Matheus Vinícius Todescato

Erick da Silva Puls, Matheus V. Todescato, Joel L. Carbonera

In recent years, we have witnessed a considerable increase in performance in image classification tasks. This performance improvement is mainly due to the adoption of deep learning techniques. Generally, deep learning techniques demand a large set of annotated data, making it a challenge when applying it to small datasets. In this scenario, transfer learning strategies have become a promising alternative to overcome these issues. This work aims to compare the performance of different pre-trained neural networks for feature extraction in image classification tasks. We evaluated 16 different pre-trained models in four image datasets. Our results demonstrate that the best general performance along the datasets was achieved by CLIP-ViT-B and ViT-H-14, where the CLIP-ResNet50 model had similar performance but with less variability. Therefore, our study provides evidence supporting the choice of models for feature extraction in image classification tasks.

Matheus Vinícius Todescato, Joel Luís Carbonera

Learning discriminative representations is a central goal of supervised deep learning. While cross-entropy (CE) remains the dominant objective for classification, it does not explicitly enforce desirable geometric properties in the embedding space, such as intra-class compactness and inter-class separation. Existing metric learning approaches, including supervised contrastive learning (SupCon) and proxy-based methods, address this limitation by operating on pairwise or proxy-based relationships, but often increase computational cost and complexity. In this work, we introduce Soft Silhouette Loss, a novel differentiable objective inspired by the classical silhouette coefficient from clustering analysis. Unlike pairwise objectives, our formulation evaluates each sample against all classes in the batch, providing a batch-level notion of global structure. The proposed loss directly encourages samples to be closer to their own class than to competing classes, while remaining lightweight. Soft Silhouette Loss can be seamlessly combined with cross-entropy, and is also complementary to supervised contrastive learning. We propose a hybrid objective that integrates them, jointly optimizing local pairwise consistency and global cluster structure. Extensive experiments on seven diverse datasets demonstrate that: (i) augmenting CE with Soft Silhouette Loss consistently improves over CE and other metric learning baselines; (ii) the hybrid formulation outperforms SupCon alone; and (iii) the combined method achieves the best performance, improving average top-1 accuracy from 36.71% (CE) and 37.85% (SupCon2) to 39.08%, while incurring substantially lower computational overhead. These results suggest that classical clustering principles can be reinterpreted as differentiable objectives for deep learning, enabling efficient optimization of both local and global structure in representation spaces.

Matheus Vinícius Todescato, Joel Luís Carbonera

While deep learning, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), has significantly advanced classification performance, its typical reliance on extensive annotated datasets presents a major obstacle in many practical scenarios where such data is scarce. Vision-language models (VLMs) and transfer learning with pre-trained visual models appear as promising techniques to deal with this problem. This paper proposes a novel zero-shot image classification framework that combines a VLM and a pre-trained visual model within a self-learning cycle. Requiring only the set of class names and no labeled training data, our method utilizes a confidence-based pseudo-labeling strategy to train a lightweight classifier directly on the test data, enabling dynamic adaptation. The VLM identifies high-confidence samples, and the pre-trained visual model enhances their visual representations. These enhanced features then iteratively train the classifier, allowing the system to capture complementary semantic and visual cues without supervision. Notably, our approach avoids VLM fine-tuning and the use of large language models, relying on the visual-only model to reduce the dependence on semantic representation. Experimental evaluations on ten diverse datasets demonstrate that our approach outperforms the baseline zero-shot method.