Neural Network-Powered Computer Vision and Data Infrastructure

Project

Neural Network-Powered Computer Vision

Project Requirements

Use AI to identify an electronics component using only an image sent from a customer, computer vision, and a database of millions of replacement parts.

Result

Early tests on a small subset of components have shown great promise, and scaling the solution has begun, though working with an 8 million part catalog is a challenge for the internal team since data must be collected and set up in a usable format. Processing times for images is also a key part of scaling the solution and is being monitored.

Early trial success has provided a window into how the tool will be rolled out, and how the parts specialist team will use the approved neural network to help identify certain categories of components. Capabilities will increase over time, assisting in forecasting future efficiencies and budget allocation.

Overview

Every day, the client’s 8000 employees fill three jumbo jets with electronics components for customers around the world. Speed and accuracy are high priorities for the client’s business. Indeed, many of their customers need electronics components shipped overnight to meet production quotas or repair downed systems. In turn, those quickly delivered components need to fit perfectly and match the specifications of the given design.

Currently, the client employs a team of parts specialists who are experts in electronics components. These specialists use their years of experience to identify parts and suggest replacements when customers have inquiries. However, this process can still be time-consuming, especially for rare parts or components with highly specific applications. In those cases, parts specialists must comb through hundreds or thousands of entries in the client’s catalog of over 8 million components.

To speed up the parts discovery process, the client has commissioned a special team to investigate the potential of using neural networks to classify images of parts. Ultimately, they hope that the customer can take a picture of the component they need to replace, and the neural network will be able to match that component to likely replacement components in the catalog.

Process

The High Tech client engaged Intertech to source expertise in neural networks and data infrastructure and work alongside their in-house expert on the small exploratory team.

Broadly, the process for developing a new solution falls into two key tasks. First, the team must create, train, and test a neural network for image recognition of electronics components to a high degree of accuracy. Second, the team must lay the infrastructure for storing, retrieving, and processing hundreds of millions of training images the neural network will use to learn the required detail.

To successfully identify an electronic component from a photo, the team began by taking images from all angles of a small set of test components. For each component, the team has 96-192 images, rotated in all possible directions. From there, they use the imgaug Python library to simulate various levels of distortion, saturation, brightness, and blur in the photos. Ultimately, the team uses PyTorch and TensorFlow models written in Python to create a neural network that then consumes the training images in order to learn. Tweaking and honing those models over time to achieve maximum accuracy is a significant challenge for this project.

Additionally, the team must find a way to store hundreds of millions of training images for all the components in the client’s catalog. Initially, they used Amazon Web Services for storage and processing, but they found it was too expensive to be feasible. So, the team ordered in custom hardware with high-end GPUs to do the image processing in-house.

Finally, the team is also looking for ways to optimize their training speed. It could take the neural network months to process a million images. A few thousand images already take several hours. As such, every small bit of efficiency the team can gain will matter to how quickly the model can be trained and retrained if new data or a better model becomes available.