Enhanced Feature Representation for Low-resolution Fine-grained Image Recognition via Categorical Knowledge Guidance

Low-resolution (LR) fine-grained image recognition requires recognizing the subcategories of LR samples with limited fine-grained details. Existing methods do not make full use of the guiding and constraining role of category-related knowledge in recovering and extracting fine-grained features of LR data, thus limiting the capability of these methods to learn global and local fine-grained features of LR data. We propose an enhanced feature representation network (EFR-Net) based on categorical knowledge guidance to capture delicate and reliable fine-grained feature descriptions of LR data for improving the recognition accuracy. First, to overcome the challenges of limited fine-grained details in LR data, we design a class-wise distillation loss. It transfers the high-quality features of class-specific high-resolution (HR) samples into the feature learning of the same-category LR samples by using a memory bank. In this way, the global representation of LR images is close to the meaningful and high-quality features. Second, considering that fine-grained discriminative features are often hidden in object parts, we preset a group of part queries to learn the positional information where the discriminative cues exist across all categories, and then use it to decode diverse and discriminative part features. The global representation combines with local discriminative features to form more comprehensive and meaningful feature descriptions of LR fine-grained objects, thus improving the recognition performance. Extensive comparison experiments on four LR datasets demonstrate the effectiveness of EFR-Net.