Yue Zhang, Ji-Won Kim, L. Minh Dang, Wenqi Zhang, Yanan Wang, Xingshi Gan, Hyeonjoon Moon
언어
영어(ENG)
URL
https://www.earticle.net/Article/A468921
원문정보
초록
영어
This study presents a lesion-guided four-channel ConvNeXt model for tomato plant disease recognition. By segmenting lesion areas in the HSV color space, grayscale masks are generated and combined with RGB images to form a four-channel input. The proposed ConvNeXt4Channel network, optimized for this input, enhances spatial feature extraction. Experiments on the PlantVillage tomato dataset (train-test split: 8:2) show that the model, trained with cross-entropy loss and Adam optimizer (learning rate = 1e-4), achieves 96.81% accuracy—surpassing conventional models by approximately 2.5%. Grad-CAM visualizations indicate improved lesion localization, confirming the effectiveness of lesion-guided enhancement. This method provides a robust and interpretable solution for automated crop disease diagnosis.
목차
Abstract 1. Introduction 2. Related Work 3. Methodologies 3.1 Disease Spot Region Extraction 3.2 Construction of Four-Channel Input 3.3 Training and Optimization Strategy 4. Experiments 4.1 Experimental Settings 4.2 Overall Performance 4.3 Comparison with Other Models 4.4 Grad-CAM Visualization Analysis 5. Conclusion Acknowledgement References