Automatic Leaf Counting Using Improved YOLOv3

In the recent years, image-based plant phenotyping has become an active research area. It can help investigating the complex association between genomes and phenotypic traits. However, traditional plant phenotyping involves manual measurement of plant traits and is quite time-consuming. With the help of automatic image analysis, we can significantly reduce time and human labor in this process. Working towards the ultimate goal of automatic plant phenotyping, we begin with developing an automatic method for leaf counting. The existing approaches for image-based leaf counting can be divided into two categories. One is based on instance segmentation and the other one is based on regression prediction. In contrast to the existing approaches, we propose to utilize object detection techniques. In particular, we modify the YOLOv3 algorithm so that it can better fit the leaf counting task. In order to evaluate our proposed method, we have collected cauliflower images from the ASBCST (Academia Sinica - Biotechnology Center in Southern Taiwan) and the Arabidopsis images from the CVPPP (Computer Vision Problems in Plant Phenotyping) dataset. Our proposed method achieves promising results on these datasets.