AI Inventory Analysis

During my time writing articles for New Elementary, I found the manual process of transcribing and auditing item inventory pages to be extremely time and labor intensive. With an interest in exploring recent advancements in machine learning, I set about training my own local model to accurately perform the task of processing inventories.

Synthetic Data Generation Pipeline

Synthetic Data Generation

Prior to training the model, I first had to generate synthetic training data to help it accurately recognize the structure of inventory pages.

Using publicly accessible digital instruction data (available in PDF format), I was able to effectively automate the process of generating annotation data indicating the bounding boxes and textual contents of individual inventory entries on a page.

During the testing process, I quickly discovered that the pristine pages from the digital instructions were inadequate for representing the imperfect printing of physical instructions.

To remedy this, I developed a system for applying a custom image filter to simulate the CMYK printing process, including halftone shading, fibers, and gamma distortion.

Computer Vision

The annotation and filtered image data was subsequently used to train a local machine learning model, based on the PyTorch framework.

After several rounds of local training, the model was capable of recognizing paired quantity and element ID information with 96% accuracy across images of varying quality.

This system helped to dramatically accelerate the delivery of articles for New Elementary,  and allowed me to do in minutes what would previously have taken hours.