Parts-per-object count in agricultural images: Solving phenotyping problems via a single deep neural network

Solving many phenotyping problems involves not only automatic detection of objects in an image, but also counting the number of parts per object. We propose a solution in the form of a single deep network, tested for three agricultural datasets pertaining to bananas-per-bunch, spikelets-per-wheat-spike, and berries-per-grape-cluster. The suggested network incorporates object detection, object resizing, and part counting as modules in a single deep network, with several variants tested. The detection module is based on a Retina-Net architecture, whereas for the counting modules, two different architectures are examined: the first based on direct regression of the predicted count, and the other on explicit parts detection and counting. The results are promising, with the mean relative deviation between estimated and visible part count in the range of 9.2% to 11.5%. Further inference of count-based yield related statistics is considered. For banana bunches, the actual banana count (including occluded bananas) is inferred from the count of visible bananas. For spikelets-per-wheat-spike, robust estimation methods are employed to get the average spikelet count across the field, which is an effective yield estimator.