Link: http://drug.mfds.go.kr/html/index.jsp#
Image Crop/Resizing ( Final Image Size: 1024 × 512 × 3 )
Label Naming
Model: Baisc ResNet18
Augmentation: Vertical Flip / Tilt / Scaling / Shear / Elastic Distortion
1. Shape (11 Types)
Validation Acc: 91.87%
2. Front Color (16 Types)
Validation Acc: 90.23%
3. Back Color (16 Types)
Validation Acc: 91.08%
________________________________________________________________________________________________________________
Total Acc: 75.50%
(Each model was trained with 50 epochs)
1. Shape (11 Types)
Validation Acc: 91.55%
2. Front Color (16 Types)
Validation Acc: 91.70%
3. Back Color (16 Types)
Validation Acc: 90.97%
________________________________________________________________________________________________________________
Total Acc: 76.37%
(Model was trained with 150 epochs)
Loss Weight
Shape : Front Color : Back Color = 1 : 1 : 1
1. Shape
Model: Resnet18
Drop-out rate: 0.1
Aleatoric Uncertainty: 0.0085
Epistemic Uncertainty: 0.0001
Validation Acc: 86.63% (8 epoch Train)
________________________________________________________________________________________________________________
2. Front Color
Model: Resnet18
Drop-out rate: 0.1
Aleatoric Uncertainty: -
Epistemic Uncertainty: -
Validation Acc: -
________________________________________________________________________________________________________________
3. Back Color
Model: Resnet18
Drop-out rate: 0.1
Aleatoric Uncertainty: -
Epistemic Uncertainty: -
Validation Acc: -
________________________________________________________________________________________________________________
Total Acc: -
1. All (Shape + Front Color + Back Color)
Model: Resnet18
Drop-out rate: 0.1
Aleatoric Uncertainty: -
Epistemic Uncertainty: -
Validation Shape Acc: 88.09%
Validation Color1 Acc: 84.08%
Validation Color2 Acc: 85.16%
________________________________________________________________________________________________________________
Total Acc: 63.32%
Loss Weight
Shape : Front Color : Back Color = 1 : 1 : 1
Reference Paper:
1. Real-world Pill Segmentation based on Superpixel Merge using Region Adjacency Graph
(Link: http://www.scitepress.org/Papers/2017/61358/61358.pdf)