Algorithm 1.

Unsupervised Few-shot Feature Learning via Self-supervised Training (UFLST)

Input: Unlabeled data set $\mathcal{X}=\left\{{\text{x}}_i\right\}$, the few-shot feature embedding $f_{{\theta }^0}$, the training iteration T.

Output: Trained few-shot embedding $f_{{\theta }^T}$

1: t = 0

2: repeat

3:     Clustering:

4:     Extracting features {zi} of {xi} using the feature extractor $f_{{\theta }^t}$.

5:     Calculating KRJD Jij based on the K-reciprocal nearest neighbors of any data pairs zi and zj.

6:     Clustering data using DBSCAN and generating pseudo labels {yi}.

7:     Removing outliers and obtaining the pseudo labeled data set $\left\{(\widetilde{{\text{x}}_i},\widetilde{y_i})\right\}$.

8:     Episodic Training:

9:     Constructing a set of episodic tasks $\left\{{\mathcal{T}}_s\right\}$; for each task, randomly sampling M classes with K+Q examples per class from $\left\{(\widetilde{{\text{x}}_i},\widetilde{y_i})\right\}$.

10:       Updating model parameters θt by training the few-shot learner on the series of episodic tasks $\left\{{\mathcal{T}}_s\right\}$.

11:       t = t + 1

12: until t = T