Texture, a significant visual attribute in images, plays an important role in many pattern recognition tasks. While Convolutional Neural Networks (CNNs) have been among the most effective methods for texture analysis, alternative architectures such as Vision Transformers (ViTs) have recently demonstrated superior performance on a range of visual recognition problems. However, the suitability of ViTs for texture recognition remains underexplored. In this work, we investigate the capabilities and limitations of ViTs for texture recognition by analyzing 25 different ViT variants as feature extractors and comparing them to CNN-based and hand-engineered approaches. Our evaluation encompasses both accuracy and efficiency, aiming to assess the trade-offs involved in applying ViTs to texture analysis. Our results indicate that ViTs generally outperform CNN-based and hand-engineered models, particularly when using strong pre-training and in-the-wild texture datasets. Notably, BeiTv2-B/16 achieves the highest average accuracy (85.7%), followed by ViT-B/16-DINO (84.1%) and Swin-B (80.8%), outperforming the ResNet50 baseline (75.5%) and the hand-engineered baseline (73.4%). As a lightweight alternative, EfficientFormer-L3 attains a competitive average accuracy of 78.9%. In terms of efficiency, although ViT-B and BeiT(v2) have a higher number of GFLOPs and parameters, they achieve significantly faster feature extraction on GPUs compared to ResNet50. These findings highlight the potential of ViTs as a powerful tool for texture analysis while also pointing to areas for future exploration, such as efficiency improvements and domain-specific adaptations.
Effectively training deep learning models relies heavily on large datasets, as insufficient instances can hinder model generalization. A simple yet effective way to address this is by applying modern deep learning augmentation methods, as they synthesize new data matching the input distribution while preserving the semantic content. While these methods produce realistic samples, important issues persist concerning how well they generalize across different classification architectures and their overall impact in accuracy improvement. Furthermore, the relationship between dataset size and model accuracy, as well as the determination of an optimal augmentation level, remains an open question in the field. Aiming to address these challenges, in this paper, we investigate the effectiveness of eight data augmentation methods—StyleGAN3, DCGAN, SAGAN, RandAugment, Random Erasing, AutoAugment, TrivialAugment and AugMix—throughout several classification networks of varying depth: ResNet18, ConvNeXt-Nano, DenseNet121 and InceptionResNetV2. By comparing their performance on diverse datasets from leaf textures, medical imaging and remote sensing, we assess which methods offer superior accuracy and generalization capability in training models with no pre-trained weights. Our findings indicate that deep learning data augmentation is an effective tool for dealing with small datasets, achieving accuracy gains of up to 17%.