Vision release notes

Safe Search model update

We will be updating the SAFE_SEARCH_DETECTION feature model to improve quality.

We'll support both the current model and the new model for the next 90 days. After 90 days, the new model will become the default. The current model can still be accessed by specifying "builtin/legacy" for an additional 90 days before it's deprecated.

To use the new model, specify "builtin/latest" in the model field of a Feature object.

New label detection model

An improved model is now available for Label Detection. Along with the improved model, the topicality field is now populated correctly.

Specify "builtin/latest" in the model field of a Feature object to use the new model. We'll support both the current model and the new model the next 90 days. After 90 days, the new models will become the default. The current models can still be accessed by specifying "builtin/legacy" for an additional 90 days before they are deprecated.

Updated feature models

Improved models are now available for the following features:

• Text detection and documentation text detection (OCR)
• Web detection
• Logo detection
• Object localization

Specify "builtin/latest" in the model field of a Feature object to use the new models.

We'll support both the current model and the new model the next 90 days. After 90 days, the new models will become the default. The current models can still be accessed by specifying "builtin/legacy" for an additional 90 days before they are deprecated.

Change in OCR model behavior when detecting line break hyphens

Soft hyphens are used for the purpose of breaking words across lines. The previous ("builtin/legacy") version used the BreakType HYPHEN to represent a soft hyphen. The new stable ("builtin/latest") version represents it as a new symbol with literal text "-" and BreakType EOL_SURE_SPACE.

Consider the following example text and OCR model output:

Sample text

"Mr. White has had considerable experience as a Veteri-

nary surgeon, and will attend practice in that line."

Output ("builtin/legacy" model):

[...]
{
  "property": {
    "detectedBreak": {
      "type": "HYPHEN"
    }
  },
  "boundingBox": {
    "vertices": [
      [...]
    ]
  },
  "text": "i"
}
[...]

Output ("builtin/latest" model):

{
  "boundingBox": {
    "vertices": [
      [...]
    ]
  },
  "text": "i"
},
{
  "property": {
    "detectedBreak": {
      "type": "EOL_SURE_SPACE"
    }
  },
  "boundingBox": {
    "vertices": [
      [...]
    ]
  },
  "text": "-"
}

Landmark Detection Upgrade

Specify "builtin/latest" in the model field of a Feature object to use the new model.

We'll support both the current model and the new model the next 90 days. After 90 days the current model will be deprecated and only the new model will be used for all landmark detection requests.

Face Detection Upgrade

Specify "builtin/latest" in the model field of a Feature object to use the new model.

We'll support both the current model and the new model the next 90 days. After 90 days the latest model will become the default model. The old model will be available for another 90 days by specifying "builtin/legacy".

OCR model migration

The TEXT_DETECTION and DOCUMENT_TEXT_DETECTION models have been upgraded to newer versions. The API interface and client library will be the same as the previous version. The API follows the same Service Level Agreement.

The legacy models can still be accessed until August 20 2022. Specify "builtin/legacy" in the model field of a Feature object to get the old model results. After August 20, 2022 the legacy models will no longer be offered.

OCR model migration reverted

We have switched the "builtin/stable" model back to the original version temporarily while we fix a bug resulting from this migration. The week of May 16th, we will update the "builtin/stable" model used for OCR again with the model from "builtin/latest" and create a new release note.

You will be able to use the original model as "builtin/legacy" for 90 more days after we upgrade "builtin/stable".

OCR model migration

The TEXT_DETECTION and DOCUMENT_TEXT_DETECTION models have been upgraded to newer versions. The API interface and client library will be the same as the previous version. The API follows the same Service Level Agreement.

The legacy models can still be accessed until August 02 2022. Specify "builtin/legacy" in the model field of a Feature object to get the old model results. After August 02, 2022 the legacy models will no longer be offered.

OCR model update

We have updated the "builtin/latest" OCR model with quality improvements. Consequently, customers can continue to test this model for 90 additional days.

Please note that you have 90 days from today to test the new model by specifying "builtin/latest" in the model field of the Feature object. At the end of that period, it will be promoted to the default model accessible as "builtin/stable". After that event, the original models will still be available for another 90 days using "builtin/legacy". If you encounter problems with this upgrade, please contact Vision API engineering team by submitting a ticket in the private issue tracker.

For the original announcement of this change, see the October 1, 2021 release note.

Region forwarding from the global to the regional endpoint has been deprecated. For more information, see the October 1, 2021 release note.

OCR Model Update

An improved model is now available for Text Detection (OCR). The new model can be used with TEXT_DETECTION and DOCUMENT_TEXT_DETECTION features. The same model is used for requests sent to both features. With the new model, the distribution of confidence scores of responses will change. For more information, see Service announcements.

Please note that you have 90 days from today to test the new model by specifying "builtin/latest" in the model field of the Feature object. At the end of that period, it will be promoted to the default model accessible as "builtin/stable". After that event, the original models will still be available for another 90 days using "builtin/legacy". If you encounter problems with this upgrade, please contact Vision API engineering team by submitting a ticket in the private issue tracker.

Deprecate region forwarding In 90 days, specifying the location "us" or "eu" in the request to the global endpoint vision.googleapis.com will no longer be supported. Instead you should directly call the "us" or "eu" region endpoints (us-vision.googleapis.com or eu-vision.googleapis.com). You can find more information in the Multi-regional support section of the feature pages.

New multi-regional support for features

The Vision API now offers multi-regional support (us and eu) for the LABEL_DETECTION and SAFE_SEARCH features.

EXIF rotation featured fixed

EXIF rotation is now disabled.

For more information, see the March 8, 2021 release note.

EXIF rotation feature fix