Aprendizado de métricas para filtragem não local e classificação de imagens tomográficas de sementes agrícolas

Abstract

Image denoising is a recurrent problem across a wide range of applications, from preprocessing in computer vision to computer-aided diagnosis. This work proposes and evaluates an enhanced Non-Local Means (NLM)–based filtering approach, termed Dual Non-Local Means (Dual NLM), which incorporates information-theoretic divergence measures to quantify similarities between image regions. We investigated the Kullback–Leibler, Bhattacharyya, Hellinger, and Cauchy–Schwarz divergences with the objective of preserving edges and fine details while more effectively reducing Gaussian noise. Experiments were conducted on 18 tomographic images from different agricultural crops (sunflower, chickpea, wheat, seed mix, jatropha, and soybean), corrupted with additive Gaussian noise of variance 𝜎𝑛 = 10. Quantitative evaluation using the Peak Signal-to-Noise Ratio (PSNR) demonstrated that the Dual NLM variants based on Cauchy–Schwarz and Kullback–Leibler divergences outperform, on average, the standard NLM and classical filters (Wiener, Anisotropic Diffusion, and Total Variation). In a subsequent stage, we investigated the usefulness of the restored images for pattern analysis. Local descriptors (patches) were extracted, and two dimensionality-reduction strategies were compared: SEI-ISOMAP (with stochastic diffusion distances) and PCA. The resulting embeddings were evaluated using trustworthiness, Pearson correlation between pairwise distances, and silhouette index, and subsequently used as input to supervised classifiers (KNN, SVM with RBF kernel, and Random Forest), whose performance was assessed in terms of accuracy and macro-𝐹1 under stratified cross-validation. The results indicate that SEI-ISOMAP provides better structural preservation and class separability than PCA, and that the combination Dual NLM + SEI-ISOMAP + Random Forest achieves the best average classification performance. Overall, the findings highlight the potential of combining information-theoretic similarity measures with diffusion-based dimensionality reduction for restoration and discriminative analysis of tomographic images corrupted by Gaussian noise.