Foreground-background Separation metric for Few-shot Fine-grained Classification
Jintao Li, Bing Wei, Xin Hu, Kuangrong Hao, Xiaodan Hong, Angang Chen · 2023
Few-shot fine-grained (FSFG) image classification is focused on the task of classifying fine-grained images with limited labeled samples. The primary challenge lies in enhancing the extraction of more discriminative features from the images. Moreover, the disturbance from the background can negatively impact the final classification, often overlooked by many researchers. Facing this circumstances, we try a new way that employs a foreground-background separation metric. This method consists of three key components: a deep embedding module, an attention generation module, a foreground-background separation metric module. Specifically, the deep embedding module extracts the features of input images. In order to emphasize the more discriminative features, we introduce an attention generation module. This module adjusts the feature map of the query sample based on the support samples. The foreground-background separation module will divide the feature map of the query sample into the foreground feature descriptors and background feature descriptors, and then measure the cosine distance of these two parts with support feature descriptors respectively. We extensively evaluate our method on three benchmark datasets, and the results demonstrate its competitive performance compared to state-of-the-art methods.