https://github.com/rahul-samedavar/doodlemind
https://github.com/rahul-samedavar/doodlemind
cnn convolutional-neural-networks machine-learning multiclass-classification
Last synced: 12 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/rahul-samedavar/doodlemind
- Owner: Rahul-Samedavar
- License: mit
- Created: 2025-03-12T17:50:58.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2025-03-26T08:50:23.000Z (over 1 year ago)
- Last Synced: 2025-03-26T09:40:52.837Z (over 1 year ago)
- Topics: cnn, convolutional-neural-networks, machine-learning, multiclass-classification
- Language: Jupyter Notebook
- Homepage:
- Size: 24 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: readme.MD
- License: LICENSE
Awesome Lists containing this project
README
# Doodle Classification using CNN
This project focuses on developing a Convolutional Neural Network (CNN) model for classifying hand-drawn doodles into 35 different categories. The model is trained using a subset of [Google's Quick, Draw!](https://quickdraw.withgoogle.com/data) dataset and aims to recognize user-drawn sketches with high accuracy. The Project contains trained models, weights and an Interactive UI.
## Features
- Hand-drawn sketch recognition with real-time classification
- Preprocessing pipeline to normalize input data
- Custom-trained CNN model for efficient doodle classification
- Interactive UI with Dark Neon theme
- Confidence-based prediction results with a probability bar visualization
## Deployment
- The Model and GUI has been deployed in HuggingFace space.
- Access it from [here](https://huggingface.co/spaces/Rahul-Samedavar/DoodleClassifier)
## 🛠 Installation & Setup
### 1️⃣ **Clone the Repository**
```bash
git clone https://github.com/Rahul-Samedavar/DoodleMind.git
cd DoodleMind
```
### 2️⃣ **Create Virtual Environment(Optional)**
create virtual environment to avoi version conflicts.
```bash
python -m venv env
```
This creates a Virtual environment called env in the current directory. You need to activate this virtual environment to use the project.
windows users
```bash
env/Scripts/activate
```
Linux, Unix and Mac users
```bash
source env/bin/activate
```
### 3️⃣ **Install Dependencies**
to install the dependencies run:
```bash
pip install -r requirements.txt
```
### 4️⃣ **Run the Application**
```bash
python app.py
```
The application will be available at `http://localhost:5000`.
## Methodology of Training the Model
### **1. Dataset Preparation**
- The dataset consists of 35 doodle categories from Google's Quick, Draw!
- Subset containing 12000 doodles for each class was choosen too balance the dataset.
- Images were resized to **28x28 grayscale** format for model compatibility.
- Data was normalized to a **[0,1]** range to improve generalization.
- numpy.float16 data type was used for memory efficiency.
- Test split : 0.15
### **2. Model Architecture**
A CNN architecture was designed with the following layers:
```python
model = Sequential([
Conv2D(16, (3,3), activation='relu', input_shape=(28, 28, 1)),
MaxPooling2D(2,2),
Conv2D(64, (3,3), activation='relu'),
MaxPooling2D(2,2),
Flatten(),
Dropout(0.2),
Dense(128, activation='relu'),
Dropout(0.2),
Dense(64, activation='relu'),
Dropout(0.2),
Dense(35, activation='softmax')
])
```
```bash
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 26, 26, 16) 160
max_pooling2d(MaxPooling2D) (None, 13, 13, 16) 0
conv2d_1 (Conv2D) (None, 11, 11, 64) 9280
max_pooling2d_1 (MaxPooling (None, 5, 5, 64) 0
2D)
flatten (Flatten) (None, 1600) 0
dropout (Dropout) (None, 1600) 0
dense (Dense) (None, 128) 204928
dropout_1 (Dropout) (None, 128) 0
dense_1 (Dense) (None, 64) 8256
dropout_2 (Dropout) (None, 64) 0
dense_2 (Dense) (None, 35) 2275
=================================================================
Total params: 224,899
Trainable params: 224,899
Non-trainable params: 0
_________________________________________________________________
```
### **3. Loss Function & Optimization**
- **Loss Function**: Categorical Crossentropy (since it's a multi-class classification problem)
- **Optimizer**: Adam (tuned for better convergence)
- **Learning Rate**: Tested over a range of values for initial learning rate and found sweet spot at 0.001
```python
model.compile(
optimizer=Adamax(0.001),
loss='categorical_crossentropy',
metrics=['accuracy', TopKCategoricalAccuracy(3), Precision(), Recall()]
)
```
### **4. Training Process**
- **Training Data Split**: 80% training, 20% validation,
- **Batch Size**: 32 for memory efficiency
- **Epochs**: 50 epochs with saving model at each epoch through checkpoints.
```python
from tensorflow.keras.callbacks import ModelCheckpoint
checkpoint = ModelCheckpoint(
filepath = "Models/model_epoch_{epoch:02}.keras",
save_weights_only = False,
save_best_only = False,
monitor = 'val_loss',
mode = 'min',
verbose = 0
)
```
- **Evaluation Metrics**: Accuracy, Top K Categorical Accuracy with k = 3, Precision, and Recall for model effectiveness
### **5. Testing and Model Selection**
- Models saved in each epoch are evluated using the test dataset.
- Model from epoch 50 showed the best accuracy of 86.29%.
- Model from epoch 44 showed the best Top K Categorical Accuracy of 95.31%.
- This scores are acceptable as they not only indicate good predictions, but also indicate the similarities between some classes. This can be understood from Confusion Matrix.
- As an example, we can see that the model predicted nearly half of the school busses as bus and 523 busses as school bus
- Although this reduces accuracy, it's a good sign considering both classes look nearly the same in doodles.
### **6. Deployment Considerations**
- Model and Weights exported in Keras and h5 format respectively.
- Integrated with a real-time drawing interface for inference
- Optimized preprocessing to match training data conditions (grid-based downscaling to 28x28 format)
---