遙感圖像數字水印系統優化方案

1. 引言

遙感圖像在現代地理信息系統、環境監測、軍事偵察等領域發揮著重要作用。為了保護遙感圖像的版權和完整性，數字水印技術被廣泛應用。然而，現有的遙感圖像水印方案往往在不可見性、魯棒性和容量之間存在權衡，難以同時滿足所有要求。

本文針對現有遙感圖像水印代碼效果不佳的問題，提出了一套全面的優化方案。我們將從算法選擇、預處理技術、嵌入策略、提取方法和性能評估等多個方面進行深入分析和優化，旨在提升水印系統的整體性能。

2. 現有問題分析

在優化之前，我們需要全面分析現有代碼可能存在的問題：

不可見性差：水印嵌入后導致圖像質量明顯下降
魯棒性不足：水印難以抵抗常見圖像處理攻擊
容量限制：嵌入的水印信息量有限
計算效率低：嵌入和提取過程耗時過長
適應性差：對不同類型遙感圖像表現不一致

3. 優化方案設計

3.1 算法選擇與改進

3.1.1 離散小波變換(DWT)與奇異值分解(SVD)結合

傳統LSB(最低有效位)方法雖然簡單，但魯棒性較差。我們采用DWT+SVD組合算法，在頻域中嵌入水印，提高魯棒性。

import numpy as np
import pywt
import cv2
from scipy.linalg import svd
from skimage.metrics import peak_signal_noise_ratio as psnr
from skimage.metrics import structural_similarity as ssim
import matplotlib.pyplot as pltclass EnhancedWatermark:def __init__(self, wavelet='haar', level=2, alpha=0.1):self.wavelet = waveletself.level = levelself.alpha = alpha  # 水印強度因子def dwt2(self, image):"""二維離散小波變換"""coeffs = pywt.wavedec2(image, self.wavelet, level=self.level)return coeffsdef idwt2(self, coeffs):"""二維離散小波逆變換"""return pywt.waverec2(coeffs, self.wavelet)def svd_embed(self, host_block, watermark_block):"""使用SVD在主機圖像塊中嵌入水印"""U, S, Vh = svd(host_block, full_matrices=False)# 修改奇異值嵌入水印S_modified = S + self.alpha * watermark_block# 重構圖像塊host_modified = np.dot(U, np.dot(np.diag(S_modified), Vh))return host_modified, U, S, Vhdef svd_extract(self, host_modified, U, S, Vh):"""從修改后的圖像塊中提取水印"""U_mod, S_mod, Vh_mod = svd(host_modified, full_matrices=False)watermark_extracted = (S_mod - S) / self.alphareturn watermark_extracted

3.1.2 自適應水印強度調整

根據圖像局部特征自適應調整水印強度，在紋理復雜區域使用較強水印，在平滑區域使用較弱水印。

def calculate_texture_map(image, block_size=8):"""計算圖像紋理圖，用于自適應水印強度"""height, width = image.shapetexture_map = np.zeros_like(image, dtype=np.float32)for i in range(0, height, block_size):for j in range(0, width, block_size):block = image[i:i+block_size, j:j+block_size]if block.size > 0:# 使用標準差作為紋理復雜度的度量texture = np.std(block)texture_map[i:i+block_size, j:j+block_size] = texture# 歸一化紋理圖texture_map = (texture_map - texture_map.min()) / (texture_map.max() - texture_map.min() + 1e-10)return texture_mapdef adaptive_alpha(texture_map, base_alpha=0.05, max_alpha=0.2):"""根據紋理圖計算自適應水印強度"""return base_alpha + (max_alpha - base_alpha) * texture_map

3.2 預處理優化

3.2.1 圖像預處理

def preprocess_image(image, target_size=None, normalize=True):"""圖像預處理函數:param image: 輸入圖像:param target_size: 目標尺寸 (width, height):param normalize: 是否歸一化到[0,1]:return: 預處理后的圖像"""# 轉換為灰度圖像（如果原是彩色）if len(image.shape) == 3:image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)# 調整尺寸if target_size is not None:image = cv2.resize(image, target_size, interpolation=cv2.INTER_AREA)# 歸一化if normalize:image = image.astype(np.float32) / 255.0return imagedef preprocess_watermark(watermark, target_shape):"""水印預處理函數:param watermark: 輸入水印，可以是二值圖像或文本:param target_shape: 目標形狀:return: 預處理后的水印圖像"""if isinstance(watermark, str):# 文本水印轉換為圖像watermark_img = text_to_image(watermark, target_shape)else:# 圖像水印處理watermark_img = cv2.resize(watermark, (target_shape[1], target_shape[0]))# 二值化處理if len(watermark_img.shape) == 3:watermark_img = cv2.cvtColor(watermark_img, cv2.COLOR_BGR2GRAY)# 歸一化到[0,1]watermark_img = watermark_img.astype(np.float32) / 255.0return watermark_imgdef text_to_image(text, shape, font_scale=0.8, thickness=2):"""將文本轉換為圖像"""img = np.zeros(shape, dtype=np.uint8)font = cv2.FONT_HERSHEY_SIMPLEXtext_size = cv2.getTextSize(text, font, font_scale, thickness)[0]text_x = (shape[1] - text_size[0]) // 2text_y = (shape[0] + text_size[1]) // 2cv2.putText(img, text, (text_x, text_y), font, font_scale, (255, 255, 255), thickness)return img

3.2.2 基于人眼視覺系統(HVS)的優化

利用人眼視覺系統特性，在不易察覺的區域嵌入更強水印。

def calculate_hvs_mask(image, block_size=8):"""基于人眼視覺系統計算掩碼在紋理復雜和邊緣區域可以嵌入更強水印"""# 計算梯度幅值sobelx = cv2.Sobel(image, cv2.CV_64F, 1, 0, ksize=3)sobely = cv2.Sobel(image, cv2.CV_64F, 0, 1, ksize=3)gradient_magnitude = np.sqrt(sobelx**2 + sobely**2)# 計算局部方差local_variance = np.zeros_like(image)for i in range(0, image.shape[0], block_size):for j in range(0, image.shape[1], block_size):block = image[i:i+block_size, j:j+block_size]if block.size > 0:local_variance[i:i+block_size, j:j+block_size] = np.var(block)# 歸一化gradient_magnitude = (gradient_magnitude - gradient_magnitude.min()) / (gradient_magnitude.max() - gradient_magnitude.min() + 1e-10)local_variance = (local_variance - local_variance.min()) / (local_variance.max() - local_variance.min() + 1e-10)# 結合梯度和方差信息hvs_mask = 0.7 * gradient_magnitude + 0.3 * local_variancereturn hvs_mask

3.3 水印嵌入優化

def embed_watermark_optimized(host_image, watermark, wavelet='haar', level=2, base_alpha=0.05):"""優化的水印嵌入函數:param host_image: 主機圖像:param watermark: 水印圖像或文本:param wavelet: 使用的小波基:param level: 小波分解層數:param base_alpha: 基礎水印強度:return: 含水印圖像、嵌入參數"""# 預處理主機圖像host_processed = preprocess_image(host_image)# 預處理水印watermark_processed = preprocess_watermark(watermark, host_processed.shape)# 計算HVS掩碼和自適應alphahvs_mask = calculate_hvs_mask(host_processed)alpha_map = adaptive_alpha(hvs_mask, base_alpha=base_alpha)# 小波分解coeffs = pywt.wavedec2(host_processed, wavelet, level=level)cA = coeffs[0]  # 近似系數# 在近似系數中嵌入水印watermark_resized = cv2.resize(watermark_processed, (cA.shape[1], cA.shape[0]))# 使用SVD嵌入U, S, Vh = svd(cA, full_matrices=False)S_modified = S + alpha_map[:S.shape[0], :S.shape[1]] * watermark_resized# 重構近似系數cA_modified = np.dot(U, np.dot(np.diag(S_modified), Vh))# 更新系數coeffs_modified = list(coeffs)coeffs_modified[0] = cA_modified# 小波重構watermarked_image = pywt.waverec2(coeffs_modified, wavelet)# 后處理watermarked_image = np.clip(watermarked_image, 0, 1)# 保存嵌入參數用于提取embed_params = {'U': U,'S': S,'Vh': Vh,'alpha_map': alpha_map,'wavelet': wavelet,'level': level,'watermark_shape': watermark_resized.shape}return watermarked_image, embed_paramsdef block_based_embedding(host_image, watermark, block_size=64, wavelet='haar'):"""基于分塊的水印嵌入，提高局部適應性"""height, width = host_image.shapewatermarked_image = np.copy(host_image)embed_params = {}# 確保水印可以被分塊整除wm_height, wm_width = watermark.shapeblocks_vertical = height // block_sizeblocks_horizontal = width // block_size# 調整水印尺寸watermark_resized = cv2.resize(watermark, (blocks_horizontal, blocks_vertical))param_blocks = {}for i in range(blocks_vertical):for j in range(blocks_horizontal):# 提取圖像塊block = host_image[i*block_size:(i+1)*block_size, j*block_size:(j+1)*block_size]if block.size == 0:continue# 計算該塊的紋理復雜度texture = np.std(block)alpha = 0.05 + 0.15 * (texture / (np.std(host_image) + 1e-10))# 小波分解coeffs = pywt.wavedec2(block, wavelet, level=2)cA = coeffs[0]# 嵌入水印位if i < watermark_resized.shape[0] and j < watermark_resized.shape[1]:wm_bit = watermark_resized[i, j]# SVD分解U, S, Vh = svd(cA, full_matrices=False)S_modified = S + alpha * wm_bit# 重構cA_modified = np.dot(U, np.dot(np.diag(S_modified), Vh))# 更新系數coeffs_modified = list(coeffs)coeffs_modified[0] = cA_modified# 逆小波變換block_modified = pywt.waverec2(coeffs_modified, wavelet)# 放回圖像watermarked_image[i*block_size:(i+1)*block_size, j*block_size:(j+1)*block_size] = block_modified# 保存參數param_blocks[(i, j)] = {'U': U,'S': S,'Vh': Vh,'alpha': alpha}embed_params['param_blocks'] = param_blocksembed_params['block_size'] = block_sizeembed_params['wm_shape'] = watermark_resized.shapereturn watermarked_image, embed_params

3.4 水印提取優化

def extract_watermark_optimized(watermarked_image, embed_params, original_image=None):"""優化的水印提取函數:param watermarked_image: 含水印圖像:param embed_params: 嵌入時保存的參數:param original_image: 原始圖像（非盲水印時需要）:return: 提取的水印"""# 預處理圖像watermarked_processed = preprocess_image(watermarked_image)if 'param_blocks' in embed_params:# 分塊水印提取return extract_block_based(watermarked_processed, embed_params, original_image)else:# 全局水印提取return extract_global(watermarked_processed, embed_params, original_image)def extract_global(watermarked_image, embed_params, original_image):"""全局水印提取"""wavelet = embed_params['wavelet']level = embed_params['level']U = embed_params['U']S = embed_params['S']Vh = embed_params['Vh']alpha_map = embed_params['alpha_map']wm_shape = embed_params['watermark_shape']# 小波分解coeffs = pywt.wavedec2(watermarked_image, wavelet, level=level)cA_modified = coeffs[0]# SVD分解修改后的近似系數U_mod, S_mod, Vh_mod = svd(cA_modified, full_matrices=False)# 提取水印watermark_extracted = (S_mod - S) / alpha_map[:S.shape[0], :S.shape[1]]# 調整尺寸watermark_extracted = cv2.resize(watermark_extracted, (wm_shape[1], wm_shape[0]))return watermark_extracteddef extract_block_based(watermarked_image, embed_params, original_image):"""分塊水印提取"""param_blocks = embed_params['param_blocks']block_size = embed_params['block_size']wm_shape = embed_params['wm_shape']wavelet = embed_params.get('wavelet', 'haar')height, width = watermarked_image.shapewatermark_extracted = np.zeros(wm_shape)for i in range(wm_shape[0]):for j in range(wm_shape[1]):if (i, j) not in param_blocks:continue# 提取圖像塊block = watermarked_image[i*block_size:(i+1)*block_size, j*block_size:(j+1)*block_size]if block.size == 0:continue# 獲取該塊的參數params = param_blocks[(i, j)]U = params['U']S = params['S']Vh = params['Vh']alpha = params['alpha']# 小波分解coeffs = pywt.wavedec2(block, wavelet, level=2)cA_modified = coeffs[0]# SVD分解U_mod, S_mod, Vh_mod = svd(cA_modified, full_matrices=False)# 提取水印wm_value = (S_mod - S) / alphawatermark_extracted[i, j] = np.mean(wm_value)return watermark_extracteddef blind_extraction(watermarked_image, wavelet='haar', level=2, wm_shape=(64, 64)):"""盲水印提取（不需要原始圖像）這種方法魯棒性較差，但適用于某些場景"""# 小波分解coeffs = pywt.wavedec2(watermarked_image, wavelet, level=level)cA = coeffs[0]# 假設水印嵌入在奇異值中U, S, Vh = svd(cA, full_matrices=False)# 使用統計方法估計水印# 這里使用簡單的閾值方法，實際應用中可能需要更復雜的方法watermark_estimated = np.zeros_like(cA)mean_S = np.mean(S)std_S = np.std(S)# 假設水印是二值的，通過統計異常值來提取threshold = mean_S + std_Swatermark_estimated = (S > threshold).astype(np.float32)# 調整尺寸watermark_extracted = cv2.resize(watermark_estimated, (wm_shape[1], wm_shape[0]))return watermark_extracted

3.5 后處理與增強

def postprocess_watermark(extracted_watermark, method='threshold', **kwargs):"""后處理提取的水印，增強可讀性"""if method == 'threshold':# 二值化threshold = kwargs.get('threshold', 0.5)processed = (extracted_watermark > threshold).astype(np.float32)elif method == 'median_filter':# 中值濾波kernel_size = kwargs.get('kernel_size', 3)processed = cv2.medianBlur((extracted_watermark * 255).astype(np.uint8), kernel_size)processed = processed.astype(np.float32) / 255.0elif method == 'wiener_filter':# 維納濾波processed = cv2.filter2D(extracted_watermark, -1, wiener_kernel(5))else:processed = extracted_watermarkreturn processeddef wiener_kernel(size=5):"""生成維納濾波核"""kernel = np.ones((size, size)) / (size * size)return kerneldef enhance_watermark_quality(watermark, enhance_method='histogram_equalization'):"""增強水印質量"""if enhance_method == 'histogram_equalization':# 直方圖均衡化watermark_uint8 = (watermark * 255).astype(np.uint8)enhanced = cv2.equalizeHist(watermark_uint8)return enhanced.astype(np.float32) / 255.0elif enhance_method == 'contrast_stretch':# 對比度拉伸min_val = np.min(watermark)max_val = np.max(watermark)if max_val > min_val:enhanced = (watermark - min_val) / (max_val - min_val)else:enhanced = watermarkreturn enhancedelse:return watermark

3.6 魯棒性測試與評估

def test_robustness(watermarked_image, original_image, embed_params, attacks=None):"""測試水印魯棒性"""if attacks is None:attacks = {'gaussian_noise': lambda img: add_gaussian_noise(img, mean=0, sigma=0.01),'salt_pepper': lambda img: add_salt_pepper_noise(img, amount=0.01),'gaussian_blur': lambda img: cv2.GaussianBlur(img, (5, 5), 0.5),'jpeg_compression': lambda img: jpeg_compression(img, quality=70),'cropping': lambda img: cropping_attack(img, ratio=0.1),'rotation': lambda img: rotate_image(img, angle=5),'scaling': lambda img: scaling_attack(img, ratio=0.9)}results = {}extracted_watermarks = {}for attack_name, attack_func in attacks.items():# 應用攻擊attacked_image = attack_func(watermarked_image.copy())# 提取水印try:extracted_wm = extract_watermark_optimized(attacked_image, embed_params, original_image)extracted_watermarks[attack_name] = extracted_wm# 評估提取質量（如果有原始水印）# 這里需要原始水印信息，實際應用中可能需要調整except Exception as e:print(f"攻擊 {attack_name} 失敗: {e}")extracted_watermarks[attack_name] = Nonereturn extracted_watermarksdef add_gaussian_noise(image, mean=0, sigma=0.01):"""添加高斯噪聲"""noise = np.random.normal(mean, sigma, image.shape)noisy_image = image + noisereturn np.clip(noisy_image, 0, 1)def add_salt_pepper_noise(image, amount=0.01):"""添加椒鹽噪聲"""noisy_image = np.copy(image)# 鹽噪聲salt_mask = np.random.random(image.shape) < amount/2noisy_image[salt_mask] = 1# 椒噪聲pepper_mask = np.random.random(image.shape) < amount/2noisy_image[pepper_mask] = 0return noisy_imagedef jpeg_compression(image, quality=70):"""JPEG壓縮攻擊"""# 轉換為0-255范圍image_uint8 = (image * 255).astype(np.uint8)# JPEG壓縮encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), quality]result, encimg = cv2.imencode('.jpg', image_uint8, encode_param)# 解碼decimg = cv2.imdecode(encimg, 0)return decimg.astype(np.float32) / 255.0def cropping_attack(image, ratio=0.1):"""裁剪攻擊"""h, w = image.shapecrop_h = int(h * ratio)crop_w = int(w * ratio)cropped = image[crop_h:-crop_h, crop_w:-crop_w]# 恢復原始尺寸return cv2.resize(cropped, (w, h))def rotate_image(image, angle=5):"""旋轉攻擊"""h, w = image.shapecenter = (w // 2, h // 2)M = cv2.getRotationMatrix2D(center, angle, 1.0)rotated = cv2.warpAffine(image, M, (w, h))return rotateddef scaling_attack(image, ratio=0.9):"""縮放攻擊"""h, w = image.shapescaled = cv2.resize(image, (int(w * ratio), int(h * ratio)))# 恢復原始尺寸return cv2.resize(scaled, (w, h))def calculate_metrics(original, watermarked, extracted_wm, original_wm=None):"""計算各種評估指標"""metrics = {}# 圖像質量指標metrics['psnr'] = psnr(original, watermarked, data_range=1.0)metrics['ssim'] = ssim(original, watermarked, data_range=1.0)if original_wm is not None and extracted_wm is not None:# 調整尺寸以便比較extracted_resized = cv2.resize(extracted_wm, (original_wm.shape[1], original_wm.shape[0]))# 水印相似度指標metrics['wm_correlation'] = np.corrcoef(original_wm.flatten(), extracted_resized.flatten())[0, 1]metrics['wm_ber'] = calculate_ber(original_wm, extracted_resized)return metricsdef calculate_ber(original, extracted, threshold=0.5):"""計算誤碼率(Bit Error Rate)"""original_binary = (original > threshold).astype(np.int8)extracted_binary = (extracted > threshold).astype(np.int8)if original_binary.shape != extracted_binary.shape:extracted_binary = cv2.resize(extracted_binary, (original_binary.shape[1], original_binary.shape[0]))error_bits = np.sum(original_binary != extracted_binary)total_bits = original_binary.sizereturn error_bits / total_bits

3.7 完整的水印系統類

class RemoteSensingWatermarkSystem:"""完整的遙感圖像水印系統"""def __init__(self, config=None):if config is None:config = {'wavelet': 'haar','level': 2,'base_alpha': 0.05,'max_alpha': 0.2,'block_size': 64,'embed_method': 'adaptive_global'  # 'adaptive_global' or 'block_based'}self.config = configself.embed_params = Nonedef embed(self, host_image, watermark):"""嵌入水印"""# 預處理主機圖像host_processed = preprocess_image(host_image)# 預處理水印watermark_processed = preprocess_watermark(watermark, host_processed.shape)if self.config['embed_method'] == 'adaptive_global':watermarked_image, self.embed_params = embed_watermark_optimized(host_processed, watermark_processed,wavelet=self.config['wavelet'],level=self.config['level'],base_alpha=self.config['base_alpha'])elif self.config['embed_method'] == 'block_based':watermarked_image, self.embed_params = block_based_embedding(host_processed, watermark_processed,block_size=self.config['block_size'],wavelet=self.config['wavelet'])else:raise ValueError("不支持的嵌入方法")# 轉換為0-255范圍輸出watermarked_uint8 = (watermarked_image * 255).astype(np.uint8)return watermarked_uint8def extract(self, watermarked_image, original_image=None):"""提取水印"""if self.embed_params is None:raise ValueError("請先嵌入水印或提供嵌入參數")watermarked_processed = preprocess_image(watermarked_image)if original_image is not None:original_processed = preprocess_image(original_image)else:original_processed = Noneextracted_wm = extract_watermark_optimized(watermarked_processed, self.embed_params, original_processed)# 后處理extracted_processed = postprocess_watermark(extracted_wm, method='threshold')extracted_enhanced = enhance_watermark_quality(extracted_processed)# 轉換為0-255范圍輸出extracted_uint8 = (extracted_enhanced * 255).astype(np.uint8)return extracted_uint8def test_robustness(self, watermarked_image, original_image, attacks=None):"""測試魯棒性"""watermarked_processed = preprocess_image(watermarked_image)original_processed = preprocess_image(original_image)return test_robustness(watermarked_processed, original_processed, self.embed_params, attacks)def evaluate(self, original_image, watermarked_image, original_watermark=None):"""全面評估系統性能"""original_processed = preprocess_image(original_image)watermarked_processed = preprocess_image(watermarked_image)if original_watermark is not None:original_wm_processed = preprocess_watermark(original_watermark, original_processed.shape)else:original_wm_processed = None# 提取水印用于評估extracted_wm = self.extract(watermarked_image, original_image)extracted_processed = preprocess_image(extracted_wm, target_size=original_processed.shape)# 計算指標metrics = calculate_metrics(original_processed, watermarked_processed, extracted_processed, original_wm_processed)return metrics, extracted_wm

3.8 可視化工具

def plot_results(original, watermarked, extracted_wm, original_wm=None, metrics=None):"""可視化結果"""fig, axes = plt.subplots(2, 2, figsize=(12, 10))# 原始圖像axes[0, 0].imshow(original, cmap='gray')axes[0, 0].set_title('原始圖像')axes[0, 0].axis('off')# 含水印圖像axes[0, 1].imshow(watermarked, cmap='gray')axes[0, 1].set_title('含水印圖像')axes[0, 1].axis('off')# 提取的水印axes[1, 0].imshow(extracted_wm, cmap='gray')axes[1, 0].set_title('提取的水印')axes[1, 0].axis('off')# 原始水印（如果有）if original_wm is not None:axes[1, 1].imshow(original_wm, cmap='gray')axes[1, 1].set_title('原始水印')else:axes[1, 1].axis('off')# 添加指標文本if metrics is not None:metrics_text = f"PSNR: {metrics.get('psnr', 0):.2f} dB\nSSIM: {metrics.get('ssim', 0):.4f}"if 'wm_correlation' in metrics:metrics_text += f"\n相關系數: {metrics['wm_correlation']:.4f}"if 'wm_ber' in metrics:metrics_text += f"\n誤碼率: {metrics['wm_ber']:.4f}"plt.figtext(0.5, 0.01, metrics_text, ha='center', fontsize=12, bbox=dict(facecolor='lightgray', alpha=0.5))plt.tight_layout()plt.show()def plot_robustness_test(results, original_wm=None):"""可視化魯棒性測試結果"""n_attacks = len(results)n_cols = 3n_rows = (n_attacks + n_cols - 1) // n_colsfig, axes = plt.subplots(n_rows, n_cols, figsize=(15, 5 * n_rows))axes = axes.flatten()for i, (attack_name, extracted_wm) in enumerate(results.items()):if extracted_wm is not None:axes[i].imshow(extracted_wm, cmap='gray')axes[i].set_title(attack_name)axes[i].axis('off')# 隱藏多余的子圖for j in range(i + 1, len(axes)):axes[j].axis('off')plt.tight_layout()plt.show()

4. 使用示例

# 示例使用代碼
if __name__ == "__main__":# 讀取遙感圖像image_path = "remote_sensing_image.tif"original_image = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)# 創建水印（可以是文本或圖像）watermark_text = "Copyright 2023"# 初始化水印系統watermark_system = RemoteSensingWatermarkSystem({'wavelet': 'haar','level': 2,'base_alpha': 0.05,'embed_method': 'adaptive_global'})# 嵌入水印watermarked_image = watermark_system.embed(original_image, watermark_text)# 提取水印extracted_watermark = watermark_system.extract(watermarked_image, original_image)# 評估性能metrics, extracted_wm = watermark_system.evaluate(original_image, watermarked_image, watermark_text)# 可視化結果plot_results(original_image, watermarked_image, extracted_wm, text_to_image(watermark_text, original_image.shape), metrics)# 測試魯棒性attacks = {'高斯噪聲': lambda img: add_gaussian_noise(img, sigma=0.02),'椒鹽噪聲': lambda img: add_salt_pepper_noise(img, amount=0.03),'高斯模糊': lambda img: cv2.GaussianBlur(img, (5, 5), 0.8),'JPEG壓縮': lambda img: jpeg_compression(img, quality=60),}robustness_results = watermark_system.test_robustness(watermarked_image, original_image, attacks)plot_robustness_test(robustness_results)