Three-Dimensional Plant Model Development Through Image Recognition

This paper proposes a novel method using image-recognition techniques to develop three-dimensional (3D) models of basil plants. Traditional approaches have dificulty scanning outdoor plants and stems overlapping outer leaves. In this paper, we collect 3D plant models in advance that reproduce the external and internal structures. Then, by selecting from the database the 3D model most similar to the actual plant in appearance, the proposed method develops 3D plant models using only images. However, collecting precise 3D models is a cost-intensive task. Based on the growth pattern that basil plants exhibit alternating leaves during growth, the proposed method automatically mass-produces realistic 3D plant models by assembling 3D leaf and stem models of actual plants. Additionally, we employ an image-recognition technique to extract embedding vectors from multi-angle images and assess the visual similarity between the actual plant and the realistic 3D plant model based on their cosine similarity. Finally, we construct a vector-search system incorporating k-means clustering and dimensionality reduction to limit the search scope and minimize computational complexity. Experimental results show that the proposed method efficiently obtains the most similar 3D model in the database, achieving a mean reciprocal rank of 0.90 and a search time of 0.003 s per query.