Phenotyping Problems of Parts-per-Object Count

Abstract

The need to count the number of parts per object arises in many yield estimation problems, like counting the number of bananas in a bunch, or the number of spikelets in a wheat spike. We propose a two-stage detection and counting approach for such tasks, operating in field conditions with multiple objects per image. The approach is implemented as a single network, tested on the two mentioned problems. Experiments were conducted to find the optimal counting architecture and the most suitable training configuration. In both problems, the approach showed promising results, achieving a mean relative deviation in range of $11\%$ 11 % – $12\%$ 12 % of the total visible count. For wheat, the method was tested in estimating the average count in an image, and was shown to be preferable to a simpler alternative. For bananas, estimation of the actual physical bunch count was tested, yielding mean relative deviation of $12.4\%$ 12.4 % .