【发布时间】：2019-08-26 16:34:41
【问题描述】：
我正在使用表单识别器从示例 PDF 表单 (ACORD 3101 (2012/02)) 中提取文本信息，但在某些字段上返回的“boundingBox”对我来说似乎不正确。我想知道它背后的原因是什么。
- 请说明如何计算这些boundingBox 信息。谢谢。
- 请查看返回的 json 响应。 “键值”对也不如预期。有一个“令牌”键带有多个值。但实际上，这些值中的每一个都应该在它们自己的键下。为什么？
我使用多张填写的表格和一张空白的表格对服务进行了培训。当我在训练模型上调用“/{id}/keys”时，我确实看到了被识别的键：
{
"clusters": {
"0": ["ADDITIONAL REMARKS", "ADDITIONAL REMARKS SCHEDULE", "Effective Date:", "Form Number:", "Form Title:", "Insured", "Insurer", "Intermediary", "Page", "Policy Number", "This Additional Remarks form is a schedule to ACORD form,", "__Tokens__"]
}
}
我觉得很好。然后我调用“/{id}/analyze”API 来提取样本 PDF。正如我所说，结果似乎不正确。以下是我得到的 Json 响应的一部分。
{
"status": "success",
"pages": [
{
"number": 1,
"height": 842,
"width": 595,
"clusterId": 0,
"keyValuePairs": [
{
"key": [
{
"text": "Page",
"boundingBox": [
493.2,
811.6,
514.7,
811.6,
514.7,
801.6,
493.2,
801.6
]
}
],
"value": [
{
"text": "of",
"boundingBox": [
543.6,
811.6,
552.1,
811.6,
552.1,
801.6,
543.6,
801.6
],
"confidence": 1.0
}
]
},
{
"key": [
{
"text": "__Tokens__",
"boundingBox": [
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
]
}
],
"value": [
{
"text": "1",
"boundingBox": [
62.3,
97.3,
62.8,
97.3,
62.8,
96.2,
62.3,
96.2
],
"confidence": 0.24
},
{
"text": "1",
"boundingBox": [
66.6,
97.3,
67.1,
97.3,
67.1,
96.2,
66.6,
96.2
],
"confidence": 0.24
},
{
"text": "John Doe",
"boundingBox": [
2.8,
93.9,
6.9,
93.9,
6.9,
92.8,
2.8,
92.8
],
"confidence": 0.24
},
{
"text": "Taren Liu",
"boundingBox": [
36.4,
93.8,
40.4,
93.8,
40.4,
92.8,
36.4,
92.8
],
"confidence": 0.24
},
{
"text": "23456R02",
"boundingBox": [
2.8,
90.8,
7.2,
90.8,
7.2,
89.8,
2.8,
89.8
],
"confidence": 0.24
},
{
"text": "RBA",
"boundingBox": [
2.8,
87.9,
4.7,
87.9,
4.7,
86.9,
2.8,
86.9
],
"confidence": 0.24
},
{
"text": "11/08/2019",
"boundingBox": [
48.2,
87.9,
53.0,
87.9,
53.0,
86.9,
48.2,
86.9
],
"confidence": 0.24
},
{
"text": "140001",
"boundingBox": [
10.4,
83.3,
13.6,
83.3,
13.6,
82.2,
10.4,
82.2
],
"confidence": 0.24
},
{
"text": "Hello World",
"boundingBox": [
22.6,
83.3,
27.5,
83.3,
27.5,
82.2,
22.6,
82.2
],
"confidence": 0.24
},
{
"text": "This is the second fake form. See",
"boundingBox": [
2.8,
80.9,
17.0,
80.9,
17.0,
79.8,
2.8,
79.8
],
"confidence": 0.24
},
{
"text": "if",
"boundingBox": [
17.3,
80.9,
17.8,
80.9,
17.8,
79.8,
17.3,
79.8
],
"confidence": 0.24
},
{
"text": "the form recognizer can learn from this.",
"boundingBox": [
18.0,
80.9,
34.7,
80.9,
34.7,
79.8,
18.0,
79.8
],
"confidence": 0.24
}
]
}
],
"tables": []
}
],
"errors": []
}
请注意高度和宽度值（分别为 842 和 595）是正确的。这些是以点为单位的正常 A4 纸尺寸。但是，“John Doe”和“aren Liu”的字段具有不正确的 boundingBox 信息。显然，这些边界框聚集在论文的左下角（例如，对于“John Doe”，它们是 2.8、93.9、6.9、93.9、6.9、92.8、2.8、92.8），而不是 pdf 顶部的预期位置.为什么？
Here is the sample pdf used for both training and analyzing
【问题讨论】：
标签： microsoft-cognitive form-recognizer