YOLOv8-Ghost-QWA: A Lightweight Model with Quadro-Weighted Attention for Tomato Maturity Segmentation

Robust, real-time recognition of tomato maturity in unconstrained field conditions remains a critical bottleneck for autonomous harvesting platforms, as manual methods are costly and labor-intensive. Existing approaches, such as Vision Transformer-based models and YOLO variants with attention modules, often overfit to controlled environments, struggle with surface anomalies, and demand excessive computational resources unsuitable for edge devices in robotic harvesters. To address these challenges, we introduce the modified YOLOv8-nano architecture, which incorporates Ghost convolutions to reduce feature redundancy and a lightweight adaptive fusion module (Quadro-Weighted Attention, QWA) block that adaptively fuses Efficient Channel Attention (ECA) with 3 × 3 Triplet Attention (TA) via a learnable sigmoid gate. In addition, we present the iTomato-Seg-UA dataset, which spans three ripening stages (green, ripening, and ripe) under diverse illumination, backgrounds, cultivars, and anomalies (cracks, catfacing, and zippering). The proposed YOLOv8-Ghost-QWA model demonstrates a 77% F1 score and 82% mAP on the merged test subset while requiring 39% fewer parameters and 25% fewer GFLOPs compared to vanilla YOLOv8. These results demonstrate a competitive efficiency-accuracy trade-off for maturity segmentation on edge devices, highlighting the practical potential of lightweight attention-based enhancements.