展示HN:Punge:基于YOLOv11n、CoreML和TensorFlow的设备端不适宜内容图像检测

1作者: markatlarge9 个月前原帖
我开发了一款名为 Punge 的移动应用,它可以扫描您的相册中的不适宜内容(NSFW)图像——完全在设备上进行处理。没有云端处理、无需登录,任何个人数据都不会离开您的手机。 技术细节: * 它由一个定制训练的 YOLOv11n 神经网络驱动。 * 在 iOS 上,所有推理通过 CoreML 进行;Android 则使用 TensorFlow Lite。 * 处理速度快——在 iPhone 15 上每张图像约需 10-20 毫秒。 * 在关键的不适宜内容类别中,准确率达到约 90%。 * 不会上传任何数据,应用也不收集用户信息。 这个项目最初是为了保护隐私,帮助人们在分享或交出手机之前清理手机中的内容。(您会惊讶于相册中被遗忘的内容。) 如果您对保护隐私的人工智能、本地推理感兴趣,或者只是想知道这个应用在实际使用中的效果,我非常欢迎您的反馈。 iOS: App Store Android: Play Store 演示视频: YouTube 更多信息: [https://markatlarge.com/](https://markatlarge.com/)
查看原文
I built a mobile app called Punge that scans your camera roll for NSFW images—completely on-device. No cloud processing, no login, and no personal data ever leaves your phone. Under the hood: * It&#x27;s powered by a custom-trained YOLOv11n neural network. * On iOS, all inference runs through CoreML; Android uses TensorFlow Lite. * Processing is fast—about 10–20ms per image on an iPhone 15. * Achieves ~90% confidence in key NSFW categories. * No data is uploaded, and the app doesn’t collect any user info. This started as a privacy-first project to help people clean up their phones before sharing or handing them over. (You’d be surprised what gets forgotten in a camera roll.) If you’re interested in privacy-preserving AI, local inference, or just curious how well this works in practice, I’d love your feedback. iOS: App Store Android: Play Store Demo video: YouTube More info: <a href="https:&#x2F;&#x2F;markatlarge.com&#x2F;" rel="nofollow">https:&#x2F;&#x2F;markatlarge.com&#x2F;</a>