Entropia diferencial multiescala e agregação de descritores na classificação de folhas de plantas

Abstract

Shape analysis and recognition are fundamental in the design of computer vision-based systems. The greatest challenge is to develop robust methods capable of extracting significant features from shapes in order to represent them effectively. In this context, multiscale descriptors represent a versatile and efficient alternative for shape characterization. This work presents a methodology for classifying plant leaves based on handcrafted features derived from the multiscale entropy of curvature and texture, as well as deep features obtained from convolutional neural networks (CNNs). We propose three object descriptors based on the multiscale entropy of curvature. These descriptors rely on the differential entropy of the probability distributions of multiscale curvatures to create a coarse-to-fine representation of the shape contour. Furthermore, we present a descriptor that aggregates the multiscale entropy of curvature, bending energy of curvature, and texture features to enhance the extraction of object signatures and subtle texture details in leaf images. The texture descriptor combines statistics from the local binary pattern and gray-level co-occurrence matrix. We compare our handcrafted descriptors with deep features extracted from various CNNs in a multiclass classification framework using the random forest classifier, replacing the fully connected layer of the CNNs with this classifier. Experiments were conducted on four public leaf datasets: Plantscan, MED117, Flavia, and Swedish. The results, evaluated using F1-score and accuracy metrics exceeding 99.50%, validate the aggregation strategy and demonstrate that it is competitive and robust. The findings also confirm that the proposed strategy outperformed four different sets of deep features according to both F1-score and accuracy metrics. Qualitative analysis through multidimensional data visualization further demonstrates that combining different shape features and texture details improves the description of leaf images, providing better intraclass compactness and interclass separation across the datasets.