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

• Published in: Scientific reports Vol. 15; no. 1; pp. 30517 - 12
• Main Authors: Chen, Li-Wei, Lin, Wei-Lun
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 20.08.2025; Nature Publishing Group; Nature Portfolio
• Subjects: 639/166; 639/166/987; Accuracy; Artificial intelligence; Automation; Confusion category extraction refinement; Datasets; Deep learning; Deep neural networks; Efficiency; Factor analysis; Flowers & plants; Humanities and Social Sciences; Identification; Image processing; Image recognition; Indigenous species; multidisciplinary; Neural networks; Optimization techniques; Science; Science (multidisciplinary); Image recognition; Confusion category extraction refinement; Deep neural networks
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-12502-9

This study tackles the challenge of image recognition for datasets with high inter-class similarity, using 18 native Platycerium species as a case study. Due to their substantial visual similarities, initial training with ResNet50 yielded a baseline accuracy of less than 10%. To address this, we conducted a comprehensive analysis using multidimensional confusion matrices to identify seven primary confusion factors, such as image edges, textures, and shapes, and stratified the dataset into processed and unprocessed images optimized for these factors through adjustments in saturation, brightness, and sharpening. A refinement process leveraging confusion matrices and bootstrapping was proposed to address ambiguous classes, significantly improving recognition of highly similar species. Recognition accuracy increased to approximately 60% after applying confusion factor analysis and image optimization, with further gains to over 80% using EfficientNet-b4 and over 90% using EfficientNet-b7. These findings highlight the importance of feature selection and grouped analysis in recognizing highly similar images, offering a robust framework for optimizing recognition accuracy in challenging datasets and providing valuable insights for advancing image recognition technologies.