Automated screening of canola seeds using GPU-based deep learning: A functional prototype

点击翻译按钮获取中文摘要

This Engage Plus Grant gives 20/20 Seed Labs, an Albertan company, access to unique knowledge andexpertise available in Dr. Dileepan Joseph's lab at the University of Alberta. The purpose is to address acompany-specific problem, namely the automated screening of canola seeds using optical microscopy.Canola, bred from rapeseed in the 1970s and continuously improved thereafter, is a Canadian innovation ofworldwide importance. In Canada alone, the canola industry "generates $19.3 billion annually in economicactivity and nearly a quarter of a million jobs," according to the Canola Council of Canada. Moreover, "canolaoil has the science to prove its benefits on [critical] health issues," and canola meal is an excellent livestockfeed to help meet the world's growing demand for protein.To meet trade requirements, seed purity is always a concern. The Canadian Food Inspection Agency hassome of the highest standards in the world regarding seed and 20/20 Seed Labs is the first private lab in Canadato be accredited by the International Seed Testing Association. Automating visual aspects of purity analysis,i.e., to identify unwanted seeds that resemble canola, is a prerequisite for automated cleaning, something thatcurrently is only partially accomplished by colour sorting.In a previous Engage Grant, this industry-university partnership successfully demonstrated the feasibility ofdiscriminating canola from six unwanted species using multilayer convolutional neural networks, an approachalso known as deep learning. As expected, parallel processing proved important and, after some investigation,graphics processing units, i.e., GPUs, appeared to be most suitable.The objective of this Engage Plus Grant is to make the above proof of concept functionally useful forautomated purity analysis. In addition to improving software delivered by the Engage Grant, entitled PurityAnalyzer 1.0, a prototype comprising both software and custom hardware will be designed, implemented, andtested to perform automated screening of canola seeds in real time.