Barley Variety Recognition with Viewpoint-Aware Double-Stream Convolutional Neural Networks

Varietal homogeneity is an important factor in quality of malting barley, but its inspection is difficult. Biochemical methods are expensive and inefficient, while machine vision suffers due to high variability of the grains' features. In our previous work, we have shown a convolutional neural network for simultaneous feature extraction and classification of image data basing on multiple views. It was suggested that for machine vision inspection, the observed side of a grain should be taken into account - dorsal and ventral sides of each kernel exhibit different features. In this study we present a viewpoint-aware convolutional neural network, which learns to extract specialized features from images of dorsal and ventral sides of barley grains. We show that it increases the average classification accuracy by 0.6% and sensitivity by 2.3% with respect to the viewpoint-ignorant architecture on our dataset.