关于Azure手写识别API无置信度返回值及结果准确性的问询

Hey there! Let's break down your questions about the Azure Computer Vision handwritten recognition API clearly:

1. Does the API support returning confidence values? #

Great question! By default, the standard response from the handwritten recognition (Read API) doesn't include confidence scores in the JSON output—this lines up with the response structure you shared:

"status": "Succeeded","succeeded": true,"failed": false,"finished": true,"recognitionResult": {"lines": [{"boundingBox": [ .....]}]

But there's a straightforward fix: if you use Read API version 3.1-preview.3 or later, add the includeConfidence=true parameter to your API request. Once enabled, each word object inside the lines array of the recognitionResult will include a confidence field with a value between 0 and 1 (1 representing the highest confidence level).

Here's a snippet of what the modified response would look like:

"recognitionResult": {
  "lines": [
    {
      "boundingBox": [10, 20, 60, 20, 60, 40, 10, 40],
      "text": "Hello World",
      "words": [
        {
          "boundingBox": [10, 20, 35, 20, 35, 40, 10, 40],
          "text": "Hello",
          "confidence": 0.97
        },
        {
          "boundingBox": [40, 20, 60, 20, 60, 40, 40, 40],
          "text": "World",
          "confidence": 0.95
        }
      ]
    }
  ]
}

2. Quick tips to boost recognition accuracy #

Since you mentioned some results are inaccurate, here are a few practical tweaks to try:

• Prioritize image quality: Make sure your handwritten images are well-lit, sharp, and free of blurriness or shadows that obscure text.
• Structure content clearly: Keep lines of handwriting spaced out, avoid overlapping characters, and write as neatly as possible.
• Specify the language: Use the language parameter in your request (e.g., language=en for English) to help the model focus on the correct character set.
• Split complex content: If you're scanning a long document or dense handwriting, split the image into smaller sections before sending it to the API—this helps the model process each segment more precisely.

内容的提问来源于stack exchange，提问作者user1411636