def head_localization(img_channels=1, img_rows=64, img_cols=64):

    input_img = Input(shape=(img_channels, img_rows, img_cols))

    x = Convolution2D(30, 5, 5, init='normal', subsample=(1,1), activation='tanh')(input_img)
    x = MaxPooling2D((2,2))(x)

    x = Convolution2D(30, 5, 5, init='normal', subsample=(1,1), activation='tanh')(x)
    x = MaxPooling2D((2,2))(x)

    x = Convolution2D(30, 4, 4, init='normal', subsample=(1,1), activation='tanh')(x)
    x = MaxPooling2D((2,2))(x)

    x = Convolution2D(30, 3, 3, init='normal', subsample=(1,1), activation='tanh')(x)
    x = ZeroPadding2D((2, 0))(x)
    x = MaxPooling2D((2, 2))(x)

    x = Convolution2D(120, 3, 3, init='normal', subsample=(1,1), activation='tanh')(x)

    x = Convolution2D(256, 3, 3, init='normal', subsample=(1,1), activation='tanh')(x)

    x = Convolution2D(256, 3, 3, init='normal', subsample=(1,1), activation='tanh')(x)

    x = Flatten()(x)
    x = Dense(256, init='normal', activation='tanh')(x)
    x = Dropout(0.5)(x)

    x = Dense(256, init='normal', activation='tanh')(x)
    x = Dropout(0.5)(x)

    x = Dense(2, activation='tanh')(x)

    model = Model(input=input_img, output=x)

    return model