DualNet: A Modified Network for Fine-Grained Flower Recognition

To address the challenge of fine-grained flower image classification, this paper proposes a new network named DualNet, in which we integrats EfficientNet-B1 and DenseNet201. Our model has two distinctive characteristics:(1) Achieve dynamic multi-scale feature fusion via a channel attention mechanism.(2) A focal loss function with label smoothing is introduced to enhance sample discriminability. Experiments demonstrate that DualNet achieves 98.31% accuracy and 98.61% mean average precision. Ablation studies show that the combined improvements boost accuracy by 10.1% compared to single-branch EfficientNet-B1. Notably, this DualNet outperforms ResNet50, InceptionV3, MobileNetV2, and their respective EfficientNet-B1 hybrids by 14.69%, 15.4%, $18.15%, 10.52%, 11.73%$, and 14.12% in average accuracy. Visualization analysis reveals the model's ability to focus on critical floral features, validating its effectiveness in fine-grained recognition. This work provides a robust solution for botanical image classification tasks.