OFMFS Few-shot learning (FSL) aims to enable models to classify unseen classes in data scarce situations using limited labeled samples. Recently, feature optimization has enhanced model classification by optimizing feature extraction and representation. However, these approaches ignore the complementary guidance of frequency information and are limited by the inherent structure within the samples. In this paper, we propose an optimizing feature diversity via multi perspective self-supervised task (OFMFS), which learns multi-granularity feature representations through two complementary stages. OFMFS comprises feature multi perspective training (FMT) and feature diversity learning (FDL). The FMT strategy enforces the model to predict image rotation angles and corresponding wavelet subbands, enabling multi angle and multi-scale feature exploration in the limited data. The FDL module enhances sensitivity to fine-grained features by randomly swapping patches between samples to break hidden feature dependencies. OFMFS demonstrates its effectiveness through testing on three widely used FSL benchmark datasets.