Intelligent Identification Analysis and Process Design for Highly Similar Categories- Using Platycerium as an Example

This study addresses the challenges of image recognition in datasets with high inter-class similarity, using 18 native species of Platycerium as a case study. Due to the substantial visual similarities within the dataset, initial training using the ResNet50 model yielded a baseline accuracy of less than 10%, highlighting the complexity of the task. To address this, a comprehensive analysis was conducted using multidimensional confusion matrices, identifying seven primary confusion factors, including image edges, textures, and shapes. The dataset was then stratified into processed images, optimized for confusion factors such as saturation, brightness, and sharpening, and unprocessed original images. Besides, a refinement process is also proposed to discover ambiguous classes and further processes are also performed, which resulted in significant improvements in recognition accuracy, demonstrating the efficacy of confusion factor analysis in enhancing model performance. The findings underscore the importance of feature selection and grouped analysis in recognizing highly similar images. By systematically addressing confusion factors, this study offers a robust framework for optimizing recognition accuracy in challenging datasets, providing valuable insights and strategies for advancing image recognition technologies.