Coal and Gangue Recognition Method Based on Dual-Channel Pseudocolor Image by Lidar

Fig. 8. Point cloud image generation process. (a) Original point cloud map; (b) vertical view of point cloud; (c) real point cloud projection; (d) two dimension intensity pixel image

Download full size

Fig. 9. Dual channel pseudo color image generation process. (a) Intensity point cloud; (b) distance point cloud; (c) dual channel pseudo color image

Download full size

Fig. 10. DenseNet-121 network structure

Download full size

Fig. 11. Dual channel pseudo color image used in the experiment. (a) Coal; (b) gangue

Download full size

Fig. 12. Training accuracy and training loss curves. (a) (d) DenseNet-40; (b) (e) DenseNet-121; (c) (f) DenseNet-201

Download full size

Table 1. Parameters of LMS4000
View table
Table 1. Parameters of LMS4000
Parameter Value
Scanning frequency /Hz 600
Laser wavelength /nm 660
Scanning range /（°） 70
Angular resolution /（°） 0.0833
Working distance /m 0.7‒3
Measurement accuracy /mm ±1

Table 2. Parameters of DenseNet-40 and DenseN-121

View table

Table 2. Parameters of DenseNet-40 and DenseN-121

Layer	Output size	DenseNet-40	DenseNet-121
Convolution	112×112	7×7Conv，s=2	7×7Conv，s=2
Pooling	56×56	3×3max pool，s=2	3×3max pool，s=2
Dense block（1）	56×56	$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 6$	$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 6$
Transition layer（1）	56×56	1×1Conv	1×1Conv
Transition layer（1）	28×28	2×2average pool，s=2	2×2average pool，s=2
Dense block（2）	28×28	$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 6$	$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 12$
Transition layer（2）	28×28	1×1Conv	1×1Conv
Transition layer（2）	14×14	2×2average pool，s=2	2×2average pool，s=2
Dense block（3）	14×14	$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 6$	$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 24$
Transition layer（3）	14×14		1×1Conv
Transition layer（3）	7×7		2×2average pool，s=2
Dense block（4）	7×7		$[\binom{1 \times 1 C o n v}{3 \times 3 C o n v}] \times 16$
Classification layer	1×1	7×7global average pool	7×7global average pool
Classification layer		1000D fully-connecter，softmax	1000D fully-connecter，softmax

Table 3. FLOPs and parameter quantity of common models
View table
Table 3. FLOPs and parameter quantity of common models
Model Size /10⁶ FLOPs /10⁹
AlexNet 61.0 0.7
GoogleNet 7.0 1.6
VGG-19 144.0 15.5
ResNet50 25.0 3.9
DenseNet40 0.2 1.3
DenseNet121 8.0 5.7

Table 4. Cross validation average results
View table
Table 4. Cross validation average results
Model Accuracy /% P R F₁
DenseNet-40 94.56 0.96 0.97 0.97
DenseNet-121 93.49 0.91 0.97 0.94
DenseNet-201 92.74 0.90 0.96 0.93
ResNet-101 83.53 0.82 0.86 0.84
VGG-19 92.54 0.90 0.95 0.93
Xception 90.24 0.89 0.92 0.90

Tools

Get Citation

Copy Citation Text

Yan Wang, Jichuan Xing, Yaozhi Wang. Coal and Gangue Recognition Method Based on Dual-Channel Pseudocolor Image by Lidar[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0437005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites