Compare commits

..

5 Commits

Author SHA1 Message Date
douboer
278f1e0f69 update at 2026-03-28 19:36:19 2026-03-28 19:36:19 +08:00
douboer
bd0e4a660a update at 2026-03-28 17:07:01 2026-03-28 17:07:01 +08:00
douboer
d7a198fd58 update at 2026-03-28 16:53:59 2026-03-28 16:53:59 +08:00
douboer
76caafc580 update at 2026-03-28 16:50:06 2026-03-28 16:50:06 +08:00
douboer
a1b0d6731c update at 2026-03-28 16:46:40 2026-03-28 16:46:40 +08:00
14 changed files with 1870 additions and 321 deletions

BIN
.DS_Store vendored

Binary file not shown.

92
.gitignore vendored
View File

@@ -1,36 +1,72 @@
# Build and Release Folders # Python cache and bytecode
bin-debug/ __pycache__/
bin-release/ *.py[cod]
[Oo]bj/ *$py.class
[Bb]in/
# Other files and folders # Native extensions
.settings/ *.so
node_modules/ # Packaging
public/ .Python
myenv/ build/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
share/python-wheels/
var/
wheels/
*.egg-info/
.installed.cfg
*.egg
MANIFEST
# Virtual environments
.venv/
venv/ venv/
package.json env/
package-lock.json ENV/
myenv/
.python-version
# Executables # Test, coverage, type check, lint caches
*.swf .coverage
*.air .coverage.*
*.ipa .hypothesis/
*.apk .mypy_cache/
.nox/
.pyre/
.pytest_cache/
.ruff_cache/
.tox/
htmlcov/
coverage.xml
nosetests.xml
*.cover
*.py,cover
# Project files, i.e. `.project`, `.actionScriptProperties` and `.flexProperties` # Notebook
# should NOT be excluded as they contain compiler settings and other important .ipynb_checkpoints/
# information for Eclipse / Flash Builder.
*.mov # Environment files
*.mp4 .env
.env.*
*.local
# IDE and OS files
.DS_Store
.idea/
.settings/
.vscode/
# Project runtime files
images/
output/
# Local certificates and keys
cert/nginx/*.crt cert/nginx/*.crt
cert/nginx/*.key cert/nginx/*.key
*.jpg
*.heic
*.jpeg
*.png

426
README.md
View File

@@ -1,208 +1,294 @@
# 图片去背景工具 # 图片去背景工具
使用rembg库实现的Python去背景工具 用于书画、篆刻作品的去背景与去主体补背景脚本。
默认模式优先走书画专用前景提取逻辑:
- 对墨色笔画做局部背景校正与明暗差分
- 对印章做更严格的红色检测
- 在粗掩码内进一步细化 alpha尽量减少字边残留底色
- 仅在需要时才回退到 `rembg`
## 功能概览
- 单张图片去背景,输出透明 PNG
- 批量处理单层目录中的图片
- 支持 `artwork``auto``rembg` 三种前景提取模式
- 支持 `calligraphy``seal``auto` 三种书画类型
- 支持 AOT-GAN 去主体补背景
- 支持常见格式:`jpg``jpeg``png``bmp``webp`
- 安装 `pillow-heif` 后可读取 `heic` / `heif`
## 依赖
项目当前依赖见 [requirements.txt](/Users/gavin/removeback/requirements.txt)
```txt
rembg[gpu]
pillow
pillow-heif
opencv-python
```
如果要使用 AOT-GAN 补背景,还需要额外安装 `torch` / `torchvision`,并准备 AOT-GAN 代码目录与权重文件。
## 环境准备
推荐使用现有虚拟环境:
```bash
source ~/venv/bin/activate
```
安装依赖:
```bash
pip install -r requirements.txt
```
如果你需要启用去主体补背景:
```bash
pip install torch torchvision
```
## 快速开始 ## 快速开始
```bash 处理默认 `images/` 目录中的图片,结果输出到 `output/`
# 激活虚拟环境
source ~/venv/bin/activate
# 使用默认参数处理images文件夹 ```bash
python remove_background.py python remove_background.py
```
# 处理单个文件 处理单张图片:
```bash
python remove_background.py input.jpg output.png python remove_background.py input.jpg output.png
# 查看所有参数-m, --model)
不同模型适用于不同场景:
| 模型名称 | 大小 | 适用场景 | 推荐度 |
|---------|------|---------|--------|
| **isnet-general-use** | 179MB | 通用场景 | ⭐⭐⭐⭐⭐ 默认推荐 |
| birefnet-general | 250MB | 通用场景,质量更高 | ⭐⭐⭐⭐⭐ |
| birefnet-portrait | 250MB | 人像专用 | ⭐⭐⭐⭐⭐ |
| u2net | 176MB | 经典通用模型 | ⭐⭐⭐⭐ |
| u2netp | 4.7MB | 快速处理 | ⭐⭐⭐ |
| u2net_human_seg | 176MB | 人物分割 | ⭐⭐⭐⭐ |
| isnet-anime | 179MB | 动漫角色 | ⭐⭐⭐⭐ |
| silueta | 43MB | 精简快速 | ⭐⭐⭐ |
**使用示例**:
```bash
# 使用默认模型
python remove_background.py input.jpg
# 使用人像专用模型
python remove_background.py input.jpg output.png -m birefnet-portrait
# 使用快速模型
python remove_background.py input.jpg output.png -m u2netp
``
```bash
ls -lh ~/.u2net/
``` ```
### 模型大小参考 处理指定目录:
| 模型名称 | 文件大小 | 特点 |
|---------|---------|------|
| u2net | 176MB | 通用模型 |
| u2netp | 4.7MB | 轻量级,速度快 |
| isnet-general-use | 179MB | 新一代通用,推荐 |
| birefnet-general | ~250MB | 最新通用模型 |
| birefnet-portrait | ~250MB | 人像专用 |
### 手动下载模型(网络问题时)
```bash ```bash
# 创建目录 python remove_background.py my_images/ my_output/
mkdir -p ~/.u2net/
# 下载指定模型以isnet-general-use为例
curl -L "https://github.com/danielgatis/rembg/releases/download/v0.0.0/isnet-general-use.onnx" \
-o ~/.u2net/isnet-general-use.onnx
``` ```
### 清理模型缓存 指定为印章场景:
```bash ```bash
# 删除所有已下载的模型 python remove_background.py input.jpg output.png --artwork-type seal
rm -rf ~/.u2net/
# 删除特定模型
rm ~/.u2net/u2net.onnx
``` ```
## 可调整参数说明 强制使用 `rembg`
### 1. 模型选择 (model_name) ```bash
python remove_background.py input.jpg output.png --foreground-mode rembg -m isnet-general-use
不同模型适用于不同场景:
- **u2net** (默认): 通用模型,适合大多数场景
- **u2netp**: 轻量版,速度更快但精度稍低
- **u2net_human_seg**: 专门用于人物分割
- **silueta**: 精简版u2net (43MB),速度快
- **isnet-general-use**: 新一代通用模型,效果可能更好
- **isnet-anime**: 专门用于动漫角色
- **birefnet-general**: 最新的通用模型,推荐尝试
- **birefnet-portrait**: 专门用于人像
- **birefnet-general-lite**: 轻量版birefnet
**建议**: 如果u2net效果不好试试 `isnet-general-use``birefnet-general`
### 2. Alpha Matting 参数
Alpha Matting 是后处理步骤,可以显著改善边缘质量,特别是头发、毛发等细节。
#### alpha_matting (True/False)
- **作用**: 是否启用alpha matting
- **默认**: False
- **建议**: 如果边缘不自然,启用此选项
#### alpha_matting_foreground_threshold (0-255)
- **作用**: 前景阈值,控制哪些区域被认为是前景
- **默认**: 240
- **调整建议**:
- 值越大(如270): 保留更多细节,但可能保留一些背景
- 值越小(如210): 去除更彻底,但可能丢失细节
- 如果前景被过度去除,增加此值
- 如果背景残留太多,减小此值
#### alpha_matting_background_threshold (0-255)
- **作用**: 背景阈值,控制哪些区域被认为是背景
- **默认**: 10
- **调整建议**:
- 值越大(如20-30): 去除背景更彻底
- 值越小(如5): 保留更多过渡区域
- 如果背景残留,增加此值
#### alpha_matting_erode_size (像素)
- **作用**: 侵蚀大小,用于平滑边缘
- **默认**: 10
- **调整建议**:
- 值越大(如15-20): 边缘更平滑,但可能损失细节
- 值越小(如5-8): 保留更多细节,但边缘可能不够平滑
### 3. Mask后处理 (post_process_mask)
- **作用**: 对mask进行额外的后处理
- **默认**: False
- **建议**: 可以尝试启用看效果是否改善
## 常见问题解决
### 问题1: 前景被过度去除
**解决方案**:
```python
alpha_matting = True
alpha_matting_foreground_threshold = 270 # 增加此值
alpha_matting_background_threshold = 10 # 保持较小
``` ```
### 问题2: 背景残留太多 查看完整参数:
**解决方案**:
```python ```bash
alpha_matting = True python remove_background.py -h
alpha_matting_foreground_threshold = 240 # 保持默认或减小
alpha_matting_background_threshold = 20 # 增加此值
post_process_mask = True # 启用后处理
``` ```
### 问题3: 边缘不自然、有锯齿 ## 输出规则
**解决方案**:
```python - 单张图片默认输出为 `*_nobg.png`
alpha_matting = True - 处理目录时,输出文件写入你指定的输出目录
alpha_matting_erode_size = 15 # 增加平滑程度 - 脚本内置的目录批处理只扫描输入目录的第一层文件,不递归子目录
- 如果启用 `--remove-subject`,输出文件名改为 `*_bgfill.<原扩展>``*_bgfill.jpg`
## 前景提取模式
### `artwork`
默认模式。优先适用于书法、国画、篆刻等纸本图像。
### `auto`
先尝试书画专用掩码;如果结果明显不可信,再回退到 `rembg`
### `rembg`
强制使用通用抠图模型。适合非书画类图片,或书画专用规则不适合的特殊样本。
## 书画类型
### `auto`
自动兼容书法与印章。
### `calligraphy`
更偏重墨色笔画、灰黑色文字。
### `seal`
更偏重红章、篆刻印记。
## 常用参数
### 书画专用参数
- `--foreground-mode`
- `--artwork-type`
- `--artwork-max-size`
建议:
- 大图先尝试 `--artwork-max-size 1600`
- 超大图如果速度较慢,可降低到 `1200``1000`
- 红章较多的图片优先试 `--artwork-type seal`
- 纯墨迹优先试 `--artwork-type calligraphy`
### rembg 相关参数
这些参数只在 `--foreground-mode rembg``auto` 回退到 `rembg` 时生效:
- `-m, --model`
- `-a, --alpha-matting`
- `-ft, --foreground-threshold`
- `-bt, --background-threshold`
- `-es, --erode-size`
- `-p, --post-process`
当前支持的 `rembg` 模型包括:
- `u2net`
- `u2netp`
- `u2net_human_seg`
- `silueta`
- `isnet-general-use`
- `isnet-anime`
- `birefnet-general`
- `birefnet-general-lite`
- `birefnet-portrait`
- `birefnet-dis`
- `birefnet-hrsod`
- `birefnet-cod`
- `birefnet-massive`
### 去主体补背景参数
启用:
```bash
python remove_background.py input.jpg output.jpg --remove-subject --aot-pretrain experiments/your_model.pt
``` ```
### 问题4: 毛发、头发细节丢失 常用参数:
**解决方案**:
```python - `--aot-root`
model_name = "birefnet-portrait" # 使用人像专用模型 - `--aot-pretrain`
alpha_matting = True - `--aot-device`
alpha_matting_foreground_threshold = 270 # 增加以保留细节 - `--aot-block-num`
alpha_matting_erode_size = 5 # 减小以保留细节 - `--aot-rates`
- `--aot-crop`
- `--aot-crop-pad`
- `--aot-max-size`
- `--aot-noise-prefill`
- `--aot-noise-strength`
- `--mask-dilate`
- `--mask-blur`
- `--mask-threshold`
- `--edge-grow`
- `--save-mask`
- `--black-subject`
- `--black-threshold`
- `--gray-subject`
- `--gray-saturation-threshold`
- `--gray-value-threshold`
- `--feather`
- `--feather-radius`
一个偏保守的示例:
```bash
python remove_background.py "images/inpaint/IMG_9259 2.JPG" output.jpg \
--remove-subject \
--foreground-mode artwork \
--artwork-type auto \
--aot-pretrain experiments/G0000000.pt \
--aot-crop \
--aot-crop-pad 64 \
--aot-max-size 900 \
--aot-noise-prefill \
--aot-noise-strength 1.0
``` ```
## 推荐配置 ## 典型用法
### 配置1: 高质量人像 书法图去背景:
```python
model_name = "birefnet-portrait" ```bash
alpha_matting = True python remove_background.py input.jpg output.png \
alpha_matting_foreground_threshold = 260 --foreground-mode artwork \
alpha_matting_background_threshold = 15 --artwork-type calligraphy
alpha_matting_erode_size = 10
post_process_mask = True
``` ```
### 配置2: 通用高质量 印章图去背景:
```python
model_name = "birefnet-general" ```bash
alpha_matting = True python remove_background.py input.jpg output.png \
alpha_matting_foreground_threshold = 250 --foreground-mode artwork \
alpha_matting_background_threshold = 12 --artwork-type seal
alpha_matting_erode_size = 10
post_process_mask = True
``` ```
### 配置3: 快速处理 通用图片走 `rembg`
```python
model_name = "u2netp" ```bash
alpha_matting = False python remove_background.py input.jpg output.png \
post_process_mask = False --foreground-mode rembg \
--model birefnet-general
``` ```
## 测试不同参数 ## 调参建议
建议按以下顺序调整: 如果背景没有去掉:
1. 先尝试不同的模型 - 先尝试 `--artwork-type calligraphy`
2. 启用alpha_matting - 再尝试 `--foreground-mode auto`
3. 调整foreground_threshold和background_threshold - 非书画图直接改用 `--foreground-mode rembg`
4. 最后调整erode_size
每次修改后运行脚本,对比结果。 如果字边仍有底色:
- 先确认原图是否有严重纸纹、阴影或压缩噪声
- 降低 `--artwork-max-size` 可能更快,但通常不利于细节
- 对超大图可以保留 `1600`,必要时单独抽样检查结果
如果去主体补背景不自然:
- 开启 `--aot-crop`
- 增加 `--aot-crop-pad`
- 尝试 `--aot-noise-prefill`
- 减小 `--mask-dilate``--mask-blur`
## 校验命令
项目当前可用的检查命令:
```bash
~/venv/bin/python -m pytest -q
~/venv/bin/python -m mypy remove_background.py tests/test_remove_background.py
~/venv/bin/python -m ruff check remove_background.py tests/test_remove_background.py
~/venv/bin/python -m ruff format remove_background.py tests/test_remove_background.py
```
说明:
- `mypy``ruff` 的项目配置见 [pyproject.toml](/Users/gavin/removeback/pyproject.toml)
- `images/``output/` 已在静态检查配置里排除
## 已知限制
- CLI 自带的目录批处理不递归子目录
- 书画专用规则仍然可能受极端纸色、重阴影、扫描边框影响
- 某些透明 PNG 在图片预览器里会显示白底或棋盘底,这是预览器合成效果,不代表 alpha 一定有问题
## 仓库结构
```txt
remove_background.py 主脚本
tests/test_remove_background.py
requirements.txt
pyproject.toml
images/ 输入样例
output/ 输出目录
experiments/ AOT-GAN 权重示例
```

Binary file not shown.

BIN
experiments/D0000000.pt Normal file

Binary file not shown.

BIN
experiments/G0000000.pt Normal file

Binary file not shown.

BIN
experiments/O0000000.pt Normal file

Binary file not shown.

Binary file not shown.

16
pyproject.toml Normal file
View File

@@ -0,0 +1,16 @@
[tool.mypy]
python_version = "3.12"
ignore_missing_imports = true
exclude = '^(generative_inpainting|images|output)/'
[tool.ruff]
exclude = [
"AOT-GAN-for-Inpainting",
"generative_inpainting",
"images",
"output",
"__pycache__",
".venv",
"venv",
".git",
]

File diff suppressed because it is too large Load Diff

View File

@@ -1,3 +1,4 @@
rembg[gpu] rembg[gpu]
pillow pillow
pillow-heif pillow-heif
opencv-python

1
runit.sh Symbolic link
View File

@@ -0,0 +1 @@
/Users/gavin/tools/runit.sh

View File

@@ -0,0 +1,203 @@
import importlib
import sys
import tempfile
import types
import unittest
from pathlib import Path
def _import_remove_background():
if "remove_background" in sys.modules:
del sys.modules["remove_background"]
rembg_stub = types.ModuleType("rembg")
rembg_stub.remove = lambda *args, **kwargs: None
rembg_stub.new_session = lambda *args, **kwargs: None
sys.modules["rembg"] = rembg_stub
return importlib.import_module("remove_background")
class RemoveBackgroundTests(unittest.TestCase):
def setUp(self):
self.mod = _import_remove_background()
def _make_paper_image(self, width=160, height=120):
rng = self.mod.np.random.default_rng(0)
base = self.mod.np.full((height, width, 3), 230, dtype=self.mod.np.uint8)
noise = rng.integers(-8, 9, size=(height, width, 3), dtype=self.mod.np.int16)
image = self.mod.np.clip(base.astype(self.mod.np.int16) + noise, 0, 255).astype(
self.mod.np.uint8
)
return image
def test_alpha_to_mask_threshold(self):
alpha = self.mod.np.array([[0, 9, 10, 255]], dtype=self.mod.np.uint8)
mask = self.mod._alpha_to_mask(alpha, threshold=10)
self.assertEqual(mask.shape, alpha.shape)
self.assertEqual(mask[0, 0], 0)
self.assertEqual(mask[0, 1], 0)
self.assertEqual(mask[0, 2], 0)
self.assertEqual(mask[0, 3], 255)
def test_prepare_mask_outputs_binary(self):
mask = self.mod.np.zeros((5, 5), dtype=self.mod.np.uint8)
mask[2, 2] = 255
mask_hard, mask_used = self.mod._prepare_mask(mask, mask_dilate=0, mask_blur=3)
self.assertEqual(mask_hard.shape, mask.shape)
self.assertEqual(mask_used.shape, mask.shape)
self.assertTrue(self.mod.np.all(self.mod.np.isin(mask_used, [0, 255])))
def test_resolve_aot_paths_defaults(self):
aot_root, aot_pretrain = self.mod._resolve_aot_paths(
"AOT-GAN-for-Inpainting", None
)
self.assertTrue(str(aot_root).endswith("AOT-GAN-for-Inpainting"))
self.assertIsNone(aot_pretrain)
def test_resolve_aot_paths_relative_pretrain(self):
aot_root, aot_pretrain = self.mod._resolve_aot_paths(
"AOT-GAN-for-Inpainting", "experiments/foo.pt"
)
self.assertTrue(str(aot_root).endswith("AOT-GAN-for-Inpainting"))
self.assertTrue(str(aot_pretrain).endswith("experiments/foo.pt"))
def test_parse_aot_rates(self):
rates = self.mod._parse_aot_rates("1+2+4+8")
self.assertEqual(rates, [1, 2, 4, 8])
def test_mask_bbox_empty(self):
mask = self.mod.np.zeros((3, 3), dtype=self.mod.np.uint8)
self.assertIsNone(self.mod._mask_bbox(mask))
def test_mask_bbox_and_expand(self):
mask = self.mod.np.zeros((5, 5), dtype=self.mod.np.uint8)
mask[1:3, 2:4] = 255
bbox = self.mod._mask_bbox(mask)
self.assertEqual(bbox, (2, 1, 3, 2))
expanded = self.mod._expand_bbox(bbox, 2, 5, 5)
self.assertEqual(expanded, (0, 0, 4, 4))
expanded_min = self.mod._expand_bbox_min_size(bbox, 0, 5, 5, 4)
self.assertEqual(expanded_min, (1, 0, 4, 3))
def test_remove_border_frame_components_keeps_inner_content(self):
mask = self.mod.np.zeros((12, 12), dtype=self.mod.np.uint8)
mask[0, :] = 255
mask[-1, :] = 255
mask[:, 0] = 255
mask[:, -1] = 255
mask[4:8, 5:7] = 255
cleaned = self.mod._remove_border_frame_components(
mask,
min_width_ratio=0.7,
min_height_ratio=0.7,
max_fill_ratio=0.35,
)
self.assertEqual(int(cleaned[0, 0]), 0)
self.assertGreater(cleaned[4:8, 5:7].mean(), 200)
def test_ensure_rgba_size_resizes(self):
img = self.mod.Image.new("RGB", (2, 3), color=(0, 0, 0))
out = self.mod._ensure_rgba_size(img, (4, 5))
self.assertEqual(out.mode, "RGBA")
self.assertEqual(out.size, (4, 5))
def test_extract_artwork_mask_detects_calligraphy_strokes(self):
image = self._make_paper_image()
self.mod.cv2.line(image, (20, 20), (130, 90), (20, 20, 20), thickness=6)
self.mod.cv2.line(image, (45, 15), (45, 105), (30, 30, 30), thickness=5)
mask = self.mod._extract_artwork_mask(
self.mod.Image.fromarray(image),
artwork_type="calligraphy",
max_size=120,
)
self.assertGreater(mask[25:95, 42:49].mean(), 180)
self.assertLess(mask[:15, :15].mean(), 20)
def test_extract_artwork_mask_detects_red_seal(self):
image = self._make_paper_image()
self.mod.cv2.rectangle(image, (40, 30), (120, 95), (165, 30, 40), thickness=5)
self.mod.cv2.line(image, (55, 40), (105, 85), (170, 25, 35), thickness=6)
mask = self.mod._extract_artwork_mask(
self.mod.Image.fromarray(image),
artwork_type="seal",
max_size=120,
)
self.assertGreater(mask[38:92, 38:122].mean(), 40)
self.assertLess(mask[:15, :15].mean(), 20)
def test_extract_artwork_mask_ignores_warm_paper_cast(self):
image = self.mod.np.full(
(120, 160, 3), (222, 188, 150), dtype=self.mod.np.uint8
)
self.mod.cv2.line(image, (30, 20), (30, 100), (28, 28, 28), thickness=6)
self.mod.cv2.line(image, (55, 20), (125, 95), (32, 32, 32), thickness=5)
mask = self.mod._extract_artwork_mask(
self.mod.Image.fromarray(image),
artwork_type="calligraphy",
max_size=120,
)
self.assertGreater(mask[25:100, 26:34].mean(), 180)
self.assertLess(mask[:12, :12].mean(), 20)
def test_mask_to_transparent_image_refines_alpha_inside_coarse_mask(self):
image = self.mod.np.full(
(120, 160, 3), (224, 190, 156), dtype=self.mod.np.uint8
)
self.mod.cv2.line(image, (40, 18), (40, 102), (24, 24, 24), thickness=6)
self.mod.cv2.rectangle(image, (102, 24), (132, 54), (170, 30, 40), thickness=-1)
coarse_mask = self.mod.np.zeros((120, 160), dtype=self.mod.np.uint8)
coarse_mask[10:110, 20:140] = 255
out = self.mod._mask_to_transparent_image(
self.mod.Image.fromarray(image),
coarse_mask,
)
alpha = self.mod.np.array(out.getchannel("A"))
self.assertGreater(alpha[25:95, 36:45].mean(), 110)
self.assertGreater(alpha[30:50, 108:128].mean(), 90)
self.assertLess(alpha[70:90, 70:90].mean(), 25)
def test_remove_background_artwork_mode_skips_rembg(self):
image = self._make_paper_image(width=100, height=80)
self.mod.cv2.line(image, (15, 15), (85, 60), (20, 20, 20), thickness=5)
original_remove = self.mod.remove
def _unexpected_remove(*args, **kwargs):
raise AssertionError("artwork 模式不应调用 rembg")
self.mod.remove = _unexpected_remove
try:
with tempfile.TemporaryDirectory() as tmpdir:
input_path = Path(tmpdir) / "input.png"
output_path = Path(tmpdir) / "output.png"
self.mod.Image.fromarray(image).save(input_path)
self.mod.remove_background(
str(input_path),
str(output_path),
foreground_mode="artwork",
artwork_type="calligraphy",
artwork_max_size=120,
)
out = self.mod.Image.open(output_path)
alpha = self.mod.np.array(out.getchannel("A"))
self.assertEqual(out.mode, "RGBA")
self.assertGreater(alpha[30:55, 35:70].mean(), 70)
self.assertLess(alpha[:10, :10].mean(), 10)
finally:
self.mod.remove = original_remove
if __name__ == "__main__":
unittest.main()