The client daily updates nearly 10 000 positions in the e-commerce store. Goods originate from different vendors in different formats. Website administrators spend a substantial amount of time to parse and categorize vendor files. The task was to enhance the administrator panel UX by implementing noise-resistant automatic categorization and simplification of goods accounting. This was expected to allow for the reduction of expenses and providing a clear competitive advantage. Another task was to develop a reliable price prediction for new products.
AI for photo categorizations
AI/ML Software, E-commerce
Research and Development
Tensorflow, TensorFlow Hub, Python, Keras, Google Colab, Google Drive, Django
The client collected around 150 000 images of goods with proper labels over the number of years. As a result, an extensive database divided into the training and validation sets was created. First of all, we examined existing computer vision solutions for classifying objects by photos. With the client training set, we have improved the object detection and classification algorithm to be noise-resistant to images with more than one object. Our team built a clustering algorithm based on the knowledge and specifics of the customer’s categorization hierarchy. Additionally, we implemented an algorithm persistent to changes in images quality, lighting and visual angle. The category detection pipeline was completely automated and was removed from the manual categorization workflow.
We were not able to use a popular image recognition datasets include CIFAR, ImageNet, COCO and Open Images as these datasets are rather applicable in academic research contexts, then for the client’s products images. As such, we decided to be careful when generalizing models trained on them. We used google drive as an intermediate storage for all existing 150 000 images. Originally we used an inception_v3/feature_vector tensorflow model without the top classification layer. The latter helped us to transfer learning and increase an accuracy by training model using existing images and product labels. We also rented a google colab, which offers free GPUs to train the model faster. As a final step we exported a new model with building an API around. This, in turn, enabled us to connect it to the web application.
Our client moved to the AI workflow for goods categorization and price prediction. The categorization algorithm was able to identify the correct category in 95% cases. While there was a lot of space for improvement, we learnt many useful insights. The main definition of success was that the e-commerce store administration staff began to spend less time on goods accounting and was able to concentrate on the interaction with customers and selling process.
Do you consider building a similar product?
Let's book a short meeting and immediately discuss with you all the details about your project.Book a call