Significant differences in spatial structure characteristics make the coarse-grained few-shot scenarios easier to handle. However, fine-grained scenes remain often more challenging, and current related research falls far short of human-level recognition capabilities. To address this tough challenge, we get inspiration by the structures and functions of the human visual system, and then propose a human-centric visual cognition recognition framework, named VCRNet, including a shallow cognition network VCRNet-4 and the deep cognition neural network VCRNet-12. Specifically, in this framework, we cleverly design a visual perception module to simulate powerful visual pathway feature encoding of human visual system, an attentional regional sensing module and another pixel-level sensing module to simulate the structural and shape-detailed perception capabilities of the human occipital lobe, and a cognitive recognition module to emulate the integrative cognitive abilities of human frontal lobe. Ulteriorly, we have evaluated the performances of our proposed methods on three publicly available fine-grained benchmarks: CUB-200-2011, Aircraft-Fewshot, and Stanford-Cars. The compelling comparative experiments and thoroughly justified validation study demonstrate the structural excellence and superior performance of our VCRNet framework. To facilitate subsequent research, our codes have been available.

Meta-metric learning has demonstrated strong performance in coarse-grained few-shot situations. However, despite their simplicity and availability, these metametrics are limited in effectively handling fine-grained few-shot scenarios. Fine-Grained Few-Shot Classification (FGFSC) presents significant challenges to the network's ability to extract subtle features. Equipped with the symmetrical binocular perception system and complex neural networks in the brain, humans inherently possess exceptional and resilient meta-learning abilities, facilitating superior management of fine-grained few-shot scenarios. In this paper, inspired by the human binocular visual system, we pioneer the first human-like meta-metric paradigm: Binocular Singular Hellinger Metametric (BinoHeM). Functionally, BinoHeM incorporates advanced symmetric binocular feature encoding and recognition mechanisms. Structurally, it integrates two binocular sensing feature encoders, a singular Hellinger metametric, and two collaborative identification mechanisms. Building on this foundation, we introduce two innovative metametric variants: BinoHeM-KDL and BinoHeM-MTL. These are grounded in two advanced training mechanisms: knowledge distillation learning (KDL) and meta-transfer learning (MTL), respectively. Furthermore, we showcase the high accuracy and robust generalization capabilities of our approaches on four representative FGFSC benchmarks. Extensive comparative and ablation experiments have validated the efficiency and superiority of our paradigm over other state-of-the-art algorithms. Our code is publicly available at: https://github.com/ChaofeiQI/BinoHeM.