[{"data":1,"prerenderedAt":138},["ShallowReactive",2],{"blog-\u002Fblog\u002Fzh-hans\u002Fchallenges-of-smartphone-capture":3,"blog-\u002Fblog\u002Fzh-hans\u002Fchallenges-of-smartphone-capture-surround":112,"blog-\u002Fblog\u002Fzh-hans\u002Fchallenges-of-smartphone-capture-translations":121},{"id":4,"title":5,"author":6,"body":9,"date":96,"description":97,"draft":98,"extension":99,"image":100,"meta":101,"navigation":103,"path":104,"seo":105,"stem":106,"tags":107,"__hash__":111},"blogZhHans\u002Fblog\u002Fzh-hans\u002Fchallenges-of-smartphone-capture.md","智能手机肌肤分析的技术壁垒与应对之道",{"name":7,"picture":8},"Soo Ick Cho","\u002Fimages\u002Fblog\u002Fauthor-sooickcho.jpg",{"type":10,"value":11,"toc":88},"minimark",[12,16,19,24,27,40,43,47,50,53,56,60,63,66,69,73,76,79,82,85],[13,14,15],"p",{},"上一篇文章说，肌肤要经常测才能看出走势。但这里有一个前提：每次都得在相近的条件下拍摄。如果今天在窗边正对着拍，明天在房间里斜着拍，那么两张照片之间的差异，究竟是肌肤的变化，还是拍摄的差异，就无从分辨。",[13,17,18],{},"事实上，我们在对外会议中最常被问到的也正是这个问题：“用户拍照的方式五花八门，那结果还可信吗？”这是一个合理的问题，也是我们投入时间最多的部分。",[20,21,23],"h2",{"id":22},"角度距离光线都会改变结果","角度、距离、光线都会改变结果",[13,25,26],{},"如果任由大家自由拍摄，照片就会各不相同。",[28,29,30,34,37],"ul",{},[31,32,33],"li",{},"角度：头稍微抬一点或低一点，法令纹和松弛看起来就不一样",[31,35,36],{},"距离：拍得近，面部会变形；拍得远，细节又会糊掉",[31,38,39],{},"光线：在窗边的自然光下和在荧光灯下，色素和潮红拍出来完全不同",[13,41,42],{},"问题在于，这些差异可能比真实的肌肤变化还要大。这样一来，上一篇文章里说的“点连成的线”，反映的就不是肌肤的走势，而是拍摄条件的走势。",[20,44,46],{"id":45},"而且智能手机的机型各不相同","而且，智能手机的机型各不相同",[13,48,49],{},"在此之上还要再叠加一层：每个人用的智能手机都不一样。",[13,51,52],{},"不同机型的摄像头像素和镜头特性不同，色彩处理方式也因厂商而异。同一张脸在同一个位置拍，不同设备拍出来的结果也会略有差别。这与诊所设备那种只需控制一台机器的情况，前提完全不同。",[13,54,55],{},"这是选择智能手机所要付出的代价。我们得到了“人人都能用自己手里的设备来测”这一优点，作为交换，就必须承担这种多样性。",[20,57,59],{"id":58},"所以我们在寻找理想的那一点","所以，我们在寻找“理想的那一点”",[13,61,62],{},"最简单的解法是严格控制拍摄条件：固定的距离、固定的角度、固定的光线。但这样做，准确度虽然上去了，却不会再有人每天拍。能够经常测量本是智能手机最大的优点，这等于亲手把这个优点抹掉了。",[13,64,65],{},"于是我们反其道而行：观察真实用户是怎么拍的，并基于这些数据，寻找两者交汇的那一点——把拍摄的麻烦降到最低，同时又能保持评估结果的准确度。",[13,67,68],{},"这不是一个有唯一确定答案的问题，更像是一项边看数据边持续调整的工作。在这个过程中找到的方法，我们已经申请了专利。",[20,70,72],{"id":71},"先测量偏差才能减少偏差","先测量偏差，才能减少偏差",[13,74,75],{},"因设备和环境而产生的差异，我们通过模型训练来应对。不过，并不是一味地堆数据。",[13,77,78],{},"我们把模型应用于在各种条件下拍摄的照片，然后分析结果会随条件散开到什么程度：同样的肌肤，换了机型数值会动多少；光线变了，哪个维度特别敏感——分开来看。在六个维度中，色素和潮红对光线更敏感，皱纹和松弛对角度更敏感，所以每个维度都需要不同的应对方式。",[13,80,81],{},"这样确认了偏差的大小和原因之后，我们再补充薄弱条件下的数据，重新训练模型。然后用同样的方式再测一次。因为偏差如果看不见，也就无从减少。",[13,83,84],{},"做一个只针对一种机型、一种光线条件的模型，相对容易，因为条件是固定的。难的是让人无论用哪部手机、在哪里拍，都能得到相近的数值。而如果是一项让用户各自用自己的手机来测的服务，就必须把难的这一边做成。",[13,86,87],{},"在前面的文章里，我写过用智能手机前置摄像头分析肌肤在技术上并不容易。这篇文章里写的这些，就是那份困难的真实面貌。尽管如此，我们仍然选择了这个方向，理由只有一个：我认为肌肤数据的主人应该是本人，而不是诊所的服务器。",{"title":89,"searchDepth":90,"depth":90,"links":91},"",2,[92,93,94,95],{"id":22,"depth":90,"text":23},{"id":45,"depth":90,"text":46},{"id":58,"depth":90,"text":59},{"id":71,"depth":90,"text":72},"2026-08-09","角度、距离和光线只要稍有不同，肌肤分析结果就会波动。这是一篇关于我们如何寻找那个既减少拍摄麻烦、又守住准确度的平衡点的记录。",false,"md","\u002Fimages\u002Fblog\u002FChatGPT Image 2026년 8월 9일 오후 09_32_46.png",{"slug":102},"challenges-of-smartphone-capture",true,"\u002Fblog\u002Fzh-hans\u002Fchallenges-of-smartphone-capture",{"title":5,"description":97},"blog\u002Fzh-hans\u002Fchallenges-of-smartphone-capture",[108,109,110],"智能手机拍摄","肌肤分析","INSKINVIEW","xbTlmX3a6dPOwRNWss8pgjQw9dbD4Kl39dMMh__4MLY",[113,117],{"title":114,"path":115,"stem":116,"children":-1},"精细的肌肤评估 AI 是怎样做出来的","\u002Fblog\u002Fzh-hans\u002Ffrom-grades-to-scores","blog\u002Fzh-hans\u002Ffrom-grades-to-scores",{"title":118,"path":119,"stem":120,"children":-1},"一次测量看不到的东西","\u002Fblog\u002Fzh-hans\u002Fwhy-measure-often","blog\u002Fzh-hans\u002Fwhy-measure-often",[122,126,130,134],{"code":123,"name":124,"path":125},"ko","한국어","\u002Fblog\u002Fchallenges-of-smartphone-capture",{"code":127,"name":128,"path":129},"en","English","\u002Fblog\u002Fen\u002Fchallenges-of-smartphone-capture",{"code":131,"name":132,"path":133},"ja","日本語","\u002Fblog\u002Fja\u002Fchallenges-of-smartphone-capture",{"code":135,"name":136,"path":137},"zh-hant","繁體中文","\u002Fblog\u002Fzh-hant\u002Fchallenges-of-smartphone-capture",1791649182965]