	0.8144	0.8360	0.9290	0.9362

Table 3. Time comparison results of different networks
View table
Table 3. Time comparison results of different networks
Model MIOU /% Time /ms
Canny 72.62 128
HED 76.49 98
Improved HED 77.23 81

Table 4. Ablation experiment results of different network models
View table
Table 4. Ablation experiment results of different network models
Experiment DCN DC AP /%
1 × × 85.47
2 √ × 90.23
3 √ √ 92.37

Tools

Get Citation

Copy Citation Text

Qinghua Gu, Fawen Wei, Mengli Guo, Song Jiang, Shunling Ruan. Segmentation Method of Broken Ore Image Based on Improved HED Network Model[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0210020

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites