Algorithm for Plug Seedling Classification Based on Improved Attention Mechanism Residual Network

Cong Wu; Zhiqiang Guo; Jie Yang

doi:10.3788/LOP202259.2210002

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 22 >Page 2210002 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 22, 2210002 (2022)

Algorithm for Plug Seedling Classification Based on Improved Attention Mechanism Residual Network

Cong Wu, Zhiqiang Guo, and Jie Yang^*

Author Affiliations

College of Information Engineering, Wuhan University of Technology, Wuhan 438300, Hubei , China

show less

DOI: 10.3788/LOP202259.2210002 Cite this Article Set citation alerts

Cong Wu, Zhiqiang Guo, Jie Yang. Algorithm for Plug Seedling Classification Based on Improved Attention Mechanism Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2210002 Copy Citation Text

show less

Fig. 1. Example of corrected disk image. (a) Original disc image; (b) corrected disc image

Download full size

Fig. 2. Segmentation result of plug image

Download full size

Fig. 3. Loss of plug picture segmentation and normal picture erasure processing. (a) Strong seedling of white eggplant with segmentation loss; (b) strong seedling of cauliflower with segmentation loss; (c) strong seedling of cucumber with segmentation loss; (d) strong seedling of capsicum with segmentation loss; (e) strong white eggplant seedling after erasing processing; (f) strong cauliflower seedling after erasing processing; (g) strong cucumber seedling after erasing processing; (h) strong capsicum seedling after erasing processing

Download full size

Fig. 4. Channel attention mechanism

Download full size

Fig. 5. Spatial attention module

Download full size

Fig. 6. Attention mechanism fusion residual module

Download full size

Fig. 7. Inserting CBAM module between ResNet convolution blocks

Download full size

Fig. 8. Accuracy and loss curves of different attention mechanisms in ResNet

Download full size

Fig. 9. Feature weight heat map of classified output image

Download full size

Plug data	BQ	HC	LJ	QZ	QC	XG	HG
Quantity /disc	10	29	10	29	15	9	11
Plug specification	12×6	12×6	12×6	12×6	12×6	16×8	12×6

Table 1. Images of plug seedlings collected

Plant_seed	BQ	HC	LJ	QZ	QC	XG	HG
Strong seedling	645	1888	400	2051	948	3395	503
Weak seedling	75	420	850	469	610	1165	550
K_X				3249
Total				17219

Table 2. Plant_seed dataset composition

Layer name	Output size	34-layer	50-layer	CBAM
conv1	$112 \times 112$	$7 \times 7,64, s t r i d e 2$
conv2_x	$56 \times 56$	$3 \times 3 m a x p o o l, s t r i d e 2$		$\sqrt$
conv2_x	$56 \times 56$	$[\begin{array}{l} 3 \times 3,64 \\ C B A M \\ 3 \times 3,64 \end{array}] \times 3$	$[\begin{array}{l} 1 \times 1,64 \\ C B A M \\ 3 \times 3,64 \\ C M A B \\ 1 \times 1,256 \end{array}] \times 3$	$\sqrt$
conv3_x	$28 \times 28$	$[\begin{array}{l} 3 \times 3,128 \\ C B A M \\ 3 \times 3,128 \end{array}] \times 4$	$[\begin{array}{l} 1 \times 1,128 \\ C B A M \\ 3 \times 3,128 \\ C M A B \\ 1 \times 1,512 \end{array}] \times 4$	$\sqrt$
conv4_x	$14 \times 14$	$[\begin{array}{l} 3 \times 3,256 \\ C B A M \\ 3 \times 3,256 \end{array}] \times 6$	$[\begin{array}{l} 1 \times 1,256 \\ C B A M \\ 3 \times 3,256 \\ C M A B \\ 1 \times 1,1024 \end{array}] \times 6$	$\sqrt$
conv5_x	$7 \times 7$	$[\begin{array}{l} 3 \times 3,512 \\ C B A M \\ 3 \times 3,512 \end{array}] \times 3$	$[\begin{array}{l} 1 \times 1,512 \\ C B A M \\ 3 \times 3,512 \\ C M A B \\ 1 \times 1,2048 \end{array}] \times 3$
	$1 \times 1$	Average pool，15-d fc，softmax
Gflops /MB		21.97	26.05

Table 3. Improved ResNet parameters

Architecture	Params /MB	Gflops	Accuracy /%
ResNet34	21.79	11.75	90.69±1.61
ResNet34+SE	21.95	11.75	92.04±1.36
ResNet34+SE+CBAM_basic	22.11	11.76	93.68±0.62
ResNet34+CBAM_basic	21.96	11.76	93.46±0.84
ResNet34+CBAM_conv	21.81	11.75	91.28±0.92
ResNet34+CBAM_basic_conv	21.97	11.76	93.80±0.80
ResNet50	25.56	13.16	91.58±1.82
ResNet50+SE	28.07	13.18	93.19±0.81
ResNet50+CBAM_bottle	25.87	13.18	92.32±0.98
ResNet50+CBAM_conv	25.73	13.16	92.23±0.97
ResNet50+CBAM_bottle_conv	26.05	13.19	94.89±0.61
ResNet50+SE+CBAM_bottle	28.39	13.20	95.42±0.92

Table 4. Comparison of CBAM module insertion methods

Architecture	BQW						BQS					HCW					HCS
Architecture	P			R			P			R		P			R		P		R
ResNet34	0.655			0.864			0.974			0.959		0.959			0.746		0.993		0.940
ResNet34+SE	0.905			0.864			0.975			1.000		0.814			0.833		0.974		0.935
ResNet34+CBAM_basic	0.947			0.818			0.975			0.995		0.852			0.913		0.980		0.963
ResNet34+CBAM_conv	0.889			0.727			0.974			0.984		0.869			0.738		0.973		0.943
ResNet34+CBAM_basic_conv	0.905			0.864			0.990			1.000		0.753			0.897		0.981		0.924
ResNet34+SE+CBAM_basic	1.000			0.864			0.985			0.990		0.915			0.770		0.93		0.979
Architecture		HGW			HGS			K_X			LJW			LJS				QZW
Architecture		P	R		P	R		P	R		P		R	P		R		P		R
ResNet34		0.934	0.945		0.993	0.940		0.914	0.943		0.915		0.973	0.980		0.808		0.933		0.700
ResNet34+SE		0.987	0.909		0.936	0.980		0.924	0.959		0.929		0.976	0.990		0.858		0.925		0.879
ResNet34+CBAM_basic		0.957	0.933		0.924	0.967		0.942	0.964		0.943		0.973	0.930		0.892		0.966		0.800
ResNet34+CBAM_conv		0.938	0.915		0.936	0.973		0.927	0.939		0.906		0.984	0.951		0.817		0.963		0.750
ResNet34+CBAM_basic_conv		0.952	0.97		0.966	0.953		0.956	0.962		0.955		0.988	0.991		0.900		0.873		0.979
ResNet34+SE+CBAM_basic		0.93	0.964		0.979	0.913		0.953	0.957		0.898		1.000	1.000		0.792		0.952		0.850
Architecture		QZS			QCW			QCS			XGW			XGS				AP		AR
Architecture		P	R		P	R		P	R		P		R	P		R		AP		AR
ResNet34		0.920	0.995		0.847	0.934		0.951	0.894		0.821		0.774	0.947		0.930		0.911		0.892
ResNet34+SE		0.965	0.977		0.854	0.896		0.924	0.894		0.918		0.736	0.923		0.973		0.929		0.911
ResNet34+CBAM_basic		0.956	0.992		0.896	0.945		0.960	0.930		0.895		0.734	0.934		0.966		0.937		0.919
ResNet34+CBAM_conv		0.940	0.995		0.882	0.896		0.927	0.937		0.784		0.736	0.919		0.944		0.919		0.885
ResNet34+CBAM_basic_conv		0.987	0.972		0.805	0.973		0.979	0.838		0.907		0.808	0.947		0.980		0.930		0.933
ResNet34+SE+CBAM_basic		0.963	0.985		0.909	0.874		0.918	0.940		0.877		0.857	0.964		0.961		0.945		0.913

Table 5. Accuracy and recall results of each model in Plant_seed

Architecture	Param /MB	Gflops	Accuracy /%
AlexNet^［17］	14.6	3.96	88.76±0.84
GoogleNet^［18］	5.99	10.14	87.10±1.30
RegNet_400mf^［19］	4.30	0.40	94.63±0.97
EfficientNet_B0^［20］	5.30	0.40	86.57±1.13
EfficientNetV2_S^［21］	24.00	8.80	94.85±0.82
ResNet34	21.79	11.75	90.69±1.61
ResNet34+CBAM_basic_conv	21.97	11.76	93.80±0.80
ResNet50	25.56	13.16	92.96±0.64
ResNet50+CBAM_bottle_conv	26.05	13.19	94.89±0.61

Table 6. Different network training results of Plant_seed dataset

Architecture	Error /%
ResNet34	7.67±0.54
ResNet50	6.55±0.36
Proposed（ResNet34）	5.96±0.49
Proposed（ResNet50）	5.04±0.37
Proposed（ResNet34+erasing）	4.23±0.21
Proposed（ResNet50+erasing）	4.14±0.13

Table 7. Comparison of model classification error rates

Cong Wu, Zhiqiang Guo, Jie Yang. Algorithm for Plug Seedling Classification Based on Improved Attention Mechanism Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2210002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information