Intelligent identification analysis and process design for highly similar categories using Platycerium as an example

Area of Science:

• Computer Vision
• Machine Learning
• Botany

Background:

• Image recognition struggles with datasets featuring high inter-class similarity.
• Platycerium species present a significant challenge due to their visual likeness.

Purpose of the Study:

• To develop a robust framework for optimizing image recognition accuracy in challenging datasets.
• To improve the recognition of highly similar species using advanced analytical methods.

Main Methods:

• Utilized multidimensional confusion matrices to identify key confusion factors (e.g., edges, textures, shapes).
• Stratified datasets into processed and unprocessed images, optimizing for identified factors via saturation, brightness, and sharpening adjustments.
• Employed confusion matrices and bootstrapping for refinement of ambiguous classes.

Main Results:

• Baseline accuracy with ResNet50 was below 10% for Platycerium species.
• Confusion factor analysis and image optimization increased recognition accuracy to approximately 60%.
• Further improvements to over 80% using EfficientNet-b4 and over 90% using EfficientNet-b7 were achieved.

Conclusions:

• Feature selection and grouped analysis are vital for recognizing visually similar images.
• The proposed framework significantly enhances recognition accuracy for challenging image datasets.
• Findings offer valuable insights for advancing image recognition technologies.