Wheat spikes counting using object-level data augmentation

Counting the number of wheat spikes in the field is crucial for yield estimation and is one of the most important indices for the breeders. To count the wheat spikes efficiently, there is a high demand for an accurate plant phenotyping algorithm. Previous studies showed that machine vision is one of the powerful tools in localizing and counting different objects in an image. However, the proposed algorithms need a considerable workforce to capture appropriated images and annotate them. While capturing the wheat field images is considered an easy task, annotating lots of small spikes is a highly time-consuming and labor-intensive task. The goal of this research is to propose a novel object-level data augmentation algorithm that significantly reduces the number of required training samples for the employed deep learning model. In the object-level augmentation algorithm, the input image decouples into individual objects and then augmented individually according to their label’s augmentation pipeline (spike, leaf, stem, and background). Finally, in the image inpainting phase, the augmented objects are used to generate the augmented images. The results demonstrate that using only one training image, the OLA-based model is able to localize and count the wheat spikes with R², RMSE, and F1 of 0.805, 2.652, and 0.876, while with image-level data augmentation, the model reached to the poor results of 0.05, 5.841, and 0.799, respectively.