Deep Learning Segmentation in Bulk Grain Images for Prediction of Grain Market Quality

Abstract

Pulses are classified for market value based on visual criteria, where the grain parcel is downgraded if the presence of visually defective grain exceeds a predetermined threshold. Assessments of grain quality are usually performed subjectively by manual inspection at the first point of sale. In the current work, a machine vision approach for the assessment of grain quality with use of bulk grain images was proposed. A deep learning model was used to train semantic segmentation masks for possible classes of grain, including sound and defective grain within lentil and chickpea samples. In a subsequent step, the extracted segmentation mask areas were used to develop calibrations for the %w/w o..