Python 批量切片 LabelMe 标注数据（图像+JSON 同步裁剪+重叠+填充）实战教程

在深度学习目标检测、实例分割等任务中，我们经常需要对图像和对应的 LabelMe 标注文件进行切片处理。本教程将教你如何使用 Python 批量裁剪 500×500 的图片和 JSON 标注文件，并支持：

✅ 自定义切片尺寸（x, y）
✅ 自定义重叠比例（z）
✅ JSON 同步裁剪坐标
✅ 不足区域自动黑色填充回 500×500
✅ 批量处理整个文件夹
✅ 输出图片命名规则：原名_行_列.jpg/json

一、准备工作

假设你的原始文件是 LabelMe 标注生成的：

ysl_20210820_01_9.jpg
ysl_20210820_01_9.json

这些文件都放在同一个目录：D:\plant_seg_datasets\woody_plant\lableme
处理后的切片输出到目录：D:\plant_seg_datasets\woody_plant\slice_test

 二、环境依赖

需要有以下几个包

import os
import json
from pathlib import Path
from PIL import Image

没有则需要安装：

pip install pillow

 三、完整 Python 脚本（复制即可运行）

# --------------- 代码开始 ---------------
import os
import json
from pathlib import Path
from PIL import Image

# --------------- USER CONFIG ---------------
slice_width = 256    # 切片宽度（可改）
slice_height = 256   # 切片高度（可改）
z = 0.5              # 重叠比例（0~1）

# 输入路径和输出路径（✅按你提供的设置）
input_folder = Path(r"D:\plant_seg_datasets\woody_plant\lableme")
output_folder = Path(r"D:\plant_seg_datasets\woody_plant\slice_test")
output_folder.mkdir(parents=True, exist_ok=True)

# 填充回 500×500
pad_to = (500, 500)
# -------------------------------------------

def compute_starts(image_size, slice_size, overlap):
    step = max(1, int(round(slice_size * (1 - overlap))))
    starts = []
    if slice_size >= image_size:
        starts = [0]
    else:
        cur = 0
        while cur + slice_size < image_size:
            starts.append(cur)
            cur += step
        last = image_size - slice_size
        if len(starts) == 0 or starts[-1] != last:
            starts.append(last)
    return starts

def crop_and_pad(img, x0, y0, x1, y1, pad_to=(500,500)):
    crop = img.crop((x0, y0, x1, y1))
    Wpad, Hpad = pad_to
    out = Image.new("RGB", (Wpad, Hpad), (0,0,0))
    out.paste(crop, (0,0))
    return out, crop.size

def process_labelme_json_for_crop(jdata, crop_box):
    x0, y0, x1, y1 = crop_box
    new_shapes = []
    for shape in jdata.get("shapes", []):
        pts = shape.get("points", [])
        kept = []
        for (px, py) in pts:
            if x0 <= px < x1 and y0 <= py < y1:
                kept.append([px - x0, py - y0])
        shape_type = shape.get("shape_type", "").lower()
        if shape_type == "polygon":
            min_points = 3
        elif shape_type in ("linestrip","rectangle","line"):
            min_points = 2
        else:
            min_points = 1
        if len(kept) >= min_points:
            new_shape = dict(shape)
            new_shape["points"] = kept
            new_shapes.append(new_shape)
    return new_shapes

def process_one_pair(image_path, json_path, out_dir,
                     slice_w, slice_h, overlap, pad_to=(500,500)):
    img = Image.open(image_path).convert("RGB")
    W, H = img.size
    with open(json_path, "r", encoding="utf8") as f:
        jdata = json.load(f)

    x_starts = compute_starts(W, slice_w, overlap)
    y_starts = compute_starts(H, slice_h, overlap)

    basename = image_path.stem
    out_files = []
    
    for i_row, y0 in enumerate(y_starts):
        for j_col, x0 in enumerate(x_starts):
            x1 = min(x0 + slice_w, W)
            y1 = min(y0 + slice_h, H)

            padded_img, _ = crop_and_pad(img, x0, y0, x1, y1, pad_to=pad_to)

            new_j = dict(jdata)
            new_shapes = process_labelme_json_for_crop(jdata, (x0,y0,x1,y1))
            new_j["shapes"] = new_shapes
            new_j["imagePath"] = f"{basename}_{i_row}_{j_col}.jpg"
            new_j["imageHeight"] = pad_to[1]
            new_j["imageWidth"] = pad_to[0]

            out_img_path = out_dir / f"{basename}_{i_row}_{j_col}.jpg"
            out_json_path = out_dir / f"{basename}_{i_row}_{j_col}.json"

            padded_img.save(out_img_path, format="JPEG", quality=95)
            with open(out_json_path, "w", encoding="utf8") as f:
                json.dump(new_j, f, ensure_ascii=False, indent=2)
            
            out_files.append((out_img_path, out_json_path))

    return out_files

def main():
    jpgs = sorted(input_folder.glob("*.jpg"))
    pairs = []
    for j in jpgs:
        js = j.with_suffix(".json")
        if js.exists():
            pairs.append((j, js))
        else:
            print("Warning: JSON not found for", j.name)

    total = 0
    for img_p, json_p in pairs:
        outs = process_one_pair(
            img_p, json_p,
            output_folder,
            slice_width,
            slice_height,
            z,
            pad_to=pad_to
        )
        total += len(outs)

    print(f"✅ 处理完成！共处理 {len(pairs)} 组图片，生成 {total} 个切片到：{output_folder}")

if __name__ == "__main__":
    main()
# --------------- 代码结束 ---------------

五、输出文件命名格式

ysl_20210820_01_9_0_0.jpg
ysl_20210820_01_9_0_0.json
ysl_20210820_01_9_1_0.jpg
ysl_20210820_01_9_1_0.json
...
全部保存在：D:\plant_seg_datasets\woody_plant\slice_test

六、效果特点

✔ 保留部分落入切片区域的标注
✔ 支持重叠切片
✔ 自动补黑边回 500×500
✔ 支持批量目录处理
✔ LabelMe JSON 结构不变，可直接继续训练使用

七、总结

这段脚本非常适用于：

• 目标检测 / 实例分割数据增强

• LabelMe 标注数据预处理

• 大图切片+小图标注同步

• 图像+JSON 批量裁剪保真