Plant leaf detection using modified active shape models

• Published in: Biosystems engineering 2013-09, Vol.116 (1), p.23-35
• Main Authors: Xia, Chunlei, Lee, Jang-Myung, Li, Yan, Song, Yoo-Han, Chung, Bu-Keun, Chon, Tae-Soo
• Format: Article
• Language: English
• Subjects: Agricultural machinery and engineering; Agronomy. Soil science and plant productions; Biological and medical sciences; Fundamental and applied biological sciences. Psychology; Generalities. Biometrics, experimentation. Remote sensing; greenhouses; leaves; paprika; pepper; rolling
• ISSN: 1537-5110; 1537-5129
• DOI: 10.1016/j.biosystemseng.2013.06.003

We propose an in situ detection method of multiple leaves with overlapping and occlusion in greenhouse conditions. Initially a multilayer perceptron (MLP) is used to classify partial boundary images of pepper leaves. After the partial leaf boundary detection, active shape models (ASMs) are subsequently built to employ the images of entire leaves based on a priori knowledge using landmark. Two deformable models were developed with pepper leaves: Boundary-ASM and MLP-ASM. Matching processes are carried out by deforming the trained leaf models to fit real leaf images collected in the greenhouse. MLP-ASM detected 76.7 and 87.8% of overlapping and occluded pepper leaves respectively, while Boundary-ASM showed detection rates of 63.4 and 76.7%. The detection rates by the conventional ASM were 23.3 and 29.3%. The leaf models trained with pepper leaves were further tested with leaves of paprika, in the same family but with more complex shapes (e.g., holes and rolling). Although the overall detection rates were somewhat lower than those for pepper, the rates for the occluded and overlapping leaves of paprika were still higher with MLP-ASM (ranging from 60.4 to 76.7%) and Boundary-ASM (ranging from 50.5 to 63.3%) than using the conventional active shape model (from 21.6 to 30.0%). The modified active shape models with the boundary classifier could be an efficient means for detecting multiple leaves in field conditions. •Detection of multiple plant leaves was studied in complicated natural scenes.•Overlapping and occluded plant leaves were successfully identified.•Active shape model was enhanced by employing the proposed leaf boundary classifier.•Detection rates were significantly improved by the modified active shape model.•Paprika leaves were additionally detected by the leaf model built from pepper leaves.