Hybrid Fine-Tuning Strategy for Few-Shot Classification

Figure 2

Diagram of the proposed FSLDA module. Given a C-way K-shot support set or target dataset, we first get the feature vector for each sample, the prototype for each class, and the covariance matrix for all feature vectors sequentially. Then, the rank factor is introduced to obtain the precision matrix for FSC tasks based on the weighted mean of the covariance matrix and the identity matrix. Finally, we obtain the parameter value of the last fully connected layer by and initialize it.