This talk describes a low budget, production grade data pipeline for image analysis. Producing good bacon is very time consuming and costly. Pork processors demand certain qualities from the hogs they buy and hog growers rely on, among other things, good genetics to meet those demands. Good genetics relies on good data.

Smithfield Foods, the largest pork producer in the U.S., has been using deep learning models in production for several years in order to identify the best genetics that will produce top quality pork products including bacon and loins. Ultrasound images taken on high value breeding stock are analyzed for characteristics related to meat quality and the results of those analyses are used to make crucial breeding decisions. Training a model to make accurate predictions is relatively simple and inexpensive given the variety of open source solutions available today.

This talk focuses more on an automated pipeline developed to implement the deep learning models into our hog production system. Various challenges had to be resolved in order to overcome a low budget and lack of internal technical skill. Challenges included communication issues across multiple platforms and leveraging the available coding skills in different languages: Python, R, SQL and others. The entire system, including image analysis, was built by a small team of biologists; not data architects; not IT professionals.