Coffee Beans Moisture Detection with Fusioned Triple Deep Convolutional Neural Network

• Authors: Ivan Arsyaditya Prananda, Mohamad Arsya Kaukabi, and Muhammad Naufal Ariiq

• Affiliation: Bangkit Academy

• E-mail: arsyakaukabi@gmail.com

Graphical Abstract

Fig 1 Triple deep convolutional neural network model for moisture level detection

Dataset

The dataset for this work is manually collected by our team.

The dataset used for this work is images of green coffee beans with different moisture. There is 5 different classification:

There are total 416 images splitted into train_data, validation_data, and test_data. You can find the dataset here: Google Drive Prepared Dataset

Dataset Size : 411MB

Models

Transfer learning and model fusion is used in this work. There are 3 fusioned model :

• InceptionV3
• VGG16
• DenseNet121

For an immediate simulation without the hassle of going over the previous instructions, refer to this link:

Pre-Trained Weights

PRE-TRAINED WEIGHTS FILESIZE: (344 MB)

How To Use

1. Open the TDCNN_1.ipynb file in Co-ffee_MoistureDetection/Model Trainer/
2. Import all the required libraries.
3. Build the model with transfer learning of InceptionV3, VGG16, and DenseNet121.
4. Fuse the previous three transfer learning model into one model and make sure when all this three is fused, they have the same input layer.
5. Download the dataset from the link and load it into ImageDataGenerator with .flow_from_directory
6. Start the training.
7. After all T-DCNN models are built, you may now run the testing.py from the main Co-ffee_MoistureDetection/ folder.
8. Follow through the given instructions and make sure to use the test sample from the provided /test/ folder

Results

Fig 2 Accuracy and loss graph after 25 epochs.

Co-ffee Github Repo Links

Machine Learning

Classification of Coffee Leaf Diseases

Green Coffee Beans Moisture Level Detection

Cloud Computing

Disease classification API

Coffee beans Moisture level detection API

Mobile Development

Project Android Studio

Acknowledgment

Thanks to Bangkit Academy. Without its support, this work would not have become possible.