### Faster R-CNN with ResNet-101 and Feature Pyramid Network (FPN) 以下是基于 PyTorch 的 Faster R-CNN 使用 ResNet-101 和特征金字塔网络（Feature Pyramid Network, FPN）实现多层特征融合的代码示例。此代码展示了如何构建一个多级 ROI Align 层并将其应用于目标检测任务。 #### 构建 Faster R-CNN 模型 ```python import torch from torchvision.models.detection import faster_rcnn from torchvision.models.detection.rpn import AnchorGenerator from torchvision.ops import MultiScaleRoIAlign from torchvision.models.detection.faster_rcnn import FastRCNNConvFCHead, TwoMLPHead from torchvision.models.resnet import resnet101 class CustomFasterRCNN(torch.nn.Module): def __init__(self, num_classes): super(CustomFasterRCNN, self).__init__() # 加载预训练的ResNet-101模型，并修改其结构以适配FPN backbone = resnet101(pretrained=True) layers_to_extract = ["layer1", "layer2", "layer3", "layer4"] in_channels_list = [backbone.layer1[-1].conv2.out_channels, backbone.layer2[-1].conv2.out_channels, backbone.layer3[-1].conv2.out_channels, backbone.layer4[-1].conv2.out_channels] # 创建FPN模块 out_channels_fpn = 256 self.backbone = create_feature_pyramid_network(backbone, layers_to_extract, out_channels_fpn) # 定义ROI Pooling层 roi_pooler = MultiScaleRoIAlign( featmap_names=layers_to_extract, output_size=7, sampling_ratio=2 ) # 定义Box Head部分 resolution = roi_pooler.output_size[0] representation_size = 1024 box_head = FastRCNNConvFCHead( (out_channels_fpn, resolution, resolution), [(resolution, resolution)], [representation_size], norm_layer=torch.nn.BatchNorm2d ) box_predictor = fast_rcnn.FastRCNNPredictor(representation_size, num_classes) # 组合到完整的Faster R-CNN框架中 rpn_anchor_generator = AnchorGenerator(sizes=((32,), (64,), (128,), (256,), (512,)), aspect_ratios=((0.5, 1.0, 2.0),)) self.model = faster_rcnn.FasterRCNN( backbone=self.backbone, num_classes=num_classes, rpn_anchor_generator=rpn_anchor_generator, box_roi_pool=roi_pooler, box_head=box_head, box_predictor=box_predictor ) def create_feature_pyramid_network(base_model, layers_to_extract, out_channels): """ 将基础骨干网络转换为支持FPN的结构。 :param base_model: 骨干网络实例 :param layers_to_extract: 列表形式，指定要提取哪些层作为FPN输入 :param out_channels: 输出通道数统一设置为目标值 :return: 返回带有FPN功能的BackboneWithFPN类实例 """ from torchvision.models.feature_extraction import create_feature_extractor feature_extractor = create_feature_extractor(base_model, return_nodes={name: name for name in layers_to_extract}) backbone_with_fpn = BackboneWithFPN(feature_extractor, layers_to_extract, out_channels=out_channels) return backbone_with_fpn class BackboneWithFPN(torch.nn.Module): """ 自定义FPN模块，用于处理不同尺度的特征图。 """ def __init__(self, body, layers_to_extract, out_channels): super(BackboneWithFPN, self).__init__() self.body = body self.layers_to_extract = layers_to_extract # 添加额外的卷积层来生成更高层次的特征图 self.extra_blocks = ExtraFPNBlock() self.fpn_convs = torch.nn.ModuleList([torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1) for _ in range(len(layers_to_extract))]) def forward(self, x): features = self.body(x) results = [] prev = None for i, layer_name in enumerate(reversed(self.layers_to_extract)): current_feat = features[layer_name] if prev is not None: upsampled_prev = torch.nn.functional.interpolate(prev, size=current_feat.shape[-2:], mode='nearest') current_feat += upsampled_prev conv_result = self.fpn_convs[i](current_feat) results.insert(0, {layer_name: conv_result}) # 插入到前面保持顺序一致性 prev = conv_result results.append(self.extra_blocks(results[-1])) return {"features": dict(sum(results, {}))} ``` 以上代码实现了以下几个核心组件： 1. **自定义 BackBone**: 基于 `resnet101` 并通过 `create_feature_pyramid_network` 方法创建了一个支持多层特征融合的 FPND[^1]。 2. **Multi-Scale RoI Align Layer**: 使用 `MultiScaleRoIAlign` 对多个尺度上的区域建议进行池化操作[^1]。 3. **Fast RCNN Predictor**: 结合分类器和回归器完成最终的目标类别预测与边界框调整[^4]。 这种设计充分利用了 ResNet-101 各个阶段的不同分辨率特征图，从而提高了小物体检测的能力以及整体性能[^3]。 ---