如何提高表单识别器的准确性？
How to improve the accuracy of Form Recognizer?
我在 5 张图像上训练了我的模型，但准确度不是特别高。
Link 形成：https://imgur.com/a/BOHVG7G
JSON 输出：
{
"status": "success",
"pages": [
{
"number": 1,
"height": 1055,
"width": 1225,
"clusterId": 0,
"keyValuePairs": [
{
"key": [
{
"text": "Name:",
"boundingBox": [
163.7,
987.1,
242.2,
987.1,
242.2,
963.4,
163.7,
963.4
]
}
],
"value": [
{
"text": "Luca Bassi",
"boundingBox": [
365.3,
982.1,
458.8,
982.1,
458.8,
963.4,
365.3,
963.4
],
"confidence": 1
},
{
"text": "Brassi",
"boundingBox": [
365.3,
938.7,
417.1,
938.7,
417.1,
919,
365.3,
919
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Surname:",
"boundingBox": [
166.9,
937.8,
282.8,
937.8,
282.8,
913.1,
166.9,
913.1
]
}
],
"value": [
{
"text": "19 Cider Lane",
"boundingBox": [
367.3,
719.1,
490.4,
719.1,
490.4,
698.4,
367.3,
698.4
],
"confidence": 0.8
}
]
},
{
"key": [
{
"text": "e-Mail Address:",
"boundingBox": [
164.7,
893.4,
358.1,
893.4,
358.1,
867.8,
164.7,
867.8
]
}
],
"value": [
{
"text": "brassi@brassi.com",
"boundingBox": [
364.3,
893.4,
528,
893.4,
528,
867.8,
364.3,
867.8
],
"confidence": 0.6
}
]
},
{
"key": [
{
"text": "Phone Number:",
"boundingBox": [
163.7,
849.1,
361.1,
849.1,
361.1,
822.6,
163.7,
822.6
]
}
],
"value": [
{
"text": "456-3456",
"boundingBox": [
367.3,
849.1,
451.8,
849.1,
451.8,
822.6,
367.3,
822.6
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Mobile Number:",
"boundingBox": [
164.7,
803.8,
361.1,
803.8,
361.1,
777.3,
164.7,
777.3
]
}
],
"value": [
{
"text": "456-2135",
"boundingBox": [
366.3,
803.8,
450.8,
803.8,
450.8,
777.3,
366.3,
777.3
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Street:",
"boundingBox": [
166.9,
714.1,
246.2,
714.1,
246.2,
690.5,
166.9,
690.5
]
}
],
"value": []
},
{
"key": [
{
"text": "House:",
"boundingBox": [
163.7,
668.8,
250.2,
668.8,
250.2,
645.3,
163.7,
645.3
]
}
],
"value": [
{
"text": "Detroit",
"boundingBox": [
364.3,
628.5,
427.3,
628.5,
427.3,
609.7,
364.3,
609.7
],
"confidence": 0.6
}
]
},
{
"key": [
{
"text": "Town:",
"boundingBox": [
166.9,
623.5,
241.2,
623.5,
241.2,
598.9,
166.9,
598.9
]
}
],
"value": [
{
"text": "80012",
"boundingBox": [
365.3,
585.2,
418.1,
585.2,
418.1,
565.5,
365.3,
565.5
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Postcode:",
"boundingBox": [
164.7,
580.2,
286.8,
580.2,
286.8,
554.5,
164.7,
554.5
]
}
],
"value": [
{
"text": "Russia",
"boundingBox": [
365.3,
534.8,
417.1,
534.8,
417.1,
516.2,
365.3,
516.2
],
"confidence": 0.6
}
]
},
{
"key": [
{
"text": "Comments:",
"boundingBox": [
166.9,
487.7,
305.2,
487.7,
305.2,
464,
166.9,
464
]
}
],
"value": [
{
"text": "The quick brown fox",
"boundingBox": [
366.3,
485.7,
549.4,
485.7,
549.4,
464,
366.3,
464
],
"confidence": 0.6
}
]
}
],
"tables": []
}
],
"errors": []
}
如您所见，姓氏和地址的内容并不是那么好。有没有办法更有效地训练它，还是我只需要使用更大的数据集？
我在 azure 的门户网站上四处搜寻，但我不太确定我是否忽略了一个更好地训练它的选项。
我在 Python 中用 opencv 和 matplotlib 创建了一个程序来检查你的结果，然后我发现 Surname 和 e-Mail Address 都出来了，但是House和Country不是，如下图。
这是我的绘图代码。
import cv2
import matplotlib.pyplot as plt
import json
import numpy as np
json_file = open('sample.json')
json_dict = json.load(json_file)
page = json_dict['pages'][0]
height = page['height']
keyValuePairs = page['keyValuePairs']
key_boundingBoxes = [np.int64(key['boundingBox']) for keyValuePair in keyValuePairs for key in keyValuePair['key']]
key_texts = [key['text'] for keyValuePair in keyValuePairs for key in keyValuePair['key']]
value_texts = [value['text'] for keyValuePair in keyValuePairs for value in keyValuePair['value']]
print(key_texts)
value_boundingBoxes = [np.int64(value['boundingBox']) for keyValuePair in keyValuePairs for value in keyValuePair['value']]
img = cv2.imread("sample.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
[cv2.rectangle(img, (boundingBox[0], height - boundingBox[1]),(boundingBox[4], height - boundingBox[5]) ,(0,255,0), 3) for boundingBox in key_boundingBoxes]
[cv2.rectangle(img, (boundingBox[0], height - boundingBox[1]),(boundingBox[4], height - boundingBox[5]) ,(255,0,0), 3) for boundingBox in value_boundingBoxes]
plt.figure()
plt.imshow(img)
plt.axis('off')
plt.show()
当然，跟提高准确率没关系
根据我的经验，为训练模型提供更多图像很容易提高准确性，因为您使用的是无法更改算法的 Azure 认知服务。
我在 5 张图像上训练了我的模型，但准确度不是特别高。
Link 形成：https://imgur.com/a/BOHVG7G
JSON 输出：
{
"status": "success",
"pages": [
{
"number": 1,
"height": 1055,
"width": 1225,
"clusterId": 0,
"keyValuePairs": [
{
"key": [
{
"text": "Name:",
"boundingBox": [
163.7,
987.1,
242.2,
987.1,
242.2,
963.4,
163.7,
963.4
]
}
],
"value": [
{
"text": "Luca Bassi",
"boundingBox": [
365.3,
982.1,
458.8,
982.1,
458.8,
963.4,
365.3,
963.4
],
"confidence": 1
},
{
"text": "Brassi",
"boundingBox": [
365.3,
938.7,
417.1,
938.7,
417.1,
919,
365.3,
919
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Surname:",
"boundingBox": [
166.9,
937.8,
282.8,
937.8,
282.8,
913.1,
166.9,
913.1
]
}
],
"value": [
{
"text": "19 Cider Lane",
"boundingBox": [
367.3,
719.1,
490.4,
719.1,
490.4,
698.4,
367.3,
698.4
],
"confidence": 0.8
}
]
},
{
"key": [
{
"text": "e-Mail Address:",
"boundingBox": [
164.7,
893.4,
358.1,
893.4,
358.1,
867.8,
164.7,
867.8
]
}
],
"value": [
{
"text": "brassi@brassi.com",
"boundingBox": [
364.3,
893.4,
528,
893.4,
528,
867.8,
364.3,
867.8
],
"confidence": 0.6
}
]
},
{
"key": [
{
"text": "Phone Number:",
"boundingBox": [
163.7,
849.1,
361.1,
849.1,
361.1,
822.6,
163.7,
822.6
]
}
],
"value": [
{
"text": "456-3456",
"boundingBox": [
367.3,
849.1,
451.8,
849.1,
451.8,
822.6,
367.3,
822.6
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Mobile Number:",
"boundingBox": [
164.7,
803.8,
361.1,
803.8,
361.1,
777.3,
164.7,
777.3
]
}
],
"value": [
{
"text": "456-2135",
"boundingBox": [
366.3,
803.8,
450.8,
803.8,
450.8,
777.3,
366.3,
777.3
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Street:",
"boundingBox": [
166.9,
714.1,
246.2,
714.1,
246.2,
690.5,
166.9,
690.5
]
}
],
"value": []
},
{
"key": [
{
"text": "House:",
"boundingBox": [
163.7,
668.8,
250.2,
668.8,
250.2,
645.3,
163.7,
645.3
]
}
],
"value": [
{
"text": "Detroit",
"boundingBox": [
364.3,
628.5,
427.3,
628.5,
427.3,
609.7,
364.3,
609.7
],
"confidence": 0.6
}
]
},
{
"key": [
{
"text": "Town:",
"boundingBox": [
166.9,
623.5,
241.2,
623.5,
241.2,
598.9,
166.9,
598.9
]
}
],
"value": [
{
"text": "80012",
"boundingBox": [
365.3,
585.2,
418.1,
585.2,
418.1,
565.5,
365.3,
565.5
],
"confidence": 1
}
]
},
{
"key": [
{
"text": "Postcode:",
"boundingBox": [
164.7,
580.2,
286.8,
580.2,
286.8,
554.5,
164.7,
554.5
]
}
],
"value": [
{
"text": "Russia",
"boundingBox": [
365.3,
534.8,
417.1,
534.8,
417.1,
516.2,
365.3,
516.2
],
"confidence": 0.6
}
]
},
{
"key": [
{
"text": "Comments:",
"boundingBox": [
166.9,
487.7,
305.2,
487.7,
305.2,
464,
166.9,
464
]
}
],
"value": [
{
"text": "The quick brown fox",
"boundingBox": [
366.3,
485.7,
549.4,
485.7,
549.4,
464,
366.3,
464
],
"confidence": 0.6
}
]
}
],
"tables": []
}
],
"errors": []
}
如您所见，姓氏和地址的内容并不是那么好。有没有办法更有效地训练它，还是我只需要使用更大的数据集？
我在 azure 的门户网站上四处搜寻，但我不太确定我是否忽略了一个更好地训练它的选项。
我在 Python 中用 opencv 和 matplotlib 创建了一个程序来检查你的结果，然后我发现 Surname 和 e-Mail Address 都出来了，但是House和Country不是，如下图。
这是我的绘图代码。
import cv2
import matplotlib.pyplot as plt
import json
import numpy as np
json_file = open('sample.json')
json_dict = json.load(json_file)
page = json_dict['pages'][0]
height = page['height']
keyValuePairs = page['keyValuePairs']
key_boundingBoxes = [np.int64(key['boundingBox']) for keyValuePair in keyValuePairs for key in keyValuePair['key']]
key_texts = [key['text'] for keyValuePair in keyValuePairs for key in keyValuePair['key']]
value_texts = [value['text'] for keyValuePair in keyValuePairs for value in keyValuePair['value']]
print(key_texts)
value_boundingBoxes = [np.int64(value['boundingBox']) for keyValuePair in keyValuePairs for value in keyValuePair['value']]
img = cv2.imread("sample.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
[cv2.rectangle(img, (boundingBox[0], height - boundingBox[1]),(boundingBox[4], height - boundingBox[5]) ,(0,255,0), 3) for boundingBox in key_boundingBoxes]
[cv2.rectangle(img, (boundingBox[0], height - boundingBox[1]),(boundingBox[4], height - boundingBox[5]) ,(255,0,0), 3) for boundingBox in value_boundingBoxes]
plt.figure()
plt.imshow(img)
plt.axis('off')
plt.show()
当然，跟提高准确率没关系
根据我的经验，为训练模型提供更多图像很容易提高准确性，因为您使用的是无法更改算法的 Azure 认知服务。