Real-Time Image Based Plant Phenotyping Using Tiny-YOLOv4

Sonal Jain, Dwarikanath Mahapatra, Mukesh Saini

Abstract: Image-based plant phenotyping is getting considerable attention with the advancement in computer vision technologies. In the past few years, the use of deep neural networks (DNNs) is well-known for segmentation and detection tasks. However, most DNN-based methods require high computational resources, thus making them unsuitable for real-time decision-making. This study presents a real-time plant phenotyping system using leaf counting and tracking individual leaf growth. For leaf localization and counting, a Tiny-YOLOv4 network is utilized, which provides faster processing, and is easily deployable on low-end hardware. Leaf growth tracking is performed by active contour segmentation of leaf localized using the Tiny-YOLOv4 network. The proposed system is implemented for top-view RGB images of the Arabidopsis thaliana’ plants. And its performance for leaf counting is evaluated against Tiny-YOLOv3 and Faster R-CNN using the difference in count (DiC), accuracy, and F1-score measures. The model achieves an improved accuracy of 90%, absolute DiC of 0.42, F1-score of 96%, and inference time of 15 milliseconds. Further, the segmentation accuracy measures using Dice and Jaccard scores are 0.91 and 0.86, with a computing time of 0.96 s. These obtained results depict the effectiveness of the proposed system for real-time plant phenotyping.