Flower recognition system based on YOLOv8 algorithm

Abstract

With the rapid expansion of the demand for automatic identification of floral species, although traditional flower recognition and classification models have made considerable progress, they still struggle to fully capture the fine-grained morphological differences. Based on this, this study focuses on a core issue: Can the new attention mechanism and Vision Mamba architecture significantly improve the performance of the traditional YOLOv8 algorithm on the floral recognition dataset? This study is oriented towards the fine-grained visual classification scenario and designs and implements an efficient flower recognition system. The Oxford Flowers-102 dataset is used as the benchmark. Subsequently, under the YOLOv8 framework, the lightweight YOLOv8n-cls, performance-balanced YOLOv8m-cls, and two improved models that integrate SE attention mechanism and Vision Mamba state space module are trained respectively. Experimental results show that YOLOv8n-cls achieves an accuracy of 98.20\% with only 2.7M parameters; YOLOv8m-cls boosts the accuracy to 99.18\% and the inference time to only 3.7ms; after introducing SE attention, the accuracy further increases to 99.45\%; when Vision Mamba is combined with YOLOv8m-cls, the accuracy climbs to 99.77\% and the Top-5 accuracy to 99.97\%, fully verifying the significant benefit of global dependency modeling for flower detail discrimination. According to the results, the reseaech question is validated. To facilitate the implementation of the algorithm, we have also developed a Gradio Web interface that supports switching between Chinese and English, enabling rapid loading of model weights and real-time inference. This provides an easy-to-use and highly accurate tool for education, horticulture, and biodiversity conservation.