haluksumen / image_classification_intel

Categorical Image Classification using Tensorflow/Keras on Intel Image Dataset

Intel_Image_Classification

Exploratory Data Analysis + Data Visualization + Deep Learning Modelling

1 - Abstract

In this project I made Exploratory Data Analysis, Data Visualisation and lastly Modelling. Intel Image Dataset contains 25 images in 3 different sets. Each example is 150x150 image and associated with 6 labels(targets). After examining the dataset I used ImageDataGenerator for example rescaling images and increasing the artifical training and test datasets. In modelling part, with a InceptionResNetV2 and several other layers implemented. Model trained with 10 Epochs for training the data. Also for long epochs time I implemented callback for time saving. Overally, model gives 0.9023 accuracy. Furthermore with hyperparameter tuning model can give higher accuracy or using GPU for training it will reduce time and number of epochs can be increased.

2 - Data

Intel Image Dataset contains 25,000 examples,train set of 14,000 test set of 3,000 examples and validation set of 7,000 . Each example is a 150x150 image, associated with a label from 6 labels.

Each training and test example is assigned to one of the following labels:

• 0 Buildings
• 1 Forest
• 2 Glacier
• 3 Mountain
• 4 Sea
• 5 Street

Train Dataset Example

Test Dataset Example

3 - Exploratory Data Analysis

Firstly, I checked data, which came three different dataset which are train, test and validation. Later I checked distribution of labels in datasets moreover I see all the classes(labels) equally distributed. So luckily I dont need to do Oversampling or Undersampling.

Number of images in Train Directory:

• Buildings: 2191
• Street: 2382
• Mountain: 2512
• Glacier: 2404
• Sea: 2274
• Forest: 2271

4 - Data Preprocessing

For preparing datasets to the model I used ImageDataGenerator for rescaling which is 1/255.0. Also I defined batch size 128, and for creating artifical images I used rotating that takes maximum 60 degree rotation. In the below code every parameters are visible with explanation.

train_datagen = ImageDataGenerator(
        rescale=1/255.0,            #multiply the data by the value provided
        featurewise_center=True,    #create generator that centers pixel values
        rotation_range=60,          #maximum 60 degree random rotation
        width_shift_range=0.2,      #fraction of total width, if < 1, or pixels if >= 1.
        height_shift_range=0.2,     #fraction of total height, if < 1, or pixels if >= 1.
        shear_range=0.2,            #image disortion in axis
        fill_mode='nearest')        #fill the color with nearest neighborhood for example, aaaaaaaa|abcd|dddddddd

train_generator = train_datagen.flow_from_directory(
        train_path,
        shuffle=True,              #shuffling the order of the image that is being yielded
        target_size=(150,150),     #size of image
        batch_size=128,            #size of the batches of data 
        class_mode='categorical'   #predicting more than two classes so we will use categorical
    )

5 - Modelling

I used pretrained InceptionResNetV2 model. The InceptionResNetV2 model is already trained more than 1 million images. Then I added these parameters over the pre-trained model.

x = tf.keras.layers.Dropout(0.2)(last_output)
x = tf.keras.layers.Dense(units=128, activation='relu')(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.Dense(units=128, activation='relu')(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.Dense(units=6, activation='softmax')(x)

Finally I am compiling model according these parameters, I used RMSprop class and I gave learning rate 0.0002 and momentum 0.9( A scalar or a scalar Tensor).

model.compile(
        optimizer = tf.keras.optimizers.RMSprop(learning_rate=0.0002, momentum=0.9, centered=True), 
        loss = ['categorical_crossentropy'], 
        metrics = ['accuracy']
    )

Additionally, because of higher epochs time I decided to implement Early Stopping class, which is stopping training when it doesnt improve anymore or extremly low.

    tf.keras.callbacks.EarlyStopping(
        monitor='val_loss',          #which quantity will monitor
        min_delta=0.001,             #minimum change in the monitored quantity to qualify as an improvement
        patience=5,                  #number of epochs with no improvement after which training will be stopped
        verbose=1,                   #verbosity mode
        mode='auto',                 #the direction is automatically inferred from the name of the monitored quantity.
        baseline=None,               #baseline value for the monitored quantity
        restore_best_weights=True)]  #whether to restore model weights from the epoch with the best value of the monitored quantity

6 - Result & Future Work

As a result, my model gives overally good results which is the result of InceptionResNetV2 model.

Accuracy of the Model

Loss of the Model

Test Loss is 0.2682

Test Accuracy is 0.9023

For higher accuracy hyperparameter tuning can be implemented or using GPU for training it will reduce training time and with this time saving epochs number can be increased.