Colour nowcasting: towards real time colour forecasting

Colour is a critical component of the fashion industry. Traditional mass manufacturing processes over the last 50 years has been supported by colour forecasting, where agencies attempt to predict which colours will be popular with consumers in about 18 months’ time. However, advances in technology are changing the way in which products are manufactured, resulting in much shorter supply chains driven by rapid agile manufacturing processes. These changes are putting pressure on tradition colour forecasting methods. Increasingly, companies are seeking to use the vast amounts of data that are now available to make more accurate forecasts and to produce these forecasts more quickly. This work focuses on the fashion and clothing industry and explores methods to automatically extract colour palettes from fashion images. A substantial amount of the ground-truth data has been produced using psychophysical experiments and various machine-learning models have been tested on various sets of images. A clustering method known as k-means is at the heart of all of the models used in this thesis. However, various other approaches such as people detection, background removal and semantic segmentation have been implemented and evaluated. Finally, it has been shown that a pixel-level semantic segmentation approach (with subsequent cluster analysis) can generate colour palettes that are indistinguishable from those generated by humans. The implications of this work to various design-related tasks such as colour nowcasting, colour decision supports and the extraction of colour palettes from mood boards are discussed.