Askari Fatemeh ; Fateh Amirreza ; Mohammadi Mohammad Reza

Abstract

In the context of few-shot classification, the goal is to train a classifierusing a limited number of samples while maintaining satisfactory performance.However, traditional metric-based methods exhibit certain limitations inachieving this objective. These methods typically rely on a single distancevalue between the query feature and support feature, thereby overlooking thecontribution of shallow features. To overcome this challenge, we propose anovel approach in this paper. Our approach involves utilizing a multi-outputembedding network that maps samples into distinct feature spaces. The proposedmethod extracts feature vectors at different stages, enabling the model tocapture both global and abstract features. By utilizing these diverse featurespaces, our model enhances its performance. Moreover, employing aself-attention mechanism improves the refinement of features at each stage,leading to even more robust representations and improved overall performance.Furthermore, assigning learnable weights to each stage significantly improvedperformance and results. We conducted comprehensive evaluations on theMiniImageNet and FC100 datasets, specifically in the 5-way 1-shot and 5-way5-shot scenarios. Additionally, we performed cross-domain tasks across eightbenchmark datasets, achieving high accuracy in the testing domains. Theseevaluations demonstrate the efficacy of our proposed method in comparison tostate-of-the-art approaches. https://github.com/FatemehAskari/MSENet