Abstract: Tomato-fruit detection, a key technology for intelligent harvesting, encounters significant challenges in complex field conditions characterized by background clutter, foliage occlusion, fruit overlap and the resulting scale variations. To address these difficulties, this paper proposes YOLO-Vega, a lightweight object-detection model built on the YOLO11n architecture. YOLO-Vega integrates an adaptive feature fusion network (AFFNet) that uses a multi-channel dynamic-weight mechanism to adaptively fuse multi-scale features derived from both the fruit itself and its surroundings; a mixed-space edge enhancement module (MSEE-C3k2) that reinforces blurred edges caused by occlusion and overlap through a combination of multi-scale context awareness and explicit high-frequency edge injection; and a key-weight detection module (KW-Detect) that improves feature selection under background interference and enhances recognition of occluded targets by generating and applying task-specific key-information weights. Experiments on benchmark datasets show that YOLO-Vega achieves an mAP@0.5 of 88.62% while maintaining a low computational overhead (2.65×106 parameters, 7.3 GFLOPs) and a compact model size (6.0?MB), outperforming mainstream models in these complex scenarios and offering an excellent balance between accuracy and efficiency.

Key words: tomato detection, lightweight object detection, YOLO11n, adaptive feature fusion, edge enhancement

摘要： 番茄果实检测作为智能采摘的关键技术，在背景干扰、枝叶遮挡、果实重叠及其引发的尺度变化的复杂环境中面临显著挑战。为应对这些问题，提出一种轻量级目标检测模型YOLO-Vega。该模型在YOLO11n架构基础上，提出三个针对番茄检测特性深度优化的关键模块：自适应特征融合网络（adaptive feature fusion network，AFFNet）通过其独特的多通道动态权重机制实现对番茄固有及环境诱导的多尺度特征的自适应融合；混合空间边缘增强模块（mixed-space edge enhancement module， MSEE-C3k2）通过其新颖的多尺度上下文感知与显式高频边缘信息注入设计，有效强化了因遮挡和重叠导致的模糊边缘；关键信息权重检测模块（key-weight detection， KW-Detect）则凭借其独特的关键信息权重生成与应用策略，显著提升了在背景干扰下的特征选择能力和对被遮挡目标的辨识度。在实验数据集上的评估结果表明，YOLO-Vega实现了88.62%的mAP@0.5，同时保持了较低的计算开销（参数量2.65×106，计算量7.3?GFLOPs）和模型大小（6.0?MB），在应对上述复杂环境时表现出色，综合性能优于现有主流模型。

关键词: 番茄检测, 轻量级目标检测, YOLO11n, 自适应特征融合, 边缘增强