MRFB-Net for Identifying Butterfly Species Via Images Taken in the Field Environments

It is a very challenging task to identify butterfly species though images captured in field environments. First, the mimicry properties of butterflies make it difficult to detect the butterflies out of the background. Second, the folding or overlapping of their wings make it very demanding to learn features for identifying butterfly species. Third, identifying butterfly species is a fine-grained and multi-class classification problem. To address the aforementioned challenges, the plug and play module referred to MRFB (Multi-Receptive Field Block) is proposed. Subsequently, the MRFB-net is developed by integrating the MRFB module into the RFB net framework. This innovative network demonstrates the capability to simultaneously detect butterflies of varying sizes by leveraging distinct receptive field blocks within MRFB module to learn features of butterflies. As a result, these advancements effectively address the challenges encountered in identifying butterfly species from images captured in field environments. The extensive experimental results demonstrate the exceptional performance of the MRFB-net in identifying butterfly species. Comparatively, it outperforms the state-of-the-art (SOTA) algorithm in this specific task, establishing itself as the current leading model for butterfly species identification from field images.