Cargando
Cargando
The media must manage a massive amount of information in record time, with a constant need to revisit historical material. Therefore, the proper accuracy and categorization of audiovisual material is a critical part of the business.
This task is so complex that there is a group of companies devoted specifically to it: news agencies. Among many other tasks, news agencies are responsible for producing large quantities of current material, which is properly cataloged, distributed and archived for further consultation. Making this huge and growing knowledge base useful requires detailed classification and categorization to enable effective exploration in the future.
Although this is applicable to a wide variety of documents, in this case we are interested in the processing of images, and specifically the processing and labeling of news photos.
News agencies are faced with the task of processing thousands of images every day, requiring complete teams of professionals to be dedicated to this task. This is expensive and produce poor results, as it is difficult to catalog images consistently given the diverse professionals involved.
Therefore, machine learning tools are needed to assist in this process to make it more robust, fast and economical.
The challenge for Gtres Online, a leading Spanish news photo and video agency, is to analyze images from different perspectives: the frame, the type of plane and even the types of garments present (jewelry, boots, coats, sunglasses, etc.).
The challenge presents various difficulties, including poorly represented categories and categories with frequent misclassification errors due both to human errors (mistaking swimsuits and bikinis, assigning wrong labels, missing labels) and to problems of subjectivity (it is not easy to consistently differentiate between the general plane or the full-length plane given the diversity of the human team).
Even so, Foqum’s image classification capability is quite robust in the face of this type of problem as long as certain best practices are followed. The results obtained were more than satisfactory with an accuracy over 90% for both the different types of frame and planes of the photo and over 95% for the most important labels for the customer, providing a fast, reliable and, what is more, consistent result.
As an added extra, an analysis function was proposed capable of indicating the area of the photo most relevant to each classification category, which is especially important if we believe that there is no type of label in this regard.
Remember: we cannot improve what we do not measure.
CONTACT US BOOK A STRATEGY MEETING