Repository contains image classification models with 31 classes. The models are trained using custom CNN methods and transfer learning to recognize Indonesian spices images. Five architectures that we use for transfer learning includes: InceptionV3, Xception, DenseNet121, EfficientNetV2B0, and MobileNetV3. After the training, we find out that the model with the highest accuracy is DenseNet121 architecture with 94.37% accuracy on test dataset.
There are total 6510 data and 31 classes of indonesian spices images including:
- adas
- andaliman
- asam jawa
- bawang bombai
- bawang merah
- bawang putih
- biji ketumbar
- bukan rempah
- bunga lawang
- cengkeh
- daun jeruk
- daun kemangi
- daun ketumbar
- daun salam
- jahe
- jinten
- kapulaga
- kayu manis
- kayu secang
- kemiri
- kemukus
- kencur
- kluwek
- kunyit
- lada
- lengkuas
- pala
- saffron
- serai
- vanili
- wijen
- Before training, we have splitted the data into 70% for training, 15% for validation, and 15% for test.
- Visualization some of the data
- We have trained the model and after 50 epochs, we obtained loss: 0.1228 and accuracy 96.44% for training. For validation, the val_loss is 0.3062 with 93.35% accuracy.
- We also have evaluated the model on test set, and achieved loss 0.23866 with 94.37% accuracy.
- Visualization of data prediction
- Confusion matrix
https://drive.google.com/drive/folders/1axCQ-LzEvcT-2hm-bzMSf215pR7z8iWe?usp=sharing