Enhanced Small Lesion Detection on Grape Leaves Using Optimized YOLOv7

The detection of small lesions on grape leaves plays a pivotal role in the early identification and management of plant diseases, potentially reducing significant agricultural losses. Traditional detection systems, including advanced versions such as YOLOv7, often overlook small target diseases due to resolution constraints and inadequate feature extraction capabilities. Addressing this, our research enhances the YOLOv7 architecture by incorporating a specialized prediction branch and leveraging an improved channel attention mechanism in conjunction with an Extended-Efficient Long-range Attention Network (E-ELAN). Firstly, a new detection head specifically designed for small targets enhances the resolution capabilities of the system. Secondly, the integration of asymmetric convolution into the E-ELAN facilitates refined multi-scale feature extraction, critical for distinguishing minute pathological changes. Further, we enhance the model's sensitivity to small-scale features through a revamped channel attention mechanism. Moreover, we replace the conventional Complete Intersection over Union (CIoU) with a Structured Intersection over Union (SIoU) in the loss function, significantly improving the model's localization accuracy. Experimental results from testing the optimized YOLOv7 on a dataset of grape leaves affected by three prevalent diseases show an average accuracy improvement of 2.7%, achieving a remarkable 93.5%. This substantiates our model's efficacy in enhancing small lesion detection, thereby contributing effectively to early disease intervention strategies in viticulture.