Conventional CNN-based colour detection mechanisms are not optimal due to colour overlap, for example, the CNN model can find it difficult to distinguish shades of red and orange. The proposed approach is aimed at optimizing colour detection in attribution processes using a superior, objective and mathematical approach. The proposed approach allows you to detect a much larger range of shades of colour while increasing colour detection accuracy and reducing the time it takes to predict colour without model training. The proposed approach performs better than the conventional CNN-based supervised machine learning colour detection owing to its enhanced quantifiability, increased degree of quantisation of the colour palette (i.e., an improved colour range that can be detected) and lower the complexity of time. The approach employs fundamental concepts of image processing to isolate pixels that best represents the object's colour from the region of interest (RoI), following which a mode filter is applied to the RGB pixel values of the pixels lying in the RoI to obtain a single set of [R, G, B] values that best represents the colour of the object. The mapping of this set of RGB values to the name of colour defined in a “Colour Taxonomy” is performed using Euclidean or Manhattan distance, depending upon the application. The accuracy of colour prediction is enhanced and computing time is greatly reduced. The current approach is aimed to be more accurate, more granular in detection (i.e., allows you to detect a much larger range of shades of colour and more time-effective than CNN-based colour detection and manual tagging thus greatly helping optimize attribution processes.
All Science Journal Classification (ASJC) codes
- Materials Science(all)