Unsupervised Learning Method for Plant and Leaf Segmentation

Plant phenotyping is a recent application of computer vision in agriculture and food security. To automatically recognize plants species, we need first to extract the plant and associated substructures. Manual segmentation of plant structures is tedious, error prone and expensive. Automatic plant se...

Full description

Saved in:
Bibliographic Details
Main Authors: Al-Shakarji, Noor M., Kassim, Yasmin M., Palaniappan, Kannappan
Format: Conference Proceeding
Language:English
Subjects:
Computer vision
expectation maximization
Image color analysis
Image segmentation
k-mean
leaf segmentation
Mathematical model
Measurement
Plant segmentation
Plants (biology)
Transforms
unsupervised learning
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Plant phenotyping is a recent application of computer vision in agriculture and food security. To automatically recognize plants species, we need first to extract the plant and associated substructures. Manual segmentation of plant structures is tedious, error prone and expensive. Automatic plant segmentation is useful for leaf extraction, identification, and counting. We have developed a robust and fast unsupervised approach for plant extraction and leaf detection. K-means based mask (of the pot) followed by Expectation Maximization (EM) algorithm is adapted to estimate a mixture model for identifying the foreground area for the plant. We utilized the EM with 3 RGB channels to identify the foreground verses background for plant localization. K-means has been used to extract the circular plant can as one of the intermediate result to fuse it with EM results for noise removal since the images suffered from contrast and illumination variations. For leaf segmentation, we utilized distance transform and watershed segmentation to localize the leaves individually followed by stem link algorithm to connect the stem with corresponding leaves. The results have been evaluated by the same algorithms that have been used in the contest of plant phenotyping [1]. In our work, we used A1 and A2 datasets 1 to test our algorithm. We achieved promissing score in some evaluation metrics and comparable in the others.
ISSN:2332-5615
DOI:10.1109/AIPR.2017.8457935