Processing Steps and Statistics¶
After encoding the sequences, we obtained size 4x4, 8x8, 16x16, 32x32 and 64x64 for each sequence.
Step 1 Merging¶
Then we do the merging. i.e.,
- Merge the data of block size 4x4 from four sequences together.
- Merge the data of block size 8x8 from four sequences together.
- Merge the data of block size 16x16 from four sequences together.
- Merge the data of block size 32x32 from four sequences together.
- Merge the data of block size 64x64 from four sequences together.
After merging, we obtained five csv files:
# Name of the Files Size Samples Usage 1 size_04.csv 206.7 MB 3675428 train/test/validation 2 size_08.csv 513.6 MB 2372324 train/test/validation 3 size_16.csv 1.25 GB 1439773 train/test/validation 4 size_32.csv 2.02 GB 567554 train/test/validation 5 size_64.csv 1.85 GB 125141 train/test/validation
Mode distribution data after merging are provided for downloading as text format.
mode distribution of block size 04x04
mode distribution of block size 08x08
mode distribution of block size 16x16
Step 2 Removing some modes¶
Remove mode 0, 1, 34, 35, 36. We only do deep learning for angular modes. Mode 34 is removed because mode 34 has the same direction feature as mode 2.
After removing 0, 1, 34, 35, 36:
# Name of the Files Size Samples Usage 1 m_size_04.csv 75.7 MB 1335970 train/test/validation 2 m_size_08.csv 130.8 MB 600187 train/test/validation 3 m_size_16.csv 377.3 MB 430302 train/test/validation 4 m_size_32.csv 708.7 GB 195943 train/test/validation 5 m_size_64.csv 1.37 GB 92034 train/test/validation
Percentage of non-angular-removed data:
size 4: (3675428 - 1335970) / 3675428.0 = 0.64
size 8: (2372324 - 600187) / 2372324.0 = 0.74
size 16: (1439773 - 430302) / 1439773.0 = 0.70
size 32: (567554 - 195943) / 567554.0 = 0.65
size 64: (125141 - 92034) / 125141.0 = 0.26
The mode distribution data are provided for downloading.
mode distribution of block size 04x04 AFTER non-angular removing
mode distribution of block size 08x08 AFTER non-angular removing
mode distribution of block size 16x16 AFTER non-angular removing
mode distribution of block size 32x32 AFTER non-angular removing
mode distribution of block size 64x64 AFTER non-angular removing
Step 3 Removing Smooth Blocks¶
Perform Edge Strength Analysis for each block sample of all sizes. Observing the histogram distribution.
Flat regions will trap CNN into ill condition. I decided to remove the regions where the edge strength is under 50.
And for the blocks where the edge strength is above 25000, we only consider four modes: VER, HOR, Wedgelet, Contour.
After removing the smooth areas,
# Name of the Files Size Samples Usage 1 sm_size_04.csv 36.3 MB 616281 train/test/validation 2 sm_size_08.csv 91.2 MB 403277 train/test/validation 3 sm_size_16.csv 210.9 MB 232806 train/test/validation 4 sm_size_32.csv 235.4 MB 65481 train/test/validation 5 sm_size_64.csv 271.8 MB 19244 train/test/validation
Percentage of smooth-removed data:
size 4: (1335970 - 616281) / 1335970.0 = 0.54
size 8: (600187 - 403277) / 600187.0 = 0.33
size 16: (430302 - 232806) / 430302.0 = 0.46
size 32: (195943 - 65481) / 195943.0 = 0.67
size 64: (92034 - 19244) / 92034.0 = 0.79
The mode distribution data are provided for downloading.
mode distribution of block size 04x04 AFTER smooth removing
mode distribution of block size 08x08 AFTER smooth removing
mode distribution of block size 16x16 AFTER smooth removing
Step 4 Imbalanced Learning¶
Please notice we are facing a problem of imbalanced learning which means the data sizes of each class vary in a large scale. To tackle the issue: we use equal data sizes for each class, abandon the extra data.
Step 5 Tagging¶
Tag the mode to start from 0, end with 31. Just use modeIdx - 2 to obtain
the new index of each mode for deep learning.
Final Data Description¶
# Name of the Files Size Samples Usage 1 train_04x04.csv 7.9 MB 4092*32 train 2 test_04x04.csv 1.1 MB 600*32 test 3 val_04x04.csv 1.1 MB 600*32 validation
# Name of the Files Size Samples Usage 1 train_08x08.csv 46.3 MB 6303*32 train 2 test_08x08.csv 3.8 MB 600*32 test 3 val_08x08.csv 3.8 MB 600*32 validation
# Name of the Files Size Samples Usage 1 train_16x16.csv 89.9 MB 3075*32 train 2 test_16x16.csv 15.9 MB 600*32 test 3 val_16x16.csv 15.6 MB 600*32 validation
- After trying to train data of size 04x04, it does not learn well, only top-28 accuracy is around 0.95.
- Too few data for size 32x32 and 64x64 after smooth removing. We resize them using Bilinear Interpolation for employing learned model for size 16x16