hykhhijk / Rock-classifier

Rock classifier using CNN

Rock-classifier

Rock classifier using CNN
4개의 암석(Leonardite, Lignite, Charcoal, Waste)을 같은 환경에서 최대한의 정확도를 내는 모델을 만드는 프로젝트이다.
이미지 샘플

Model shape

Input = keras.layers.Input(shape=train_generator[0][0][0].shape)
x = keras.layers.Conv2D(16, kernel_size=7, activation="relu",kernel_initializer="he_normal", padding="same")(Input)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.Conv2D(32, kernel_size=5, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.Conv2D(64, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.GlobalAveragePooling2D()(x)
Output = keras.layers.Dense(4, activation="softmax")(x)

model = keras.models.Model(inputs = Input, outputs = Output)

기존의 간단한 모델 형태로도 모델이 과대적합 되었기에 데이터 증강 및 새로운 모델 형태를 적용해야함.

Input = keras.layers.Input(shape=X[0].shape)
x = keras.layers.Conv2D(16, kernel_size=7, activation="relu",kernel_initializer="he_normal", padding="same")(Input)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.Dropout(0.2)(x)

shortcut =x
x = keras.layers.Conv2D(32, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(32, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(32, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
shortcut = keras.layers.Conv2D(32, kernel_size=5, activation="relu",kernel_initializer="he_normal", padding="same")(shortcut)
x = keras.layers.Add()([x, shortcut])
x   = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)


x = keras.layers.Dropout(0.2)(x)


shortcut =x
x = keras.layers.Conv2D(64, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(64, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(64, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
shortcut = keras.layers.Conv2D(64, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(shortcut)
x = keras.layers.Add()([x, shortcut])
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.Dropout(0.2)(x)

shortcut =x
x = keras.layers.Conv2D(128, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(128, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(128, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
shortcut = keras.layers.Conv2D(128, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(shortcut)
x = keras.layers.Add()([x, shortcut])
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.Dropout(0.2)(x)

shortcut =x
x = keras.layers.Conv2D(256, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(256, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(256, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(x)
shortcut = keras.layers.Conv2D(256, kernel_size=3, activation="relu",kernel_initializer="he_normal", padding="same")(shortcut)
x = keras.layers.Add()([x, shortcut])
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(2)(x)

x = keras.layers.GlobalAveragePooling2D()(x)
Output = keras.layers.Dense(4, activation="softmax")(x)

model = keras.models.Model(inputs = Input, outputs = Output)
model.compile(loss="categorical_crossentropy",
optimizer=keras.optimizers.Adam(learning_rate=0.0001), metrics=["accuracy"])

history_list.append(model.fit(X_train, y_train, validation_data=(X_test, y_test),callbacks=[es],epochs = 100))
print("Train: ", model.evaluate(X_train, y_train))
print("Valid: ", model.evaluate(X_test, y_test))

ResNet의 Skip-connection을 사용한 결과 약95%의 정확도를 보여주었음.

Image augmentation

Albumentations를 사용하여 이미지를 증강하였다.

데이터를 증강시킨 후 성능이 눈에 띄게 좋아졌다. 최종 제출 시에는 데이터를 보간법을 다르게 적용한 데이터셋에 증강법을 적용시켰다.