第 12 章 · 真实图片去噪重建代码实验¶

# 真实建筑照片去噪：用同一张图片的局部图块训练复原模型。
denoise_clean = np.asarray(Image.fromarray(load_sample_image("china.jpg")).resize((96, 96))).astype("float32") / 255.0
denoise_patch_size = 12
denoise_stride = 6
patches_clean = []
patch_coords = []
for row in range(0, denoise_clean.shape[0] - denoise_patch_size + 1, denoise_stride):
    for col in range(0, denoise_clean.shape[1] - denoise_patch_size + 1, denoise_stride):
        patches_clean.append(denoise_clean[row:row + denoise_patch_size, col:col + denoise_patch_size].reshape(-1))
        patch_coords.append((row, col))
patches_clean = np.array(patches_clean)

rng = np.random.default_rng(19)
noise_sigma = 0.18
patches_noisy = np.clip(patches_clean + rng.normal(0, noise_sigma, patches_clean.shape), 0, 1)
patch_denoiser = MLPRegressor(hidden_layer_sizes=(160,), max_iter=180, random_state=19)
patch_denoiser.fit(patches_noisy, patches_clean)

noisy_image = np.clip(denoise_clean + rng.normal(0, noise_sigma, denoise_clean.shape), 0, 1)
noisy_image_patches = np.array([
    noisy_image[row:row + denoise_patch_size, col:col + denoise_patch_size].reshape(-1)
    for row, col in patch_coords
])
predicted = np.clip(patch_denoiser.predict(noisy_image_patches), 0, 1)

denoised_image = np.zeros_like(denoise_clean)
denoise_weight = np.zeros(denoise_clean.shape[:2] + (1,), dtype=float)
for (row, col), patch in zip(patch_coords, predicted):
    denoised_image[row:row + denoise_patch_size, col:col + denoise_patch_size] += patch.reshape(denoise_patch_size, denoise_patch_size, 3)
    denoise_weight[row:row + denoise_patch_size, col:col + denoise_patch_size] += 1
denoised_image = denoised_image / np.maximum(denoise_weight, 1)

denoise_summary = pd.DataFrame(
    [
        {"图像": "带噪输入", "均方误差": mean_squared_error(denoise_clean.reshape(-1), noisy_image.reshape(-1))},
        {"图像": "去噪输出", "均方误差": mean_squared_error(denoise_clean.reshape(-1), denoised_image.reshape(-1))},
    ]
)
display(pd.DataFrame({
    "图片": ["china.jpg"],
    "图块数": [len(patches_clean)],
    "噪声标准差": [noise_sigma],
}))
display(denoise_summary.round(4))

# 绘制真实图片、噪声输入和模型去噪输出。
fig, axes = plt.subplots(2, 2, figsize=(8.4, 7.0))
for ax, image, title in zip(
    axes.ravel(),
    [denoise_clean, noisy_image, denoised_image, np.abs(denoise_clean - denoised_image) * 3],
    ["原图", "带噪输入", "去噪输出", "误差 x3"],
):
    ax.imshow(np.clip(image, 0, 1))
    ax.set_title(title, fontweight="bold")
    ax.set_xticks([])
    ax.set_yticks([])
fig.suptitle("真实图片去噪重建：从噪声输入恢复建筑纹理", x=0.08, ha="left", fontsize=14, fontweight="bold", color="#0f172a")
plt.tight_layout()
plt.show()