Oracle Cloud Infrastructure Documentation

Pretrained Models

There are five types of pretrained model with Vision.

The pretrained models are:

Optical Character Recognition (OCR)

Vision can detect and recognize text in a document. Language classification identifies the language of a document, then OCR draws bounding boxes around the printed or hand-written text it finds in an image, and digitizes the text.

If you have a PDF with text, Vision finds the text in that document and extracts the text. It then provides bounding boxes for the identified text. Text Detection can be used with Document AI or Image Analysis models.

Vision provides a confidence score for each text grouping. The confidence score is a decimal number. Scores closer to 1 indicate a higher confidence in the extracted text, while lower scores indicate lower confidence score. The range of the confidence score for each label is from 0 to 1.

Note

OCR support is limited to English. If you know that the text in the images is in English, set the language to Eng.
Supported features are:
  • Word extraction
  • Text line extraction
  • Confidence score
  • Bounding polygons
  • Single request
  • Batch request
Limitations are:
  • Although Language classification identifies several languages, OCR is limited to English.
OCR Example

An example of OCR use in Vision.

Input document
Figure 1. OCR Input
Receipt from a fictitious cafe, including two line items, tax, subtotal and total amounts.
API Request: 
{ "analyzeDocumentDetails":
 { "compartmentId": "",
   "document": { "namespaceName": "",
   "bucketName": "",
   "objectName": "",
   "source": "OBJECT_STORAGE" },
  "features":
             [ { "featureType": "TEXT_DETECTION" },
               { "featureType": "LANGUAGE_CLASSIFICATION",
                 "maxResults": 5 } ]
 } 
}
.
Output:
Figure 2. OCR Output
The receipt with all the fields identifed
API Response: 
{ "documentMetadata":
 { "pageCount": 1,
   "mimeType": "image/jpeg" },
   "pages":
           [ { "pageNumber": 1,
               "dimensions":
                            { "width": 361, 
                              "height": 600,
                              "unit": "PIXEL" },
                              "detectedLanguages":
                                                  [ { "languageCode": "ENG",
                                                      "confidence": 0.9999994 },
                                                    { "languageCode": "ARA", 
                                                      "confidence": 4.7619238e-7 },
                                                    { "languageCode": "NLD",
                                                      "confidence": 7.2325456e-8 },
                                                    { "languageCode": "CHI_SIM",
                                                      "confidence": 3.0645523e-8 },
                                                    { "languageCode": "ITA",
                                                      "confidence": 8.6900076e-10 } ],
                              "words":
                                                  [ { "text": "Example",
                                                      "confidence": 0.99908227,
                                                      "boundingPolygon":
                                                                        { "normalizedVertices": 
                                                                                               [ { "x": 0.0664819944598338, 
                                                                                                   "y": 0.011666666666666667 },
                                                                                                 { "x": 0.22160664819944598,
                                                                                                   "y": 0.011666666666666667 },
                                                                                                 { "x": 0.22160664819944598,
                                                                                                   "y": 0.035 },
                                                                                                 { "x": 0.0664819944598338,
                                                                                                   "y": 0.035 } ]
                                                                        } ... "detectedLanguages":
                                                                                                [ { "languageCode": "ENG", 
                                                                                                     "confidence": 0.9999994 } ], ...

Document Classification

Document Classification can be used to classify a document.

Vision provides a list of possible document types for the analyzed document. Each document type has a confidence score. The confidence score is a decimal number. Scores closer to 1 indicate a higher confidence in the extracted text, while lower scores indicate lower confidence score. The range of the confidence score for each label is between 0-1. The list of possible document types is:
  • Invoice
  • Receipt
  • Resume or CV
  • Tax form
  • Driver's license
  • Passport
  • Bank statement
  • Check
  • Payslip
  • Other
Supported features are:
  • Classify document
  • Confidence score
  • Single request
  • Batch request
Document Classification Example

An example of document classification use in Vision.

Input document
Figure 3. Document Classification Input
Receipt from a fictitious cafe, including two line items, tax, subtotal and total amounts.
API Request:
{ "analyzeDocumentDetails":
 { "compartmentId": "",
   "document":
              { "namespaceName": "",
                "bucketName": "",
                "objectName": "",
                "source": "OBJECT_STORAGE" },
   "features": 
              [ { "featureType":
                  "DOCUMENT_CLASSIFICATION",
                  "maxResults": 5 } ]
 } 
}
Output:
API Response:
{ "documentMetadata":
 { "pageCount": 1,
   "mimeType": "image/jpeg" },
  "pages":
          [ { "pageNumber": 1,
              "dimensions": 
                           { "width": 361,
                             "height": 600,
                             "unit": "PIXEL" },
              "detectedDocumentTypes":
                                      [ { "documentType": "RECEIPT",
                                          "confidence": 1 },
                                        { "documentType": "TAX_FORM",
                                          "confidence": 6.465067e-9 },
                                        { "documentType": "CHECK",
                                          "confidence": 6.031838e-9 },
                                        { "documentType": "BANK_STATEMENT",
                                          "confidence": 5.413888e-9 },
                                        { "documentType": "PASSPORT",
                                          "confidence": 1.5554872e-9 } ],
 ...
               detectedDocumentTypes":
                                      [ { "documentType": "RECEIPT",
                                          "confidence": 1 } ], ...

Table Extraction

Table extraction can be used to identify tables in a document and extract their contents. For example, if a PDF receipt contains a table that includes the taxes and total amount, Vision identifies the table and extract the table structure.

Vision provides the number of rows and columns for the table and the contents in each table cell. Each cell has a confidence score. The confidence score is a decimal number. Scores closer to 1 indicate a higher confidence in the extracted text, while lower scores indicate lower confidence score. The range of the confidence score for each label is from 0 to 1.

Supported features are:
  • Table extraction for tables with and without borders
  • Bounding polygons
  • Confidence score
  • Single request
  • Batch request
Limitations are:
  • English language only
Table Extraction Example

An example of table extraction use in Vision.

Input document
Figure 4. Table Extraction Input
Fictitious balance sheet for eight quarters
API Request:
{ "analyzeDocumentDetails":
 { "compartmentId": "",
   "document": 
              { "namespaceName": "",
                "bucketName": "",
                "objectName": "",
                "source": "OBJECT_STORAGE" },
   "features": 
              [ { "featureType": "TABLE_DETECTION" } ]
 } 
}
Output:
Figure 5. Table Extraction Output
The balance sheet with cell, column header and row identifer highlighted
API Response:
{ "documentMetadata":
 { "pageCount": 1,
   "mimeType": "application/pdf" },
  "pages":
          [ { "pageNumber": 1,
              "dimensions": 
                           { "width": 2575, 
                             "height": 1013,
                             "unit": "PIXEL" },
 ... 
  "tables":
           [ { "rowCount": 15,
               "columnCount": 9,
               "bodyRows":
                          [ { "cells":
                                      [ { "text": "Qtr1-12",
                                          "rowIndex": 0,
                                          "columnIndex": 1,
                                          "confidence": 0.92011595,
                                          "boundingPolygon":
                                                            { "normalizedVertices": 
                                                                                   [ { "x": 0.2532038834951456,
                                                                                       "y": 0.022704837117472853 },
                                                                                     { "x": 0.3005825242718447,
                                                                                       "y": 0.022704837117472853 },
                                                                                     { "x": 0.3005825242718447,
                                                                                       "y": 0.05330700888450148 },
                                                                                     { "x": 0.2532038834951456,
                                                                                       "y": 0.05330700888450148 } ]
                                                             },
                                                               "wordIndexes": [ 0 ] },
                                        { "text": "Qtr2-12",
                                          "rowIndex": 0,
                                          "columnIndex": 2,
                                          "confidence": 0.919653,
                                          "boundingPolygon":
                                                           { "normalizedVertices":
                                                                                   [ { "x": 0.33048543689320387,
                                                                                       "y": 0.022704837117472853 },
                                                                                     { "x": 0.3724271844660194,
                                                                                       "y": 0.022704837117472853 },
                                                                                     { "x": 0.3724271844660194,
                                                                                       "y": 0.05330700888450148 },
                                                                                     { "x": 0.33048543689320387,
                                                                                       "y": 0.05330700888450148 } ]
                                                          }, "wordIndexes": [ 1 ] },
 ...

Key Value Extraction (Receipts)

Key value extraction can be used to identify values for predefined keys in a receipt. For example, if a receipt includes a merchant name, merchant address, or merchant phone number, Vision can identify these values and return them as a key value pair.

Supported features are:
  • Extract values for predefined key value pairs
  • Bounding polygons
  • Single request
  • Batch request
Limitations:
  • Supports receipts in English only.
Supported fields are:
MerchantName
The name of the merchant issuing the receipt.
MerchantPhoneNumber
The telephone number of the merchant.
MerchantAddress
The address of the merchant.
TransactionDate
The date the receipt was issued.
TransactionTime
The time the receipt was issued.
Total
The total amount of the receipt, after all charges and taxes have been applied.
Subtotal
The subtotal before taxes.
Tax
Any sales taxes.
Tip
The amount of tip given by the purchaser.
The supported line item information is:
ItemName
Name of the item.
ItemPrice
Unit price of the item.
ItemQuantity
The number of each items purchased.
ItemTotalPrice
The total price of the line item.
Key Value Extraction (Receipts) Example

An example of key value extraction use in Vision.

Input document
Figure 6. Key Value Extraction (Receipts) Input
Receipt from a fictitious cafe, including two line items, tax, subtotal and total amounts.
API Request:
{ "analyzeDocumentDetails":
 { "compartmentId": "",
   "document":
              { "namespaceName": "",
                "bucketName": "",
                "objectName": "",
                "source": "OBJECT_STORAGE" },
   "features":
              [ { "featureType": "KEY_VALUE_DETECTION" } ]
 } 
}
Output:
Figure 7. Key Value Extraction (Receipts) Output
The fictitious receipt with only specific lines and fields highighted
API Response:
{ "documentMetadata":
                     { "pageCount": 1,
                       "mimeType": "image/jpeg" },
                       "pages":
                               [ { "pageNumber": 1, 
                                   "dimensions":
                                                { "width": 361,
                                                  "height": 600,
                                                  "unit": "PIXEL" },
 ...
                                   "documentFields":
                                                     [ { "fieldType": "KEY_VALUE",
                                                         "fieldLabel":
                                                                      { "name": "MerchantName" },
                                                         "fieldValue":
                                                                      { "valueType": "STRING",
                                                                        "boundingPolygon":
                                                                                          { "normalizedVertices":
                                                                                                                 [ { "x": 0.0664819944598338,
                                                                                                                     "y": 0.011666666666666667 },
                                                                                                                   { "x": 0.3157894736842105,
                                                                                                                     "y": 0.011666666666666667 },
                                                                                                                   { "x": 0.3157894736842105,
                                                                                                                     "y": 0.035 },
                                                                                                                   { "x": 0.0664819944598338,
                                                                                                                     "y": 0.035 } ]
                                                                                           },
                                                                        "wordIndexes":
                                                                                      [ 0, 1 ],
                                                                        "value": "Example cafe" } },
 ...

Optical Character Recognition (OCR) PDF

OCR PDF generates a searchable PDF file in your Object Storage. For example, Vision can take a PDF file with text and images, and return a PDF file where you can search for the text in the PDF.

Supported features:
  • Generate searchable PDF
  • Single request
  • Batch request
OCR PDF Example

An example of OCR PDF use in Vision.

Input
Figure 8. OCR ODF Input
Page from a PDF document
API Request:
{ "analyzeDocumentDetails":
 { "compartmentId": "",
   "document":
              { "source": "INLINE",
                "data": "......" },
   "features":
              [ { "featureType": "TEXT_DETECTION",
                  "generateSearchablePdf": true } ]
 } 
}
Output:
Searchable PDF.

Using the Pretrained Document AI Models

Vision provides pretrained models for customers to extract insights about their documents without needing Data Scientists.

You need the following before using a pretrained model:

  • A paid tenancy account in Oracle Cloud Infrastructure.

  • Familiarity with Oracle Cloud Infrastructure Object Storage.

You can call the pretrained Document AI models as a batch request using Rest APIs, SDK, or CLI. You can call the pretrained Document AI models as a single request using the Console, Rest APIs, SDK, or CLI.

See the Limits section for information on what is allowed in batch requests.