Grad-CAMdpi

### 使用 Grad-CAM 生成高质量 DPI 的热力图 为了生成高分辨率的热力图，可以采用 Grad-CAM 技术并调整图像处理参数来确保最终输出具有较高的 DPI (dots per inch)[^1]。 #### 准备工作 首先需要安装必要的库文件。这通常包括 PyTorch 或 TensorFlow 这样的深度学习框架以及 OpenCV 和 Matplotlib 等用于图像操作和可视化的工具包[^2]。 ```bash pip install torch torchvision opencv-python matplotlib numpy ``` #### 加载预训练模型与输入图片 加载一个已经过良好训练的目标检测或分类网络，并准备好待分析的测试样本。这里假设使用的是 ResNet50 模型作为例子： ```python import torch from torchvision import models, transforms from PIL import Image import cv2 import numpy as np import matplotlib.pyplot as plt device = 'cuda' if torch.cuda.is_available() else 'cpu' model = models.resnet50(pretrained=True).to(device) model.eval() transform = transforms.Compose([ transforms.Resize((224, 224)), transforms.ToTensor(), ]) img_path = "path_to_your_image.jpg" image = Image.open(img_path).convert('RGB') input_tensor = transform(image).unsqueeze(0).to(device) ``` #### 实现 Grad-CAM 类 定义 `GradCAM` 类以便于后续调用其方法获取特征映射及其对应的梯度信息： ```python class GradCAM: def __init__(self, model, target_layer): self.model = model self.target_layer = target_layer # Register a hook to save gradients of the activations. self.gradients = [] self.activations = [] handle_grad = target_layer.register_backward_hook(self._store_gradients) handle_actv = target_layer.register_forward_hook(self._store_activations) def _store_gradients(self, module, grad_input, grad_output): self.gradients.append(grad_output[0].detach()) def _store_activations(self, module, input, output): self.activations.append(output.detach()) @staticmethod def normalize_cam(cam): cam -= np.min(cam) cam /= (np.max(cam) + 1e-8) return cam def compute_heatmap(self, img_tensor, class_idx=None): logits = self.model(img_tensor) if not isinstance(class_idx,int): _, predicted_class = torch.max(logits.data.cpu(), 1) class_idx=predicted_class.item() one_hot = torch.zeros_like(logits).to(device) one_hot[:, class_idx] = 1. self.model.zero_grad() loss = torch.sum(one_hot * logits) loss.backward(retain_graph=True) weights = F.adaptive_avg_pool2d(self.gradients[-1], 1) weighted_sum = sum([a * w for a,w in zip(self.activations[-1],weights)]) heatmap = weighted_sum.squeeze().mean(dim=0).cpu().numpy() normalized_heatmap = self.normalize_cam(heatmap) return normalized_heatmap ``` #### 绘制高DPI热力图 最后一步就是利用上述类实例化对象计算得到热力图数据，并通过适当的方式将其可视化出来。注意设置保存图形时指定较大的尺寸和更高的 dpi 值以获得更清晰的结果: ```python target_layers = list(model.children())[:-2][-1] gradcam = GradCAM(model=model,target_layer=target_layers) heat_map = gradcam.compute_heatmap(input_tensor) fig, ax = plt.subplots(figsize=(10, 10), dpi=300) ax.imshow(cv2.resize(np.float32(image),(224,224)), alpha=.7) ax.imshow(heat_map,cmap='jet',alpha=.5) plt.axis('off') output_file_name="high_quality_heatmap.png" plt.savefig(output_file_name,bbox_inches='tight',pad_inches=0,dpi=300) print(f"Heatmap saved successfully at {output_file_name}") ```