Image recognition for product identification
RS Components – Image Recognition
RS Components, the number one high-service distributor of electronics and maintenance products across Europe and Asia Pacific.
Create an image recognition framework that can form the basis of product identification services across a range departments, tools and services offered by RS Components. Enabling their long-term strategic goal to be the market leader in customer experience and innovation.
How we did it
H&C partnered our machine learning and image recognition experts with the value added services team and sales team to form a comprehensive taskforce.
We held a workshop for the entire taskforce and discussed different approaches and data requirements. After reaching a collaborative consensus, we established initial user stories to guide the development process.
We designed and built an image recognition system and accompanying prototype self-service kiosk with imaging capabilities. These then went through a series of iterative stages to refine and improve, all the while building on user feedback. A staged rollout to the market gathered early user feedback and generated additional sample images which were used for further training.
Product identification using imagery differs from conventional software, by requiring a two-stage development approach of training followed by runtime-operation.
Developing product identification system by using images is a hard problem to solve. It’s particularly challenging when classifying user generated images taken under random conditions. To achieve more accurate classification we undertook a multifaceted approach. First, we enlarged the training data set of images, by applying transformations which mirror common photo errors, including being out of focus, unwanted shadows, distortions etc. This data set was then added to the original, and we used it as a basis for training the image recognition neural nets.
We then developed a multi-layered image recognition framework which utilised TensorFlow and convolutional neural networks. This base capability was then supplemented with the addition of bespoke colour extraction and object size detection layers, before applying a final OCR detection layer.
By utilising these multiple layers within the image recognition system, we were able to significantly increase the capability and accuracy that the system achieved.
In addition, the image recognition system has been designed as a self-learning system. Which enables it to use each and every interaction to expand its training set, making the system smarter and more accurate as more customers use it.
The image recognition system aims to increase sales by providing faster and more accurate identification of unknown products. Allowing customers to easily purchase the products they need in a more efficient and automated process.