S18 - Session O5 - Intelligent imaging methods to assess quality traits in bush bean breeding

Authors: Dirk Jollet *, Mark Müller-Linow

In the field of bush bean breeding, we observe an increasing demand for rapid and objective assessment of yield and quality parameters of bean pods, due to the fact that handmade measurements and visual scorings are costly, time consuming, and subjective, making the phenotyping process the bottleneck in many breeding programs. To analyze the bean pods we used a customized acquisition box that was introduced in Jollet et al. (Acta Hort., accepted) in the context of length and caliber measurements, complemented it with a backlight source for the detection of missing grains and added an additional luminance and color calibration for a rough assessment of both color and brightness features. We completed the processing pipeline with estimations of the pod curvature, all of them now calculated on the basis of a new classification approach for the pod regardless of tips or peduncles. We validated the computer vision estimations with manual measurements from an expert breeder, whereas the test data set covered measurements of 400 single bean pods (n=400). This dataset also included measurements of metric pod length, which needed to be re-validated due to an improved automatic identification of apical and basal pod ends. Therefore, a Mask Region Based Convolutional Neural Network (Mask R-CNN) was trained to detect tips at the apical end and check for peduncles at the basal end, and delete them by overlaying the predicted masks in the image. Peduncles together with the distal tips need to be excluded from the pod length measures with respect to marketing requirements.