xuanwulab/HaS_Image_0209_FP32

HaS Image Model (FP32)

HaS (Hide and Seek) is an on-device image privacy model that performs pixel-level detection and segmentation of sensitive regions in images.

• 📦 YOLO11-based instance segmentation, FP32, 119 MB
• 🔒 Data never leaves device — local inference, no network required
• 🎯 21 privacy categories covering biometrics, ID documents, financial cards, screens, and more
• ⚡ ~20ms per image inference speed on modern hardware

1. Core Capabilities

Traditional image redaction relies on face-only detectors or manual masking. HaS Image is an on-device privacy segmentation model — it recognizes 21 categories of sensitive content with pixel-level masks, enabling precise and automated redaction.

Capability	Description
21 Privacy Categories	Covers biometrics (face, fingerprint, palmprint), ID documents (ID card, passport, permit), financial (bank card), screens (mobile, monitor), and more
Pixel-Level Segmentation	Instance segmentation, not just bounding boxes — masks follow the exact contour of each sensitive region
Multiple Redaction Methods	Supports mosaic, Gaussian blur, and solid fill — configurable strength and style per use case
Batch Processing	Process entire directories of images in one command
Category Filtering	Selectively redact only specific categories (e.g., faces only, or faces + license plates)
High Speed	~20ms per image, suitable for real-time and batch workflows

2. Supported Privacy Categories

6 groups, 21 categories:

Group	ID	Category	Description
Biometrics	0	face	Human faces
	1	fingerprint	Fingerprints
	2	palmprint	Palm prints
ID Documents	3	id_card	National ID cards
	4	hk_macau_permit	HK/Macau travel permits
	5	passport	Passports
	6	employee_badge	Employee badges
Transportation	7	license_plate	Vehicle license plates
Financial	8	bank_card	Bank/credit cards
Security	9	physical_key	Physical keys
Documents	10	receipt	Receipts
	11	shipping_label	Shipping/courier labels
	12	official_seal	Official stamps/seals
	20	paper	Paper documents
Information Carriers	13	whiteboard	Whiteboards
	14	sticky_note	Sticky notes
	15	mobile_screen	Mobile phone screens
	16	monitor_screen	Computer monitor screens
Medical	17	medical_wristband	Hospital wristbands
Codes	18	qr_code	QR codes
	19	barcode	Barcodes

Category filtering supports English names, Chinese names, or numeric IDs (comma-separated).

3. Quick Start

Installation

pip install ultralytics

Python Usage

from ultralytics import YOLO

model = YOLO("sensitive_seg_best.pt")

# Detect all privacy regions
results = model.predict("photo.jpg", conf=0.25)

# Process results
for result in results:
    for box, mask in zip(result.boxes, result.masks):
        category_id = int(box.cls)
        confidence = float(box.conf)
        print(f"Category: {category_id}, Confidence: {confidence:.2f}")

CLI Usage (via HaS)

HaS ships with a CLI tool has-image that wraps the model into ready-to-use commands:

# Scan — detect privacy regions without modifying the image
has-image scan --image photo.jpg

# Hide — detect and redact with mosaic (default)
has-image hide --image photo.jpg

# Hide — redact faces only, with Gaussian blur
has-image hide --image photo.jpg --types face --method blur --strength 25

# Hide — batch process a directory
has-image hide --dir ./photos/ --output-dir ./redacted/

4. Redaction Methods

Method	Flag	Description
Mosaic	--method mosaic	Pixelated blocks (default). --strength controls block size
Gaussian Blur	--method blur	Smooth blur. --strength controls blur radius
Solid Fill	--method fill	Solid color overlay. --fill-color sets hex color (default #000000)

5. Usage Scenarios

Scenario	Description	Method
Photo Publishing	Redact faces and license plates before sharing photos online	hide --types face,license_plate
Document Redaction	Mask ID cards, bank cards, and seals in scanned documents	hide --types id_card,bank_card,official_seal
Dataset Anonymization	Batch-anonymize images in ML training datasets	hide --dir ./dataset/
Compliance Screening	Scan uploaded images for sensitive content before processing	scan --image uploaded.jpg
Video Frame Processing	Extract frames, redact per-frame, reassemble	hide per frame
Screen Recording Privacy	Mask phone/monitor screens in recordings	hide --types mobile_screen,monitor_screen

6. Redaction Example

Figure 1. Before and after privacy redaction with HaS Image model

7. Model Specifications

Property	Value
Architecture	YOLO11 Instance Segmentation
Precision	FP32
File Size	119 MB
Task	Instance Segmentation
Categories	21 privacy classes
Inference Speed	~20ms per image
Input	Any resolution (auto-scaled)
Output	Bounding boxes + pixel-level masks + confidence scores
Framework	Ultralytics
License	MIT

8. Related Models

Model	Type	Description
HaS Text Q8_0	Text	Text anonymization, 0.6B params, Q8_0 quantized, 639 MB
HaS Text Q4_K_M	Text	Text anonymization, 0.6B params, Q4_K_M quantized, 397 MB
HaS Image FP32	Image	Image privacy segmentation, YOLO11, FP32, 119 MB

HaS Text handles text anonymization (named entities, PII); HaS Image handles image anonymization (visual privacy regions). Together they provide a complete on-device privacy solution for both text and images.

---

中文版

HaS Image Model (FP32)

HaS（Hide and Seek） 是一个端侧部署的图像隐私模型，对图片中的隐私区域进行像素级识别和分割。

• 📦 基于 YOLO11 实例分割，FP32 精度，119 MB
• 🔒 数据不出设备，本地推理，无需联网
• 🎯 21 类隐私类别，覆盖生物特征、证件、金融卡、屏幕等
• ⚡ 单张图片 ~20ms 推理速度

一、核心能力

传统图像脱敏依赖人脸检测器或手动遮挡。HaS Image 是一个端侧隐私分割模型——识别 21 类敏感内容，提供像素级掩码，实现精准自动化脱敏。

能力	说明
21 类隐私类别	覆盖生物特征（人脸、指纹、掌纹）、证件（身份证、护照、通行证）、金融（银行卡）、屏幕（手机、显示器）等
像素级分割	实例分割而非仅边界框——掩码精确贴合每个敏感区域的轮廓
多种遮挡方式	支持马赛克、高斯模糊、纯色填充——可按场景配置强度和样式
批量处理	一条命令处理整个目录的图片
类别过滤	选择性脱敏指定类别（如仅人脸，或人脸+车牌）
高速推理	单张图片 ~20ms，适合实时和批量工作流

二、支持的隐私类别

6 大分组，21 个类别：

分组	ID	类别	中文
生物特征	0	face	人脸
	1	fingerprint	指纹
	2	palmprint	掌纹
证件	3	id_card	身份证
	4	hk_macau_permit	港澳通行证
	5	passport	护照
	6	employee_badge	工牌
交通	7	license_plate	车牌
金融	8	bank_card	银行卡
安全	9	physical_key	钥匙
文档	10	receipt	收据
	11	shipping_label	快递单
	12	official_seal	公章
	20	paper	纸张
信息载体	13	whiteboard	白板
	14	sticky_note	便签
	15	mobile_screen	手机屏幕
	16	monitor_screen	显示器屏幕
医疗	17	medical_wristband	医用腕带
编码	18	qr_code	二维码
	19	barcode	条形码

类别过滤支持英文名、中文名或数字 ID，逗号分隔。

三、快速开始

安装

pip install ultralytics

Python 使用

from ultralytics import YOLO

model = YOLO("sensitive_seg_best.pt")

# 检测所有隐私区域
results = model.predict("photo.jpg", conf=0.25)

# 处理结果
for result in results:
    for box, mask in zip(result.boxes, result.masks):
        category_id = int(box.cls)
        confidence = float(box.conf)
        print(f"类别: {category_id}, 置信度: {confidence:.2f}")

CLI 使用（通过 HaS）

HaS 配套了 CLI 工具 has-image，将模型封装为开箱即用的命令：

# 扫描——仅检测隐私区域，不修改图片
has-image scan --image photo.jpg

# 脱敏——检测并遮挡（默认马赛克）
has-image hide --image photo.jpg

# 脱敏——仅遮挡人脸，使用高斯模糊
has-image hide --image photo.jpg --types face --method blur --strength 25

# 批量处理整个目录
has-image hide --dir ./photos/ --output-dir ./redacted/

四、遮挡方式

方式	参数	说明
马赛克	--method mosaic	像素化方块（默认）。--strength 控制块大小
高斯模糊	--method blur	平滑模糊。--strength 控制模糊半径
纯色填充	--method fill	纯色覆盖。--fill-color 设置十六进制颜色（默认 #000000）

五、使用场景

场景	说明	方法
照片发布	分享照片前遮挡人脸和车牌	hide --types face,license_plate
文档脱敏	遮挡扫描件中的身份证、银行卡、公章	hide --types id_card,bank_card,official_seal
数据集匿名化	批量匿名化 ML 训练数据集中的图片	hide --dir ./dataset/
合规审查	上传图片处理前扫描敏感内容	scan --image uploaded.jpg
视频帧处理	抽帧 → 逐帧脱敏 → 重新组装	逐帧 hide
录屏隐私	遮挡录屏中的手机/显示器屏幕	hide --types mobile_screen,monitor_screen

六、脱敏示例

图 1. HaS Image 模型脱敏前后对比

七、模型规格

属性	值
架构	YOLO11 实例分割
精度	FP32
文件大小	119 MB
任务	实例分割
类别数	21 类隐私实体
推理速度	单张图片 ~20ms
输入	任意分辨率（自动缩放）
输出	边界框 + 像素级掩码 + 置信度分数
框架	Ultralytics
许可证	MIT

八、相关模型

模型	类型	说明
HaS Text Q8_0	文本	文本脱敏，0.6B 参数，Q8_0 量化，639 MB
HaS Text Q4_K_M	文本	文本脱敏，0.6B 参数，Q4_K_M 量化，397 MB
HaS Image FP32	图像	图像隐私分割，YOLO11，FP32，119 MB

HaS Text 处理文本脱敏（命名实体、个人信息）；HaS Image 处理图像脱敏（视觉隐私区域）。二者结合，提供完整的端侧文本+图像隐私保护方案。