AGHV-Net: An Attention-Guided Hybrid Vision Network for Robust Flower Classification

Abstract

Classification of fine-grained flower species is still a challenging task in computer vision because of high inter-class similarity, intra-class variability, illumination variations, and complex background conditions but Conventional CNN-based methods are good at capturing local spatial features but are generally inadequate in capturing long-range contextual dependencies, while ViT-based architectures learn well the global context but lack efficient local discriminative representation. In this paper, we propose Attention-Guided Hybrid CNN-Transformer Network (AGHV-Net) for accurate flower species classification. The proposed framework combines CNN-based local feature extraction with Transformer-based global contextual learning using an adaptive attention-guided feature fusion mechanism. First, we preprocess and augment the input flower images to improve the uniformity and generalization of the features. The CNN branch captures the discriminative local feature representations such as petal textures, edge structures, and color gradients. The Vision Transformer branch captures the long-range semantic dependencies and self-attention mechanisms. Then, an adaptive attention-guided feature fusion module is dynamically used to learn the local feature representations of CNN and Transformer to perform the local-global feature integration. Finally, we process this fused feature representation through fully connected and Softmax layers to perform multi-class classification. They conducted comprehensive experiments on the Oxford 102 Flower Dataset with accuracy, precision, recall, and F1-score as evaluation metrics. Experimental results show that the proposed AGHV-Net classification accuracy reaching 97.84% outperforms some existing CNN, Transformer, and hybrid deep learning architectures across many other methods. The comparison and ablation analysis further provides strong support for the proposed attention-based fusion mechanism that can dramatically improve discriminative feature identification, alleviate the inter-class confusion to enhance overall classification performance. Our new framework offers a practical and reliable route for fine-grained flower identification and can be extended into other complex picture classification applications.