Figure 1: Image of pixel coordinate
Filter name to use |
Data Status |
Kernel Filter size |
|||
2 x 2 |
3 x 3 |
4 x 4 |
5 x 5 |
||
|
Input Data |
28 x 28 |
28 x 28 |
28 x 28 |
28 x 28 |
Convolutional Filter1 |
Convolutional Layer1 |
27 x 27 |
26 x 26 |
25 x 25 |
24 x 24 |
Convolutional Filter2 |
Convolutional Layer2 |
26 x 26 |
24 x 24 |
22 x 22 |
20 x 20 |
Pooling Filter1(2x2) |
Pooling Layer1 |
13 x 13 |
12 x 12 |
11 x 11 |
10 x 10 |
Convolutional Filter3 |
Convolutional Layer3 |
12 x 12 |
10 x 10 |
8 x 8 |
6 x 6 |
Convolutional Filter4 |
Convolutional Layer4 |
11 x 11 |
8 x 8 |
5 x 5 |
2 x 2 |
Pooling Filter 2 (2x2) |
Pooling Layer2 |
5 x 5 |
4 x 4 |
2 x 2 |
1 x 1 |
Flatten Layer |
5 x 5 x Number of Kernel Filters |
4 x 4 x Number of Kernel Filters |
2 x 2 x Number of Kernel Filters |
1 x 1 x Number of Kernel Filters |
Table 1: Size of data generated when changing the size of a kernel filter
Therefore |
Variable |
Healthy (n=30) |
Diabetic (n=30) |
DFU (n=30) |
P-Value Interaction |
Vertical components of the earth's reaction force |
Frequency with power of 99.5% |
9.60 ± 0.45 |
8.33 ± 0.50 |
6.40 ± 0.27 |
0.000 * |
Essential number of harmonics |
20.40 ± 0.98 |
22.53 ± 1.23 |
20.56 ± 0.87 |
0.281 |
|
Medium frequency |
2.26 ± 0.09 |
2.20 ± 0.07 |
2.13 ± 0.06 |
0.490 |
|
Frequency Band width |
1.26 ± 0.95 |
1.13± 0.06 |
1.13 ± 0.06 |
0.356 |
|
toe |
Frequency with power of 99.5% |
8.90 ± 1.03 |
8.53 ± 0.80 |
15.96 ± 1.35 |
0.000 * |
Essential number of harmonics |
15.26 ± 0.92 |
18.76 ± 1.17 |
22.40 ± 0.97 |
0.000 * |
|
Medium frequency |
2.80 ± 0.13 |
2.20 ± 0.07 |
2.26 ± 0.08 |
0.000 * |
|
Frequency Band width |
1.43 ± 0.14 |
1.20 ± 0.07 |
1.26 ± 0.08 |
0.266 |
|
Toes 2 to 5 feet |
Frequency with power of 99.5% |
8.13 ± 0.72 |
7.90 ± 0.71 |
12.43 ± 0.89 |
0.000 * |
Essential number of harmonics |
16.23 ± 1.02 |
19.86 ± 1.18 |
21.16 ± 0.87 |
0.003 * |
|
Medium frequency |
2.83 ± 0.13 |
2.20 ± 0.07 |
2.30 ± 0.09 |
0.000 * |
|
Frequency Band width |
1.53 ± 0.14 |
1.20 ± 0.07 |
1.26 ± 0.08 |
0.073 |
Figure 1: Image of pixel coordinate
Figure 2: Flow of convolutional processing
Figure 3: Product categories and sample images included in the Fashion-MNIST
Figure 4.1: Relationship between the number of clusters and SSE (16 kernel filters)
Figure 4.2: Relationship between the number of clusters and SSE (32 kernel filters)
Figure 4.3: Relationship between the number of clusters and SSE (64 kernel filters)
Figure 4.4: Relationship between the number of clusters and SSE (128 kernel filters)
Figure 4.5: Relationship between the number of clusters and SSE (256 kernel filters)
Figure 5.1: Relationship between the number of clusters and log-likelihood chi-square (16 kernel filters)
Figure 5.2: Relationship between the number of clusters and log-likelihood chi-square (32 kernel filters)
Figure 5.3: Relationship between the number of clusters and log-likelihood chi-square (64 kernel filters)
Figure 5.4: Relationship between the number of clusters and log-likelihood chi-square (128 kernel filters)
Figure 5.5: Relationship between the number of clusters and log-likelihood chi-square (256 kernel filters)
Figure 6: Average value of feature vector finally extracted using 32 of 2 × 2 kernel filters
Figure 7: Number of clusters vs likelihood ratio chi-square
Tables at a glance
Figures at a glance