RT-DETR-MC: A Novel Fusion of Multi-Scale Features and Gating Mechanism for Lettuce Disease Detection

To tackle the issue of low efficiency in early detection of lettuce leaf spot disease within plant factories that feature high-density planting, this study puts forward an enhanced RT - DETR detection model. The cross - scale fusion module of RT - DETR - r18 has been restructured. A collaborative framework incorporating a multi - scale channel - aware feature fusion module (MS - CAFF) and convolutional gated linear unit (MS - CGLU) was devised. By integrating multi - scale convolution with channel attention mechanisms, the capacity for multi - scale feature extraction is boosted. The conventional loss function was swapped out for normalized Gaussian distance regression loss (NWD), which effectively addresses the difficulty of detecting overlapping lesion areas in intricate scenarios. The experimental findings underscore that the enhanced architecture achieves mAP50 and mAP50:95 metrics of 92.6% and 65.9% in tandem, culminating in a 6.6% performance elevation when juxtaposed against its progenitor framework. There's also a 4.9% rise in precision and a 4.6% increase in recall. When contrasted with well - known models such as YOLOv5 and YOLOv8, this model stands out in terms of detection precision and robustness, presenting an efficient approach for intelligent disease control in the context of plant factories.