Avaliando diferentes técnicas de inteligência artificial explicável em classificação automática de sementes de soja

Abstract

This study addresses the challenge of automatic classification of defects in soybean seeds, a task in which Artificial Intelligence (AI) models are employed. In this context, we aim to investigate the application of Explainable Artificial Intelligence (XAI) techniques, focusing on LIME and Grad-CAM, to conduct a qualitative analysis of the decision-making process of Convolutional Neural Networks (CNNs) in the classification of defects in soybean seeds. The ResNet50 and SqueezeNet architectures were evaluated, with the goal of assessing quantitative performance metrics and understanding the basis of their predictions. The results indicate that although both architectures presented similar accuracy, the XAI-based analysis revealed significant distinctions. The ResNet50 architecture demonstrated a superior ability to focus on relevant and specific semantic features of the defects, generating explanations that were visually more coherent and aligned with human interpretation. In contrast, SqueezeNet exhibited a more dispersed attentional behavior, with activations on generic textures or irrelevant regions, suggesting a reliance on spurious correlations. It is concluded that explainability is an indispensable tool for validating AI models, providing critical insights into the robustness and reliability of the system that are not captured by traditional performance metrics. The integration of qualitative analysis is, therefore, essential for the development of safer and more transparent artificial intelligence systems.