Few-shot fine-grained image classification remains challenging due to limited supervision and the need to distinguish subtle differences among visually similar categories. We propose MP-DistillFormer, a three-stage framework that enhances discrimination and generalization through multimodal prototype distillation and self-supervised refinement. In the first stage, a CNN-based teacher is trained with stochastic augmentation to capture diverse local features. In the second stage, knowledge is distilled into our proposed TeSMo-KAN student model. TeSMo-KAN unifies convolutional tokenization for spatial precision, lightweight local refinement for detail preservation, and nonlinear decision modeling within a Transformer backbone. To enrich semantic representation, TeSMo-KAN is guided by multimodal prototypes that fuse complementary visual and textual embeddings, enabling more discriminative learning under few-shot settings. Finally, a rotation-based self-supervised fine-tuning stage improves robustness under data-scarce conditions. Extensive experiments on three fine-grained benchmarks including CUB-200-2011, Stanford Dogs, and Stanford Cars demonstrate that MP-DistillFormer consistently outperforms state-of-the-art methods in both 1-shot and 5-shot scenarios. The source code is available at https://github.com/annym-ai00/MP-DistillFormer .

Few-shot fine-grained image classification remains a challenging task due to the subtle inter-class variations and the scarcity of labeled data. Existing few-shot fine-grained methods often struggle to generalize effectively under low-data regimes and fail to discriminate between visually similar samples. To address these challenges, we propose DistillCvT, a two-stage framework to enhance model generalization and feature discrimination through self-supervised learning and self-distillation within prototypical network structure. Convolutional Vision Transformer (CvT) is employed as the feature extractor, which integrates convolutional locality with transformer-based global modeling to capture fine-grained details while preserving global context. DistillCvT mitigates these issues through a two-stage strategy. In the first stage, the model is trained using a prototypical distance-based classification loss, where class probabilities are computed from the distances between query samples and class prototypes, complemented by standard data augmentation to mitigate data scarcity and learn discriminative class representations. In the second stage, an identical student model is trained with a self-supervised jigsaw task to enhance feature discrimination and robustness by aligning support samples with their augmented query counterparts. Simultaneously, the teacher model from stage one guides it through logit-based self-distillation to further improve generalization. Experiments on CUB-200-2011, Stanford Dogs, and Stanford Cars demonstrate that DistillCvT consistently improves performance and outperforms recent state-of-the-art few-shot fine-grained image classification methods.

Over the past decade, deep learning has made significant advancements in image classification. However, these models struggle with data scarcity and distribution shifts, commonly referred to as the few-shot image classification (FSIC) problem. FSIC aims to recognize novel classes using only a limited number of labeled samples, posing challenges for conventional deep learning models that rely on large datasets for optimal performance. This paper provides a comprehensive review of FSIC methodologies, categorizing them into five main approaches: meta-learning, transfer learning, data augmentation, attribute-related, and vision-language foundation model adaptation. Meta-learning approaches are further classified into metric-based, model-based, and optimization-based methods, while transfer learning approaches are divided into hybrid and non-hybrid methods. Vision-language foundation model adaptation approaches are grouped into few-shot parameter tuning, dynamic or unsupervised tuning, and training-free adaptation methods. Beyond general FSIC, this paper also explores specialized FSIC methods in fine-grained classification, cross-domain classification, and class-incremental learning. Additionally, it reviews commonly used few-shot image datasets and compares the performance of representative methods through experimental results. Practical applications of FSIC across various domains are also discussed, highlighting its potential to address real-world challenges. Finally, the research trends of FSIC are identified, offering insights into the state-of-the-art FSIC methods and guiding future advancements in this field.

Typically, hyperspectral image suffers from redundant information, data scarcity, and class imbalance problems. This letter proposes a hyperspectral image classification framework named a two-stream residual separable convolution (2SRS) network that aims to mitigate these problems. Principal component analysis (PCA) is first employed to reduce the spectral dimension of the hyperspectral image. Subsequently, the data scarcity and class imbalance problems are overcome via spatial and spectral data augmentations. A novel spectral data creation from image patches is proposed. The augmented samples are fed into the proposed 2SRS network for hyperspectral image classification. We evaluated the proposed method on three benchmark datasets, namely, 1) Indian Pines (IP); 2) Pavia University; and 3) Salinas Scene (SA). The proposed method achieved state-of-the-art performance in terms of overall accuracy (OA), average accuracy (AA), and Kappa coefficient (Kappa) for both 30% and 10% training set ratios.