Experimental performance records
Params
Pre-trained weights: imagenet - resnet50
Input shape: (224, 224, 3)
Encoder trainable: False
Batch size: 32
Epochs: 533
Model
Input Image (224, 224, 3) → resnet50 (224, 224, 32) → Result (4)
Total params: 49,725,436
Trainable params: 40,912,012
Non-trainable params: 8,813,424
Dataset
Preprocessing
- Remove malformed images from entire dataset.
- Fill missing polygons from label data using CVAT.
- Exports its ROI, binary masks and transformed original images.
- Select pixel from binary masks.
Compile options
Loss
- categorical_crossentropy
Metrics
- categorical_accuracy
Optimizer
- Adam