YOLOv8s-Swin: Enhanced Tomato Ripeness Detection for Smart Agriculture

Abstract

Accurate object detection and classification are paramount in precision agriculture for assessing ripeness stages and optimizing yield, particularly for high-value crops like tomatoes. Traditional manual inspection methods are laborious, time-consuming, and error-prone. Furthermore, existing deep learning models often struggle with real-world agricultural challenges such as varying lighting, occlusions from foliage or other fruits, and dense clustering of small objects. To address these limitations and enhance tomato production efficiency and quality in diverse agricultural conditions, this study introduces YOLOv8s-Swin, an advanced object detection model. YOLOv8s-Swin integrates the powerful YOLOv8s architecture with a Swin Transformer module (C3STR) to capture global and local contextual information, crucial for robust small object detection. It also incorporates Focus, Depthwise Convolution (DWconv), Spatial Pyramid Pooling with Contextual Spatial Pyramid Convolution (SPPCSPC), and C2 modules for preserving fine details, reducing computational overhead, enhancing multi-scale feature fusion, and improving high-level semantic feature extraction, respectively. The Wise Intersection over Union (WIoU) loss function is adopted to enhance localization and address convergence issues. Evaluated on a comprehensive tomato image dataset, YOLOv8s-Swin demonstrated superior performance with a mean Average Precision (mAP@0.5) of 88.3%, precision of 84.4%, recall of 79.9%, and an F1-Score of 0.821. This significantly surpasses the base YOLOv8s (84.7% mAP@0.5, 0.795 F1-Score) and other models like Faster R-CNN, SSD, YOLOv4, YOLOv5s, and YOLOv7, all under identical conditions. Maintaining a competitive inference speed of 166.67 FPS, YOLOv8s-Swin offers a robust and efficient solution for AI-driven crop management and sustainable food production.