🏕️Praktikum 4

Penerapan metode Self-Organizing Map (SOM) untuk segmentasi citra Lenna.

Download dataset terlebih dahulu

463KB

Lenna.png

image

Open

instalasi minisom

pip install minisom

Nanti akan muncul tampilan seperti dibawah ini:

Collecting minisom
  Downloading MiniSom-2.3.1.tar.gz (10 kB)
  Preparing metadata (setup.py) ... done
Building wheels for collected packages: minisom
  Building wheel for minisom (setup.py) ... done
  Created wheel for minisom: filename=MiniSom-2.3.1-py3-none-any.whl size=10588 sha256=c45ce374f32484fbd730daef8874570c408920644f2a0e7f17402eae6556fd9d
  Stored in directory: /root/.cache/pip/wheels/c7/92/d2/33bbda5f86fd8830510b16aa98c8dd420129b5cb24248fd6db
Successfully built minisom
Installing collected packages: minisom
Successfully installed minisom-2.3.1

import numpy as np
import matplotlib.pyplot as plt
from skimage import io

# Fungsi untuk menginisialisasi bobot SOM
def initialize_weights(input_shape, output_shape):
    return np.random.rand(output_shape[0], output_shape[1], input_shape[2])

# Fungsi untuk menghitung jarak antara vektor input dan bobot SOM
def calculate_distance(input_vector, weights):
    return np.linalg.norm(input_vector - weights, axis=2)

# Fungsi untuk menemukan indeks unit pemenang (unit dengan bobot terdekat)
def find_winner_unit_in_image(input_vector, weights):
    distances = calculate_distance(input_vector, weights)
    return np.unravel_index(np.argmin(distances), distances.shape)

# Fungsi untuk memperbarui bobot SOM
def update_weights(input_vector, weights, winner, learning_rate, neighborhood_radius):
    distances = np.linalg.norm(np.indices(weights.shape[:2]).T - np.array(winner).reshape(1, -1), axis=2)
    influence = np.exp(-distances / (2 * neighborhood_radius**2))
    weights += learning_rate * influence[:, :, np.newaxis] * (input_vector - weights)

# Fungsi untuk melatih SOM
def train_som(image, num_epochs, initial_learning_rate, initial_neighborhood_radius):
    input_shape = image.shape
    som_shape = (10, 10, input_shape[2])  # Ukuran SOM sesuai dengan jumlah saluran warna
    weights = initialize_weights(input_shape, som_shape)

    for epoch in range(num_epochs):
        # Update parameter pembelajaran dan radius tetangga
        learning_rate = initial_learning_rate * np.exp(-epoch / num_epochs)
        neighborhood_radius = initial_neighborhood_radius * np.exp(-epoch / num_epochs)
        # Pemrosesan SOM
        for i in range(input_shape[0]):
            for j in range(input_shape[1]):
                input_vector = image[i, j, :]
                winner = find_winner_unit_in_image(input_vector, weights)
                update_weights(input_vector, weights, winner, learning_rate, neighborhood_radius)

    return weights

# Load citra Lenna (Anda bisa mengganti ini dengan citra lain jika diperlukan)
Lenna_path = "Lenna.png"
Lenna = io.imread(Lenna_path) / 255.0  # Normalisasi intensitas piksel menjadi rentang [0, 1]

# Latih SOM
num_epochs = 100
initial_learning_rate = 0.1
initial_neighborhood_radius = 5
trained_weights = train_som(Lenna, num_epochs, initial_learning_rate, initial_neighborhood_radius)

# Visualisasi bobot SOM
plt.imshow(trained_weights)
plt.title('Trained SOM Weights for Lena')
plt.show()

Nanti akan muncul tampilan seperti dibawah ini: