ALSDet: a global context-enhanced network for detecting small-target diseases on apple leaves

Accurate detection of small-target diseases on apple leaves is of great importance for optimizing orchard management and facilitating precision agriculture. Small-target disease detection remains challenging due to insufficient features and complex backgrounds. This work proposes an effective detector for apple leaf small-target diseases called ALSDet. The global context module is integrated into Stage2 to Stage4 of the ResNet-50 backbone, yielding a refinement of the feature-extraction architecture. In the bottleneck blocks of the Stage2 and Stage3, dilated convolution is used in place of the normal 3×3 convolution to enlarge the receptive field for small targets and strengthen feature extraction. During the model training phase, a multi-scale training strategy combined with the online hard example mining method is adopted to focus on learning hard samples and enhance adaptability for various scale targets. According to the experimental results, ALSDet obtains a mean average precision (mAP) of 65.6% and an average recall (AR) of 71.2% on the dataset. The proposed model achieves the highest levels in both mAP and AR when compared to the popular object detection models, such as Cascade R-CNN, Faster R-CNN, GFL, Grid R-CNN, Libra R-CNN, FCOS, VFNet, RetinaNet, SSD, YOLOv7, and YOLOv8. For small-target diseases like rust and frog eye leaf spot, the average precision surpasses 87% with an intersection over union (IoU) threshold of 0.5. These results confirm that ALSDet achieves stable performance against existing methods, demonstrating its potential as a practical tool for intelligent orchard disease management.