Leaf Counting with Deep Convolutional and Deconvolutional Networks

Abstract

In this paper, we investigate the problem of counting rosette leaves from an RGB image, an important task in plant phenotyping. We propose a data-driven approach for this task generalized over different plant species and imaging setups. To accomplish this task, we use state-of-theart deep learning architectures: a deconvolutional network for initial segmentation and a convolutional network for leaf counting. Evaluation is performed on the leaf counting challenge dataset at CVPPP-2017. Despite the small number of training samples in this dataset, as compared to typical deep learning image sets, we obtain satisfactory performance on segmenting leaves from the background as a whole and counting the number of leaves using simple data augmentation strategies. Comparative analysis is provided against methods evaluated on the previous competition datasets. Our framework achieves mean and standard deviation of absolute count difference of 1.62 and 2.30 averaged over all five test datasets.

Pheno‐Deep Counter: a unified and versatile deep learning architecture for leaf counting

“… A paired t ‐test between our method and Aich and Stavness (the only two approaches from the CVPPP Workshop 2017) shows statistically significant differences ( p value <0.0001). …”

“…Note that the single input model of our deep architecture achieved the best results on the CVPPP 2017 dataset in the LCC. A paired t ‐test shows statistically significant gains when compared with Aich and Stavness () ( P ‐value <0.0001; last column of Table ). Figure collates results across all images as: (i) the correlation between ground truth and prediction, showing the high agreement of our method ( R 2 = 0.96); (ii) the distribution of error in leaf count, where it can be seen that in about 80% of cases the error is confined within the ±1 leaf range (for comparison Giuffrida et al .…”

“…We report quantitative results in Table , comparing our performance with other deep learning methods for leaf counting (Aich and Stavness, ) and leaf counting via segmentation (Romera‐Paredes and Torr, ; Ren and Zemel, ), as well as with the machine‐learning algorithm that won CVPPP LCC 2015 (Giuffrida et al ., ). The CVPPP 2017 dataset contains as a subset data from previous competitions, allowing comparisons across the years and methods (but not on all data).…”

“…We report quantitative results in Table 1, comparing our performance with other deep learning methods for leaf counting (Aich and Stavness, 2017) and leaf counting via segmentation (Romera-Paredes and Torr, 2016; Ren and Zemel, 2017), as well as with the machine-learning algorithm that won CVPPP LCC 2015 (Giuffrida et al, 2015). The CVPPP Note that the single input model of our deep architecture achieved the best results on the CVPPP 2017 dataset in the LCC.…”

Simulated Plant Images Improve Maize Leaf Counting Accuracy

“…Approaches to extracting quantitative or qualitative measurements of plant traits from images or other sensor data (such as point clouds produced by LIDAR sensors) can be broadly divided into two categories: 1) those where the programmer or scientist is responsible for telling the computer how to process the images [5] and 2) those where the computer learns how to process images from sets of annotated training data [6,7]. Both approaches require a combination of image/sensor data and ground truth measurements.…”