This repository builds a Convolutional Neural Network (CNN) using TensorFlow/Keras to perform multi-class image classification. The model classifies images into three categories: driving_license, social_security, and others. It utilizes a combination of convolutional layers, batch normalization, max pooling, and dropout to achieve high performance in classifying these images.

The aim is to develop a sequential CNN model that can accurately classify images into one of three predefined classes. The model is built using TensorFlow/Keras and trained on a dataset of images from the categories driving_license, social_security, and others.

The dataset consists of images categorized into three folders:

• driving_license
• social_security
• others

Each image is resized to 224x224 pixels, and data augmentation techniques such as rotation, zoom, and flips were applied to improve generalization and reduce overfitting.

The CNN model follows the architecture outlined below:

# Build the CNN model
model= Sequential()

# Input layer
model.add(Input(shape= (224,224,3)))

# First convolutional layer
model.add(Conv2D(32, 3, padding= 'same', activation= 'relu'))

# Batch normalization to normalize the inputs to the next layer
model.add(BatchNormalization())

# Max pooling layer to reduce spacial dimensions of the feature map
model.add(MaxPool2D())

# Dropout layer to prevent overfitting
model.add(Dropout(0.4))

# Flatten the feature map into a 1D vector for the fully connected layer
model.add(Flatten())

# Fully connected layer
model.add(Dense(128, activation= 'relu'))

# Output layer with 3 units (one for each class) and softmax activation for multi-class classification
model.add(Dense(3, activation= 'softmax'))

# Compile the model
model.compile(optimizer= Adam(learning_rate= 0.00001), loss= tf.keras.losses.SparseCategoricalCrossentropy(from_logits= True), metrics= ['accuracy'])

• Convolutional Layers: Extract features from the input images using 3x3 kernels.
• Batch Normalization: Normalizes the output of the convolutional layers, speeding up training and improving model stability.
• Max Pooling: Reduces the spatial dimensions of the feature maps, helping to make the model more efficient.
• Dropout: A dropout rate of 40% is used to prevent overfitting by randomly disabling some neurons during training.
• Fully Connected Layer: After flattening the feature maps, a dense layer with 128 neurons is applied.
• Output Layer: A softmax layer with 3 units for multi-class classification, producing probabilities for each class.

The model achieved an accuracy of 86% on the validation set, with the following detailed performance per class:

• Social Security: The model achieved high recall (0.95), meaning it correctly identified most social_security cases, though precision was slightly lower (0.76), suggesting some false positives.
• Others: The model showed strong precision (0.97) but slightly lower recall (0.78), indicating it missed a few true positives for this class.
• Driving License: A balanced performance with precision and recall around 0.89 and 0.85, respectively.

This CNN model provides a robust approach to multi-class image classification with strong overall performance. With further fine-tuning, such as enhancing precision for certain classes, this model could be adapted for broader applications. Advanced techniques such as Transfer Learning or additional data augmentation could further improve model generalization.