Deeplab v3 turf datasets 훈련

데이터 폴더 생성

데이터 선정

훈련에 적합하다고 판단되는 데이터 20장 선정 후 훈련 데이터 폴더에 복사

훈련 : 검증 = 0.85 : 0.15 = 17 : 3

훈련

Backbone Model	Resnet50
Image Shape	(224 x 224 x 3)
Mask Shape	(224 x 224 x 2)
Number of output channels	2 (0 – background, 1 – abnormal)
Number of Epochs Trained	50
Batch Size	3
Optimizer	Adam
Learning Rate	0.0001
Loss Function/s	Binary Cross entropy Intersection over union (BCE + IoU)
Evaluation Metric	Intersection over Union (IoU)

python main_train.py --epochs 50 --batch 3

검증

python main_test.py --input ../input/Turf_Dataset/valid_images/

ori_data	valid_data

.

ori_data	valid_data

테스트

python main_test.py --input ../input/Turf_Dataset/test_images/

참조

Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation