Retail companies want to show customers products which are similar to the ones bought in the past. This type of problems can be solved by ranking images according to their similarity and do a content-based recommendation.
The Clothes texture data sets collected from search engine are used as example for building image recommendation models. The dataset used in this demonstration contains clothes texture images with three labels, dotted, leopard, and striped. The goal is to build a recommendation model to recommend new clothes images to users.
This pre-built model has used the Python language to build a recommendation model to represent the knowledge discovered leveraging CNN for image featurization and SVM for similarity ranking. The knowledge representation is easy to understand. The computation time is much shorter on GPU based DLVM (1 minutes) than that on CPU based DSVM (hours).
The print will display a textual summary of the pre-trained CNN model and its build parameters.
The demo applies the pre-built model to a demo data set with around 60 images and shows the top-1 recommendation results.
The display will visualize the image recommendation results.
Run the score command to provide recommendation for new clothes texture data.
The following link provides a deep guidance on how the model is built using Microsoft Cognitive Toolkit (CNTK). [<https://github.com/Azure/ImageSimilarityUsingCntk>]