Global beauty company Coty collects many images of fragrances, photographed by consumers. To be useful for any analytics purpose, these pictures need to be labelled and associated with a brand and a name. This process is very manual and requires Coty employees to spend a large amount of time going through thousands of pictures.
Our challenge was to automate this task as much as possible with image recognition algorithms.
Like in most image recognition applications, the algorithm must be trained appropriately before the model can identify and label pictures.
To achieve this, we selected and trained the algorithm with a stratified sample of 42 pictures with different quality standards – ranging from ‘very well taken’ to almost unrecognisable. To replicate a real-life scenario, we added 2,000 randomly selected pictures unrelated to the fragrances we were trying to recognise.
To reach a minimum 60% accuracy level, we trained and tested several algorithms.