Plant Species Identification from Occluded Leaf Images.

Area of Science:

• Computer Vision
• Botany
• Pattern Recognition

Background:

• Contour-based 2D shape matching is well-established, but matching occluded objects, particularly plant leaves, to complete databases remains a significant challenge.
• Classifying occluded plant leaves is complex due to structural variations and the inherent difficulty of matching incomplete data against complete references.

Purpose of the Study:

• To develop an effective and efficient algorithm for identifying plant species from occluded leaf images using contour information.
• To address the challenge of matching incomplete leaf contours against a database of complete leaf shapes.

Main Methods:

• Representing 2D contour points using β-Spline curves and extracting interest points via the Discrete Contour Evolution (DCE) algorithm.
• Employing subgraph matching with DCE points and computing similarity transformation parameters to overlay occluded curves with database entries.
• Utilizing a refined energy functional incorporating curvature, Shape Context, and String Cut features, minimized via convex-concave relaxation for accurate matching.

Main Results:

• The proposed algorithm demonstrates effectiveness and efficiency in identifying plant species from occluded leaf images across multiple public databases.
• The method outperforms existing state-of-the-art techniques in matching leaf contours, even when up to 50% of the contour is occluded.
• Experimental results validate the algorithm's capability to accurately identify the best full leaf match despite significant contour occlusion.

Conclusions:

• The developed approach provides a robust solution for plant species identification from occluded leaf images, overcoming limitations of previous methods.
• The combination of contour representation, feature extraction, and energy minimization offers a powerful tool for botanical image analysis.
• This research contributes a significant advancement in handling partial shape matching problems within the domain of plant identification.