Evaluation of SE, ECA, and CA Modules in ResNet-50 for the Classification of Grape Leaf Diseases

Context: Plant diseases pose a major threat to agricultural production, as they hinder crop growth and significantly reduce yields. In this context, the automatic detection of diseases using artificial intelligence techniques represents an effective solution. Objective: This work presents a comparative study on the integration of attention mechanisms into the ResNet-50 architecture, applied to the classification of grape leaf diseases. Method: Three attention modules were considered: SE (Squeeze-and-Excitation), ECA (Efficient Channel Attention), and CA (Coordinate Attention), each integrated separately into the model to evaluate its impact. Results: The performance of the different configurations was assessed by considering the quality of the results, the model complexity, and the processing speed. Experimental results show that the ECA module achieved the best performance with a reduced number of parameters, followed by SE and then CA. Conclusion: This study highlights the contribution of attention mechanisms in improving the effectiveness of CNN architectures for plant disease detection.