跳到正文
DEV Community· Bonggi Cho·· 2 天前AI 評分41

在韓國旅行:用手機鏡頭閲讀菜單、標誌和告示

Traveling in Korea: reading menus, signs and notices with a phone camera

AI 導讀

一名在首爾獨立開發 Android 應用程式的開發者製作了 Photo Read,可辨識相片中的文字,並翻譯成韓文、英文、中文或日文。辨識和翻譯均在手機上處理;下載語言檔後可離線使用,應用程式免費並以廣告支援,無需帳戶。

正文

Korea is easy to travel in until the moment everything is written in Korean. A menu on a wall, a notice taped to a shop door, a form on a counter, the label on a medicine packet you were handed at a pharmacy: you can point at things, but you cannot read them.

I am a solo Android developer in Seoul, and I built a small free app for that. But most of this post is about the problem and the trades involved in reading text with a phone camera, because that part applies whatever app you use.

The three places Korean-only text catches a visitor

  1. Food. Menus often have pictures, but the prices, the portion size, the "spicy" warning and whether a dish contains pork or beef are in text.
  2. Signs and notices. "Closed today", "cash only", "elevator out of order", opening hours printed in a font size nobody over forty can read at arm's length.
  3. Paper handed to you. A receipt, a delivery slip, a prescription envelope, a form at a clinic. This is the case where translation is not a convenience but the difference between understanding and guessing.

What to look for in a camera reader

Four things decide whether a camera OCR app is useful on a trip.

  • Does it work without a connection? Roaming data is expensive, and subway platforms are exactly where you need to read something.
  • Where does the photo go? If recognition happens on a server, the photo of your prescription is now on someone else's computer.
  • How many taps from pocket to answer? A reader that needs an account, a login and a language picker before it shows anything is not going to be used at a restaurant table.
  • Does it speak? A spoken translation is easier to follow while you are standing in front of a person who is waiting for your answer.

Why on-device processing changes the experience

The app I built, Photo Read, runs recognition and translation on the phone. The photo is not uploaded, and after each language file has been downloaded once, translation keeps working offline. That decision has a cost: the models ship with or get downloaded to the app, so the first use of a language needs an internet connection, and on-device models are not the largest ones available.

What you gain is speed, privacy and the ability to work in a basement restaurant with no signal. On-device inference is also the reason the app can be free with ads and still keep a session short: no round trip to a server, nothing to bill per request.

What the app does, concretely

  • Take a photo of text: a menu, a sign, a document.
  • The app recognises the text and translates it into Korean, English, Chinese or Japanese.
  • One tap reads the translation out loud, after a short ad.
  • One screen: four language buttons and one big capture button, no settings to get lost in.

Google Play: https://play.google.com/store/apps/details?id=io.cho.photo_read

It is free, ad-supported, has no account and no in-app purchase. The same person also publishes BboshaCam, a free camera app, but that is a different story.

Practical tips for reading Korean in the wild

  • Fill the frame with the text block, not the whole menu board. Recognition works on what is in focus.
  • Keep the phone parallel to the paper. Perspective is harder than blur.
  • Photograph one paragraph or one column at a time. A page of dense Korean is a harder problem than three lines of it.
  • Take the photo before you ask the question. Standing at a counter with a screen is easier for everyone than pointing at the wall twice.
  • Treat the output as a strong hint, not a legal translation. Numbers and short phrases are reliable; a long complicated sentence is where machine translation stops being a translator and becomes a paraphrase.

A note for other developers

If you are building something similar, the constraint that improved the app most was deleting options. The first version had a settings screen and a language picker; every person I handed the phone to opened it, changed a language and got lost. Four fixed buttons and one capture button did more for real usage than any accuracy improvement.

The second was treating offline as a feature rather than a fallback. The moment the app works with no signal, it stops being a tool for a planned task and becomes something you can reach for in the middle of a day.

If you travel here and try it, I would like to know what it got wrong first. That is the most useful feedback I can get.

來源:DEV Community · dev.to