DPDB-YOLO: A lightweight YOLOv13 cherry tomato ripeness detection method with adaptive extraction module and multi-scale feature fusion architecture

In this paper, we propose a lightweight and efficient cherry tomato ripeness detection model, named DPDB-YOLO, based on YOLOv13n, for fast and accurate detection in natural environments. The improvements are as follows: first, the DPC3k2 (DWConv+PConv+C3k2) module replaces the DSC3k2 module in the backbone and neck as well as the A2C2f module, constructing a compact feature extraction unit that improves accuracy and reduces parameter overhead. Secondly, structural DLAE (Depth Light-weight Adaptive Extraction) is introduced in the backbone instead of ordinary convolution to enhance adaptive learning in key regions and reduce computation. In addition, structural BSMFM (Bounded Sigmoid Modulation Fusion Module) is used in the neck instead of FullPaD to strengthen spatial perception and semantic discrimination. Experiments show the model improves accuracy by 4.88 %, recall by 4.84 %, F1 score by 4.86 %, mAP50 by 3.13 %, mAP50–95 by 8.13 %, with parameters reduced by 40 %, model size by 38 %, and GFLOPS by 20 % compared with the original. Compared to the SSD model, the EfficientDet model, and other YOLO series models, it achieves superior detection with fewer parameters, validating its effectiveness for embedded devices and providing accurate support for automated harvesting. • A novel DPC3k2 module forms an efficient feature extraction unit. • The Depth Light-weight Adaptive Extraction (DLAE) architecture enhances region-adaptive learning. • A Bounded Sigmoid Modulation Fusion Module (BSMFM) to enhance the distinction between spatial and semantic features. • The DPDB-YOLO model demonstrates its immense potential for real-world edge deployment on the NVIDIA Jetson Orin Nano.