method: Conv_Attention2019-06-02
Authors: Daehyun Nam
Description: Convolution-based feature extractor followed by patch-based prediction with attention. Feature etractor is pre-trained with ImageNet. Roll-back method was applied to enhance low-level features. CE loss with label smoothing.
Confusion Matrix
Detection | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|
Arabic | Latin | Chinese | Japanese | Korean | Bangla | Hindi | Symbols | None | ||
GT | Arabic | 4783 | 156 | 20 | 49 | 30 | 37 | 38 | 29 | 0 |
Latin | 324 | 53107 | 865 | 2846 | 930 | 435 | 614 | 1516 | 0 | |
Chinese | 4 | 107 | 4119 | 463 | 30 | 11 | 6 | 10 | 0 | |
Japanese | 29 | 531 | 356 | 6914 | 138 | 38 | 35 | 116 | 0 | |
Korean | 23 | 699 | 188 | 403 | 11473 | 62 | 75 | 69 | 0 | |
Bangla | 2 | 43 | 5 | 12 | 5 | 2447 | 27 | 4 | 0 | |
Hindi | 5 | 65 | 5 | 14 | 4 | 86 | 4024 | 21 | 0 | |
Symbols | 18 | 172 | 19 | 41 | 8 | 7 | 28 | 3722 | 0 | |
None | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |