Automatically learns task-specific features for high-performance image classification
Mingcheng Chen, Xin Lin, Mina Zhang · 2025
Image classification is a core task in computer vision, but it becomes particularly difficult when faced with small sample data sets. The limited amount of data restricts the model's ability to capture and learn important features. To address the problem, the research proposes a new method called Set-CL, which introduces a set-based feature extraction method to replace the traditional vector-based approach. Instead of representing each image as a single feature vector, Set-CL decomposes the image into a set of feature vectors, each capturing unique aspects of the image. This richer and more discriminative representation improves the ability of the model to produce new classes, particularly in learning scenarios of few-shots. Besides set-based feature extraction, Set-CL employs a matching index of set-to-set for classification, where the similarity in feature sets of an image which is a query and the support images is measured to enhance classification accuracy. Furthermore, Set-CL integrates contrastive learning, encouraging the model to differentiate between similar and dissimilar instances, thereby improving its ability to learn robust feature representations from limited data. Extensive experiments on the CUB (Caltech-UCSD Birds-200-2011) dataset, a benchmark for fine-grained image classification, demonstrated that Set-CL significantly outperforms seven well-known algorithms, including ProtoNet, MAML, and RelationNet, 1-shot, and 5-shot classification tasks have higher accuracy.