Catalogue management is a very important aspect in the field of ecommerce as it helps the visitors in efficiently selecting the necessary interest items. In every retail website, all the items in the catalogue are in a particular order and orientation of different categories whose manual grouping and ordering takes a lot of time. Secondly, image quality assessment plays a very important part in catalogue management since the quality of the images sent by the vendor are not always of adequate quality which when displayed on the website results in customer dissatisfaction.
In this work, we have developed an entire pipeline where the first task is to automatically classify the various orientations (front view, side view, top view etc.) of the images sent by the vendor using transfer learning. In the second part of our pipeline, we have eased the process of catalogue management with the image quality assessment of the vendor images using Structural similarity index and Deep learning techniques.
The vendor-sent images act as an input to the pipeline from which the image embeddings are being extracted. Histogram of oriented gradients, multiple pre-trained CNN features, Kernel PCA embeddings are being extracted from the image in the task of image orientation classification with Multinomial Logistic regression achieved an validation accuracy of 95% which is significantly higher than baseline models. Secondly, for image quality assessment ,various types of noise signals were added to synthetically generate noisy dataset and image embeddings previously computed were used to classify the same using Deep learning model . Bottleneck features from the deep learning model have been extracted which can capture the quality aspects of the images. Structural Similarity indices (SSIM) were computed for each of the noisy image from the reference image and using the quality embedding as features and SSIM as response, a ridge regression model has been fitted which predicts the image quality in terms of SSIM given an image with an accuracy of 84%.
This model has been tested for various categories in USA. Automated image-orientation classification will help the business hugely to improve in terms of time required to manually classify them.
Moreover the image quality assessment of the images sent by vendor would help the business to properly manage the catalogue and reduce customer dissatisfaction which can happen due to poor image quality displayed. Also it will help in supplier evaluation.
Also, SSIM gives human perceived quality understanding which was very important for business to understand.
Challenges overcame :
For many classes of items, the dataset size is small in which majority of the modelling technique fails and hence we built a robust methodology which can handle minority class problem as well.
Popular image comparative metrics like MSE,SNR etc failed to give human perceived quality and hence we went for SSIM.
Souradip Chakraborty currently works as a Statistical Analyst in Walmart Labs in International Sourcing domain. He has over 3 US patents filed (as a part of Walmart Labs) in the field of AI and Machine Learning applications in Retail Domain. He has also been a part of ANZ Bank, Tata Steel Research and Development and Petrabytes in the field of Data Science.
He did his Masters from Indian Statistical Institute Bangalore (Batch Topper) and have publications in ISI Journal in the field of Multivariate Statistics and Quality Control.
Apart from that he has several publications in the field of Advanced Recommender systems using Natural Language Processing ,Computer Vision and Mathematical Morphology , also an IEEE publication in Stability Analysis of Network Control System.
Did his graduation from Jadavpur University in the field of Electronics engineering and has worked as a research analytical engineer in Amec Foster Wheeler Pvt Ltd for 2 years and received special mention from Clients like Exxon Mobil.