Modelagem granular convolucional evolutiva para classificação de fluxo de imagens

Abstract

Recent advances in machine learning for computer vision and image classification emphasize two main aspects: (i) the explainability or interpretability of deep neural models for classification; and (ii) the ability for continuous online learning of the model after its deployment in a dynamic environment, as observed in a stream of images. In this work, we present a framework of Convolutional Evolving Granular Neural Network aimed at advancing the understanding and application of machine learning in computer vision, specifically in image recognition and classification. The network is equipped with an incremental algorithm, which addresses both issues (i) and (ii), providing a higher level of interpretability to the neural model and enabling lifelong continuous learning. The proposed modeling, named Convolutional Evolving Granular Neural Network (CEGNN), combines part of a Convolutional Neural Network (CNN) called VGG-16 with an evolving granular network (EGNN). The connectionist structure and the information granule parameters of EGNN are gradually developed and updated based on the analysis of principal components (PCAs) of latent variables that may represent features that are not directly observable, such as edges, textures, shapes, or objects, extracted from the stream of images. In particular, the VGG-16 CNN is exploited to generate a compact feature space, which refers to a representation of data features in a lower-dimensional space that preserves relevant information for a specific task, such as image classification, while the EGNN, composed of trapezoidal fuzzy granules and T-norm and S-norm aggregation functions, is used to capture patterns and classify images. The Principal Component Analysis (PCA) technique is implemented at the integration point between VGG-EGNN, aiming to represent the abstract features that influence the observed data, reducing data processing and online training time. This approach not only allows for efficient handling of images or video frames at relatively higher frequencies but also highlights that the accuracy and interpretability of the global model are enhanced by the reconfiguration of connections resulting from PCA transformation in the latent space. This is possible because by reducing the dimensionality of the data, information loss is minimized. The results obtained indicate that the CEGNN model is efficient and interpretable in the task of classifying images into ten distinct classes, achieving an accuracy of 78.88% and a precision of 0,79 in image classification. These results highlight the effectiveness of the proposed approach in dealing with the complexity of classification tasks, emphasizing its viability and relevance in various practical applications, such as analysis of brain images, radiological images, satellite images, mobile robots, and autonomous vehicles, among others.