Google's commitment to machine learning and artificial intelligence is considerable, as explained in the article What is the significance of ML for Google? The influence on SEO is also clearly visible, as explained in the article What is the significance of machine learning for SEO? Experts dare to look into the future .
Image recognition and the identification of texts and their meaning play an important role for Google.
Are the days of unreadable text in images soon over? I asked myself denmark phone number data this question while following the news in the area of image and text recognition from Google.
At the end of this post I have compiled a selection of Google patents related to image and video recognition and interpretation.
Thanks to machine learning and the availability of suitable learning data, the development of existing systems is rapid. The question is, how advanced are Google's products and can this provide information about the possible use of the Google crawler ?
Difficulties in the readability of texts in images
The problem with the development of identifying and interpreting text in images is not the feasibility or the patents, but the error-free implementation and the meaningful further processing. This is because an intelligent abstraction ability would have to be included, which Google is currently still training .
The problem to be solved in the case of text recognition in images is distinguishing between real text and supposed text, such as structures that are shaped like a letter or a number and, in the worst case, are in close proximity to real characters. For the crawler, this would mean that it would have to distinguish between text and elements without error , transcribe them and assign them to the image as a tag or create sitelinks from them or determine the true author of the website. With an imprint and a simple image, very little abstraction is necessary and the implementation works very well, but how do you separate the different image types from each other?
It would be possible to achieve this using machine learning and by comparing with similar images that have already been verified. It would also be possible to compare transcribed data with typical keywords or addresses in order to establish a connection to companies and to correct minor errors in the survey.
Another problem with text recognition in images is the weighting of the automatically created tag and the automated thematic classification. This process, which is called generalization, decides how relevant the text is in relation to the image motif and what semantic relationship the text has to the image. The comparison with image captions and manual keywording must also be taken into account in order to determine relevance if necessary. In the worst case, your vacation photo on the beach would be associated with the name of the local garbage can manufacturer. There was already an explosive scandal in image recognition and automatic classification this year .
It would be easier with images such as detailed infographics that would provide enough text for Google to classify, or data such as addresses that have a relatively fixed structure. Web 3.0 will play a key role in this problem in conjunction with image and object recognition such as the Cloud Vision API and deep learning . The publication of Google's API and its use in apps will in turn lead to even larger data volumes and multifunctional applications. The process can be refined with the confirmation of calculated scoring and classifications by the user. In this way, existing systems continue to develop.
In summary, the question is not only about ultimate feasibility, but also what Google will/can ultimately do with the data collected in this way, or what Google sees as a sensible use of the data. Ultimately, it's about generating tangible added value for users, and this only works with lead time and the interaction of several systems. In order to test the theoretical ideas on a real product, here is a small study.