<![CDATA[Diseases affecting bird's eye chili (Capsicum frutescens L) leaves are a primary cause of reduced harvest productivity. A major challenge in implementing deep learning-based automated detection systems is the drop in accuracy during cross-domain testing, where smartphone images captured in the field contain complex background noise compared to laboratory datasets. This study proposes an SE-MobileNetV2 architecture that integrates a Squeeze-and-Excitation (SE) module to enhance cross-domain classification robustness while maintaining memory efficiency for mobile devices. Testing was conducted on 675 cross-domain images (200 from a Kaggle benchmark dataset and 75 smartphone photos) categorized into five leaf condition classes. Experimental results demonstrate that SE-MobileNetV2 achieved the highest overall accuracy at 91.27%, with a weighted average precision of 92.09%, recall of 91.27%, and F1-score of 91.41%. These results outperform the standard MobileNetV2 (90.55% accuracy) and significantly surpass the VGG16 model (84.00% accuracy). Integrating the channel attention mechanism proved effective in mitigating background interference from soil and shadows, thereby boosting precision for the "Spot" class to 84.75% and the "Whitefly" class to 100.00%. Regarding computational efficiency, SE-MobileNetV2 requires only ~14.8 MB of memory, making it far more compact than VGG16 (~528 MB). Thus, SE-MobileNetV2 achieves an optimal balance between robust field accuracy and file size efficiency, making it an ideal model for offline implementation on farmers' smartphones.]]>
