### PADIM Anomaly Detection 的 PyTorch 实现 PADIM (Padding-based Image Manipulation) 是一种基于无监督学习的方法，用于异常检测。它通过提取图像特征并计算重建误差来识别异常区域。以下是使用单张训练图片和单张测试图片的完整代码示例。 #### 依赖库安装 在运行代码之前，请确保已安装所需的 Python 库： ```bash pip install torch torchvision matplotlib numpy scikit-learn ``` #### 完整代码实现 以下是一个完整的 PyTorch 实现： ```python import torch from torch import nn from torchvision import transforms from PIL import Image import numpy as np import matplotlib.pyplot as plt class FeatureExtractor(nn.Module): def __init__(self, model_name='resnet18', pretrained=True): super(FeatureExtractor, self).__init__() resnet = torch.hub.load('pytorch/vision:v0.10.0', model_name, pretrained=pretrained) layers = list(resnet.children())[:-2] # Remove the last two layers (avgpool and fc) self.feature_extractor = nn.Sequential(*layers) def forward(self, x): features = self.feature_extractor(x) return features def load_image(image_path, transform=None): img = Image.open(image_path).convert('RGB') if transform is not None: img = transform(img) return img.unsqueeze(0) # Add batch dimension # Define transformations transform = transforms.Compose([ transforms.Resize((224, 224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) # Load a single training image train_img_path = 'path_to_single_training_image.jpg' train_img_tensor = load_image(train_img_path, transform=transform) # Initialize feature extractor device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = FeatureExtractor().to(device) model.eval() with torch.no_grad(): train_features = model(train_img_tensor.to(device)) # Flatten the features to compute statistics train_features_flattened = train_features.view(-1).cpu().numpy() mean_train = np.mean(train_features_flattened) std_train = np.std(train_features_flattened) # Load a single test image test_img_path = 'path_to_single_test_image.jpg' test_img_tensor = load_image(test_img_path, transform=transform) with torch.no_grad(): test_features = model(test_img_tensor.to(device)) # Compute anomaly score based on Mahalanobis distance or simple z-score test_features_flattened = test_features.view(-1).cpu().numpy() z_scores = abs((test_features_flattened - mean_train) / std_train) anomaly_score = np.max(z_scores) print(f"Anomaly Score: {anomaly_score}") # Visualize results plt.figure(figsize=(10, 5)) plt.subplot(1, 2, 1) plt.title("Training Image") plt.imshow(Image.open(train_img_path)) plt.axis('off') plt.subplot(1, 2, 2) plt.title("Test Image") plt.imshow(Image.open(test_img_path)) plt.axis('off') plt.show() ``` 上述代码实现了以下几个功能： 1. 使用 ResNet 提取图像特征[^1]。 2. 计算训练图像的均值和标准差作为统计量。 3. 对于测试图像，计算其特征向量并与训练图像的统计量比较，得到异常分数。 4. 可视化输入图像以便直观理解结果。 #### 关键点说明 - **模型选择**：此处选择了 `ResNet` 作为特征提取器，也可以尝试其他预训练网络如 `VGG` 或 `EfficientNet`。 - **数据标准化**：为了提高性能，通常会对图像进行归一化处理[^3]。 - **异常评分机制**：可以进一步优化评分方法，例如引入协方差矩阵或更复杂的距离度量算法[^2]。 --- ###