RDA-YOLOv8n: A Directionally Refined and Structurally Optimized Approach for Tomato Maturity Detection

Maturity assessment in color-sensitive crops such as tomatoes is a critical requirement for automated harvesting and quality management, yet it remains challenging due to subtle phenotypic differences that are easily affected by illumination variation, occlusion, and complex field backgrounds. These factors limit the effectiveness of lightweight object detectors, while aggressive data augmentation and generic attention mechanisms often provide limited benefits or introduce additional computational overhead. To address these challenges, this study presents a lightweight tomato maturity detection framework based on YOLOv8n that integrates detection-head semantic refinement with structural compression using Ghost-style and depthwise–pointwise convolutions. A Residual Directional Attention (RDA) module was introduced at the detection head to enhance high-level semantic representation through directional depthwise spatial filtering, channel-selective gating, and controlled residual modulation, enabling improved discrimination between visually adjacent ripeness stages without altering color distributions. To preserve real-time efficiency, a compression strategy replaces redundant high-capacity operators in the detection head with lightweight alternatives, reducing model size and computational cost without modifying the detection pipeline. Experimental results on a tomato ripeness dataset show that the proposed framework Yolov8n +RDA-lite improves detection precision from 79.8% to 85.4% and increases localization accuracy from 86.8% to 88.0% [email protected] and from 74.7% to 75.9% [email protected]:0.95, while reducing the parameter count from 3.01 million to 2.68 million. Performance gains are most pronounced for intermediate ripeness stages characterized by high interclass similarity, while stable behavior is maintained for clearly separable classes. The results demonstrate that targeted detection-head refinement combined with structural model compression provides an effective balance between detection accuracy and computational efficiency for tomato maturity assessment.