Research on a Multi-Scale Tomato Ripeness Detection Method Based on an Enhanced YOLOv13s

Intelligent tomato harvesting robots are expected not only to satisfy basic harvesting needs but also to increase picking accuracy and efficiency, where tomato ripeness detection is especially important. This study investigates target recognition of tomato fruit ripeness under diverse cultivation environments. To improve detection accuracy and recall rate, an enhanced YOLOv13 algorithm (YOLOv13-SSW) is proposed. First, the SimAM attention mechanism is used to replace the last backbone layer of the YOLOv13s model. SimAM adaptively adjusts pixel weights by computing the similarity between each pixel and its neighboring pixels in the feature map, thereby highlighting salient features while suppressing irrelevant information. The DS-C3k2 convolution module in the backbone and neck structures is replaced with the DS-C3k2_SAConv module. This module adopts SAConv to reduce spatial and channel redundancy during feature extraction and fusion, thereby increasing detection speed and decreasing model parameters. Finally, the loss function is optimized by replacing CIoU-Loss with WIoUv3 to accelerate convergence and improve convergence precision. Based on tests using 5,000 data samples over 300 epochs, the results show that the YOLOv13-SSW model achieves an accuracy of 91.612%, a recall rate of 89.678%, and an average precision of 92.319%. Compared with YOLOv13s, this corresponds to a 5.553% increase in accuracy, a 4.62% gain in recall rate, and an 8.033% rise in average precision. Comparative experiments with common models indicate that YOLOv13-SSW provides strong generalization ability, localization performance, detection performance, and robustness. Ablation experiments also confirm the necessity of each improvement, demonstrating that no redundant operations are introduced. These methods offer a useful reference for addressing similar tasks.