SE-AF CNN: An Attention-Enhanced Lightweight Network for TomatoLeaf Disease Detection

Abstract

To reduce crop losses and increase agricultural productivity, tomato leaf diseases must be identified early. In order to automatically classify tomato leaf diseases, this research introduces SE-AF CNN, a lightweight convolutional neural network that combines dual-path Attention-Fusion (AF) and Squeeze-and-Excitation (SE) channel attention. A balanced subset of the PlantVillage dataset with ten tomato leaf condition classes is used to test the suggested architecture, which is completely trained from scratch. On the validation set, experimental results show a weighted recall of 96.80%, weighted precision of 97.01%, and overall classification accuracy of 96.80%. The suggested design maintains a small model size appropriate for settings with limited resources while achieving good classification performance. The suggested model, which has about 8.2 million parameters, offers a good trade-off between computing efficiency and accuracy, making it appropriate for use in precision agriculture applications with limited resources. In order to increase the model's practical resilience and generalisation, future research will concentrate on assessing it in real-world scenarios.