A ResNet34-Based Dual-Attention Network for Tomato Leaf Disease Recognition

Abstract

Tomato leaf diseases are important factors affecting tomato yield and quality. Accurate and efficient disease recognition is therefore essential for intelligent crop protection and agricultural monitoring. Although convolutional neural networks have achieved promising performance in plant disease recognition, complex field environments still present challenges such as background interference, illumination variation, and subtle lesion differences. To improve the discriminative capability of tomato leaf disease recognition, a ResNet34-based dual-attention network is proposed in this study. ResNet34 is adopted as the backbone network for hierarchical feature extraction, and the original classification structure is replaced by an attention-enhanced classification framework. A dual channel-spatial fusion attention module is introduced to enhance lesion-related feature responses by jointly modeling channel dependency and spatial saliency. In addition, depthwise convolutional operations are incorporated to improve feature adjustment and computational feasibility. Experiments are conducted on a seven-class tomato leaf dataset containing six disease categories and healthy leaves. The proposed method achieves an accuracy of 98.12%, precision of 98.82%, recall of 98.10%, and F1-score of 98.22%, outperforming ResNet34, ResNet50, DenseNet121, and ShuffleNetV2 under the same experimental conditions. The ablation results and Grad-CAM visualization further demonstrate that the proposed attention mechanism can enhance lesion localization and suppress irrelevant background responses. These results indicate that the proposed model provides an effective method for tomato leaf disease recognition in intelligent agricultural monitoring scenarios.