Leaf Classification

This project aims to classify leaf types using deep learning with various CNN architectures. There are two types of preprocessing: standard and multicolor (specifically for Dracaena/lucky bamboo leaves).

---

Dataset

You can get dataset from:

• IEEE Dataport - Five Class Indonesian Plant Species Leaf Image Dataset
• Kaggle - Five Class Indonesian Plant Species Leaf

---

Research

You can check my research about species classification from: Academia.edu

---

Preprocessing

1. Standard Preprocessing

Used for most leaves:

• Cropping & Rotate: Crop and rotate images to center the leaf object.
• Brightening: Increase image brightness.
• Median Blur (Denoising): Reduce noise in the image.
• Thresholding Binary (Binarize): Convert the image to binary (black and white).
• Get Biggest Scratch: Extract the largest leaf area.
• Replace White: Replace the white background.

2. Multicolor Preprocessing (For Dracaena Only)

Used for Dracaena/lucky bamboo leaves:

• Cropping & Rotate
• Brightening
• Binarize Multicolor
• Denoising Colored
• Gaussian Blur
• Conversion to HSV
• Thresholding Yellow Color
• Morphology Closing
• Get Biggest Scratch
• Replace White

Note:
Multicolor preprocessing is only used for Dracaena/lucky bamboo leaves.

---

Workflow

1. Split Data

   • Split the data into train, validation, and testing sets.

2. Preprocessing

   • Apply standard preprocessing for regular leaves.
   • Apply multicolor preprocessing for Dracaena leaves.

3. Training

   • Train the model with various CNN architectures (AlexNet, LeNet5, ResNet50, LeafNet, Custom1, Custom2).

4. Testing

   • Testing the model using the testing data.

5. Predict

   • Predict the leaf type on new data using the trained model.

---

Requirements

• Python 2.x & Python 3.x
• Keras & TensorFlow
• OpenCV
• NumPy
• scikit-learn
• Google Colab (recommended)

---

Model Accuracy & Loss Results

Architecture	Top 1 Acc	Top 1 Loss	Top 1 Val Acc	Top 1 Val Loss	Top 5 Acc	Top 5 Loss	Top 5 Val Acc	Top 5 Val Loss
LeNet5	0.9853	0.0446	0.2423	10.4603	0.9915	0.0248	0.2924	10.2835
AlexNet	0.9722	0.0818	0.9499	0.1995	0.9775	0.0587	0.9788	0.1767
LeafNet	0.9698	0.0928	0.7428	1.1867	0.9728	0.0998	0.9237	0.2735
ResNet50	0.9848	0.0427	0.7839	0.9675	0.9890	0.0408	0.9499	0.1173
Custom1	0.9706	0.0733	0.9453	0.2361	0.9754	0.0737	0.9537	0.1060
Custom2	0.9738	0.0689	0.9060	0.2579	0.9765	0.0638	0.9198	0.2935

Description:

• Acc: Accuracy on training data
• Loss: Loss on training data
• Val Acc: Accuracy on validation data
• Val Loss: Loss on validation data

---

The table above shows the performance of each model architecture for leaf classification, both for Top 1 and Top 5 best predictions.

---

Confusion Matrix

Below are confusion matrices and accuracy values from several CNN architectures used:

AlexNet	LeNet5	ResNet50

LeafNet	Custom1	Custom2

Each image shows the confusion matrix and the final accuracy value of the corresponding model.

API Implementation

The API implementation for this project is available at https://github.com/hafiizhekom/daunesia-api.

---