Table 4.

The comparison of the encoding ability in different depth of the proposed CNN.

Layer	Type	Discarding Accuracy (%)	Feature Length	Time (ms)
1	Input	85.61 ± 2.43	21,600	21.00
2	Attention	85.35 ± 2.61	21,600	20.76
3	Convolution	77.65 ± 2.86	118,800	104.56
4	BN	72.37 ± 3.24	118,800	104.09
5	ReLu	72.19 ± 3.18	118,800	104.09
6	MP	75.32 ± 2.92	29,700	27.57
7	Attention	73.36 ± 3.21	29,700	28.48
8	Convolution	71.37 ± 3.17	59,400	52.91
9	BN	70.82 ± 3.23	59,400	53.02
10	ReLu	76.00 ± 3.26	59,400	53.74
11	MP	83.58 ± 2.68	15,840	15.67
12	Convolution	90.79 ± 1.30	36,720	33.68
13	BN	91.70 ± 0.95	36,720	33.37
14	ReLu	89.48 ± 1.83	36,720	33.28
15	MP	95.88 ± 0.86	18,360	18.31
16	Convolution	96.52 ± 0.70	10,080	10.99
17	BN	96.55 ± 0.68	10,080	10.82
18	ReLu	96.27 ± 1.09	10,080	10.90
19	MP	95.42 ± 1.33	3360	5.00
20	Convolution	80.92 ± 2.66	960	2.25
21	BN	84.39 ± 3.07	960	2.25
22	ReLu	90.23 ± 1.92	960	2.41
23	Convolution	77.75 ± 2.57	8160	9.08
24	BN	77.48 ± 2.22	8160	9.10
25	ReLu	72.51 ± 2.50	8160	9.04
26	GAP	26.79 ± 1.75	68	1.33
27	FC	16.83 ± 1.23	1	1.05

Attention refers to the radial attention layer.