Proso millet is a grain crop that has been increasingly popular in the food industry in recent years due to its numerous health benefits, while also enabling eco-friendly cultivation with its robust growth. During the processing, proso grains go through several stages of cleaning. The most significant losses occur in the final stage, which separates unhulled proso grains from the hulled proso millet, as it is challenging to distinguish them using sieves and suction systems, while manual sorting is extremely time-consuming.
We addressed this problem by developing a mechanical system that handles grains in a controlled manner and removes foreign particles, based on visual analysis of images. Image data is collected with an embedded computer vision system, specially designed for quality inspection. Images of grains are processed with one-class learning anomaly datection algorithms, which classify the grains as either good or bad. We first tested state-of-the-art methods PaDiM and CS-Flow. We further improved anomaly detection rates by combining the advantages of both approaches. As a result, we achieved less than 5 % hulled proso millet among the removed anomalies at 1 % unpeeled proso grains in the final product. Throughout the system's development, a dataset of hulled proso millet and unhulled proso grains was created.
|
