Articles
Computer vision for carrot root phenotyping on smartphone images taken during crop evaluation
Article number
1393_31
Pages
239 – 246
Language
English
Abstract
Within seed companies, many processes are based on human visual expertise: from the phenotyping of harvested carrot roots to the evaluation of germination tests through visual evaluation of contamination of the seed batch with weed seeds.
These visual analyses often consider many characters that can be subjective.
This generates variability in judgment between people (inter person variability), but also variability for the same person (intra person variability). Therefore, seed companies have invested in computer vision approaches, making it possible to provide quantified data, in addition to accelerating and modernizing processes.
In this paper we will focus on the phenotyping of harvested carrot roots.
Our objective is to be able to phenotype carrots using images taken in the field under current crop evaluation conditions by breeders and technicians, with smartphones or standard cameras, without particular sensor settings or light conditions.
When developing computer vision methods, and in particular deep learning ones, a significant amount of manually annotated images are required.
However, the annotation work needed to create the training data sets is time-consuming, tedious, and a bottleneck.
We took advantage of existing manual annotations of a given species for the benefit of another.
We have developed a CycleGAN pipeline that efficiently turns cucumbers into carrots and creates a synthetic carrot data set for instance segmentation.
The network is able to change color, texture and shape during translation.
With this work we were able to constitute a sufficient training data set by mixing images incorporating many synthetic carrots and images of real harvested carrots.
This allowed us to develop a performing carrot root segmentation model using Mask RCNN. After segmentation, many phenotypic traits are extracted, such as number of carrots, length, diameter, shape, and many others.
Our segmentation model shows an F1 score of 93% in segmenting carrots correctly.
These visual analyses often consider many characters that can be subjective.
This generates variability in judgment between people (inter person variability), but also variability for the same person (intra person variability). Therefore, seed companies have invested in computer vision approaches, making it possible to provide quantified data, in addition to accelerating and modernizing processes.
In this paper we will focus on the phenotyping of harvested carrot roots.
Our objective is to be able to phenotype carrots using images taken in the field under current crop evaluation conditions by breeders and technicians, with smartphones or standard cameras, without particular sensor settings or light conditions.
When developing computer vision methods, and in particular deep learning ones, a significant amount of manually annotated images are required.
However, the annotation work needed to create the training data sets is time-consuming, tedious, and a bottleneck.
We took advantage of existing manual annotations of a given species for the benefit of another.
We have developed a CycleGAN pipeline that efficiently turns cucumbers into carrots and creates a synthetic carrot data set for instance segmentation.
The network is able to change color, texture and shape during translation.
With this work we were able to constitute a sufficient training data set by mixing images incorporating many synthetic carrots and images of real harvested carrots.
This allowed us to develop a performing carrot root segmentation model using Mask RCNN. After segmentation, many phenotypic traits are extracted, such as number of carrots, length, diameter, shape, and many others.
Our segmentation model shows an F1 score of 93% in segmenting carrots correctly.
Authors
C. Torres, D.N. Diaz Estrada, M. Kresović, O. Robert
Keywords
computer vision, Mask RCNN, GANs, CycleGan, carrot, phenotyping
Online Articles (40)