Developing an AI-based image editing solution for one of India’s largest online jewelry marketplaces- eJohri

Team Daffodil analyzed the existing business challenges and as a solution, tried computer vision techniques to remove the background from the images. Edge detection, image thresholding, unsupervised learning (K-clustering), etc. were some of the techniques that were examined individually and in combination to resolve the problem. However, these techniques worked differently on different types of jewelry and posed accuracy concerns. For that reason, computer vision techniques were dropped and machine learning models were chosen as the solution to resolve the problem.

One of the prime challenges in building the ML model was the limited data set. To augment the data set and train the model for maximum accuracy, the transfer learning technique was utilized. This allowed team Daffodil to train the model on a generalized data set of jewelry.

This fundamental model training overcame the limitations of the finite data set available. The machine learning model thus created had an accuracy rate of 95% and saved several man-hours for the eJohri team in the image editing process.