SE-YOLO: A sobel-enhanced framework for high-accuracy, lightweight real-time tomato detection with edge deployment capability

Abstract

Tomato detection is one of the most crucial steps in automated tomato harvesting. However, detection in agricultural settings faces challenges from complex environments and computational constraints. Here, we propose SE-YOLO, a lightweight framework with fundamentally different technical approach compared to existing methods. While current SOTA methods like RT-DETR rely on complex transformer architectures for global attention and Hyper-YOLO employs hypergraph computation for high-order feature relationships, SE-YOLO introduces a more direct and efficient paradigm through explicit edge perception. This approach includes: (i) SPStem’s integration of Sobel operators at network entry to directly extract edge features—a departure from the conventional approach of implicitly learning edges through multiple convolutional layers used in existing frameworks; (ii) ADown’s efficient channel-split architecture; (iii) SEDFF’s residual-based edge enhancement throughout the backbone that preserves both edge and semantic information; and (iv) WFPIoU’s adaptive spatial supervision that dynamically adjusts loss weights across training phases—contrasting with the static penalty terms in conventional IoU losses like CIoU and EIoU. These innovations create a detection framework specifically optimized for agricultural environments, capturing crucial boundary information for occluded fruits while maintaining minimal computational overhead. SE-YOLO achieves 93.6% [email protected] and 67.3% [email protected]:0.95, outperforming state-of-the-art detectors while using only 4.3% of RT-DETR’s computational cost Subsequently, the developed SE-YOLO model is deployed on standard edge hardware in INT8-quantized form. The deployment achieves 18.7 frames per second (FPS), with an 88.8% [email protected] and a 61.6% [email protected]:0.95, showing gains of 3.2% and 2.4%, respectively, over the YOLO11n model. Finally, the generalization performance of SE-YOLO is tested and validated across other datasets, demonstrating its effectiveness in detecting occluded and partially visible tomatoes, as well as in maturity-related detection tasks.