为何部分图像能显示Bounding Box，部分无法显示？（Keras模型）

我参考Keras官方目标检测模型评估章节，用预训练模型绘制Bounding Box，但相同代码下，部分图像能正常显示框，部分图像有预测标签却看不到Bounding Box。已知图像显示黑色是数据类型问题，无需关注，求代码调试帮助。

我的代码：

model_load = load_model('stenosis_model_transfer.h5')
# 计算IoU（交并比）
def bounding_box_intersection_over_union(box_predicted, box_truth):
    # 获取边界框交集的坐标
    top_x_intersect = max(box_predicted[0], box_truth[0])
    top_y_intersect = max(box_predicted[1], box_truth[1])
    bottom_x_intersect = min(box_predicted[2], box_truth[2])
    bottom_y_intersect = min(box_predicted[3], box_truth[3])

    # 计算交集面积
    intersection_area = max(0, bottom_x_intersect - top_x_intersect + 1) * max(
        0, bottom_y_intersect - top_y_intersect + 1
    )

    # 计算预测框和真实框的面积
    box_predicted_area = (box_predicted[2] - box_predicted[0] + 1) * (
        box_predicted[3] - box_predicted[1] + 1
    )
    box_truth_area = (box_truth[2] - box_truth[0] + 1) * (
        box_truth[3] - box_truth[1] + 1
    )

    # 返回IoU值
    return intersection_area / float(
        box_predicted_area + box_truth_area - intersection_area
    )


i, mean_iou = 0, 0

# 对比测试集中前5张图像的结果
for input_image in test_images[:5]:
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 15))
    im = input_image
    plt.tight_layout()

    # 显示图像
    ax1.imshow(im.astype("uint8"))
    ax2.imshow(im.astype("uint8"))

    input_image = cv2.resize(
        input_image, (224, 224), interpolation=cv2.INTER_AREA
    )
    input_image = np.expand_dims(input_image, axis=0)
    preds = model_load.predict(input_image)[0]

    (h, w) = (im).shape[0:2]

    top_left_x, top_left_y = int(preds[0] * w), int(preds[1] * h)
    bottom_right_x, bottom_right_y = int(preds[2] * w), int(preds[3] * h)

    box_predicted = [top_left_x, top_left_y, bottom_right_x, bottom_right_y]
    # 创建预测边界框
    rect = patches.Rectangle(
        (top_left_x, top_left_y),
        bottom_right_x - top_left_x,
        bottom_right_y - top_left_y,
        facecolor="none",
        edgecolor="red",
        linewidth=1,
    )
    ax1.add_patch(rect)
    ax1.set_xlabel(
        "Predicted: "
        + str(top_left_x)
        + ", "
        + str(top_left_y)
        + ", "
        + str(bottom_right_x)
        + ", "
        + str(bottom_right_y)
    )

    # 获取真实框坐标
    top_left_x, top_left_y = int(test_targets[i][0] * w), int(test_targets[i][1] * h)
    bottom_right_x, bottom_right_y = int(test_targets[i][2] * w), int(test_targets[i][3] * h)
    box_truth = top_left_x, top_left_y, bottom_right_x, bottom_right_y

    mean_iou += bounding_box_intersection_over_union(box_predicted, box_truth)
    # 创建真实边界框
    rect = patches.Rectangle(
        (top_left_x, top_left_y),
        bottom_right_x - top_left_x,
        bottom_right_y - top_left_y,
        facecolor="none",
        edgecolor="red",
        linewidth=1,
    )
    ax2.add_patch(rect)
    ax2.set_xlabel(
        "Target: "
        + str(top_left_x)
        + ", "
        + str(top_left_y)
        + ", "
        + str(bottom_right_x)
        + ", "
        + str(bottom_right_y)
        + "\nIoU: "
        + str(bounding_box_intersection_over_union(box_predicted, box_truth))
    )
    i = i + 1

plt.show()

调试建议 #

• 修复边界框尺寸合法性问题：当模型预测的右下角坐标小于左上角坐标时，bottom_right_x - top_left_x或bottom_right_y - top_left_y会为负数，导致patches.Rectangle生成不可见的框。添加坐标修正逻辑：

  # 修正预测框坐标
  width = bottom_right_x - top_left_x
  height = bottom_right_y - top_left_y
  if width < 0:
      top_left_x, bottom_right_x = bottom_right_x, top_left_x
      width = -width
  if height < 0:
      top_left_y, bottom_right_y = bottom_right_y, top_left_y
      height = -height
  # 重新创建矩形
  rect = patches.Rectangle((top_left_x, top_left_y), width, height, facecolor="none", edgecolor="red", linewidth=2)
  

• 验证模型输出范围：模型预测的坐标值可能超出[0,1]范围，缩放后会超出图像边界。添加打印检查preds值，若超出则做裁剪：

  print(f"Predicted raw values: {preds}")
  preds = np.clip(preds, 0.0, 1.0)  # 将预测值限制在0-1之间
  

• 确认数据索引匹配：检查test_images和test_targets的遍历索引是否严格对应，避免因索引错位导致目标框坐标异常。
• 提升框线可见性：将linewidth从1改为2或3，避免因线宽过细在部分图像中被忽略。

内容的提问来源于stack exchange，提问作者NevMthw