MKCNet: Learning Multi-Granularity Key Cues for FGVC for Intelligent Monitoring Robot via Transformers

FGVC plays a major role in the field of robotic vision, which has a significant impact, for example, on the conservation of endangered and rare birds, It aims to distinguish similar subcategories. In this paper, we propose a novel Multi-Granularity Key Cue Learning Network (MKCNet) for fine-grained visual classification. First, we design a multi-granularity feature extraction module to capture local information at different scales and enhance high-level feature representations with discriminative local features across different granularities. Next, to extract partial features of different shapes at each granularity, we introduce a partial sampling attention mechanism that comprehensively samples implicit semantic parts on the feature map. This partial sampling not only takes into account the importance of sample localization, but also applies local loss to mitigate the overfitting problem. Through extensive experimental validation on the CUB-200-2011 and NABirds datasets, our approach demonstrates superior performance compared to current top techniques.