Computer Vision Damage Detection

Working at TVS Supply Chain Solutions (TVS SCS) was my first professional job after university. I was excited to apply what I had learned at university to this setting. I think this role, and even more so this first project, was very fitting to a graduate, it allowed me to handle projects end-end without diving head-first into a messy ancient code-base getting overwhelmed and not learning much.

Near the start of my tenure, we had a tour of one of the warehouses. There was a very large section dedicated to smart meter triage; Damaged smart-meters would come in and it was the responsibility of the worker to assess the damage and manually log in the machine the way in which the smart meter was damaged, it could be for a variety of reasons, such as: cracked screen, burnt parts, missing components, etc. The purpose of the role was to investigate potential improvements possible with AI as an emerging technology, this seemed like a prime use case.

This manual imputation could clearly be accelerated by some computer vision. The smart meter triage line was already using a large white camera box to take pictures of the meters and send them back to the manufacturer. If I could use these images to train a model, I could predict the damages in a simple classification model. Each side of the box had a camera (6 total), I opted to a 'Branched Convolutional Neural Network' to train and predict incoming images. Each image passes through convolutional and batching layers until they have a single dimension, at which point the six feature sets are concatenated into one and passed through a Multi-layer Perceptron (MLP) to obtain a classification.

The images and result were passed through to each other using an API, as well as logged on the server the model ran on. This showed both office and warehouse staff the results in real time. This project massively improved warehouse efficiency allowing staff to parallelise the meter triage process. The project also became the basis of future machine vision classification tasks.