Answers are grounded in this paper and cite their sources.
To address rising false negatives in automated cherry‑tomato sorting, multi‑scale detection challenges, and the difficulty of deploying advanced models on edge devices, we propose an improved YOLOv10m‑based model. First, the neck’s original concatenation (Concat) is replaced with an Adaptive Multi‑Scale Feature Fusion (AMSFF) module to enhance feature fusion. A dynamic scale‑selection mechanism is introduced to improve detection of challenging, small, and multi‑scale targets. Partial Convolution (PConv) is then applied to the neck’s downsampling operators to reduce computational cost while retaining essential performance. Finally, Ghost Depthwise Convolution (GhostDWConv) substitutes the baseline Depthwise Convolution (DWConv) to better balance efficiency and representational capacity. Compared with the baseline, the enhanced architecture increases recall by 7.06% and improves mean Average Precision by 2.12% (mAP@50) and 2.49% (mAP@[50:95]). Meanwhile, model parameters, model size, and FLOPs are reduced by 26.14%, 25.97%, and 28.55%, respectively. When deployed on the NVIDIA Jetson Orin Nano, the proposed model attains an inference throughput of 58.6 FPS, outperforming industrial camera frame rates and fulfilling real-time constraints for inline cherry tomato sorting. This advancement not only enhances detection precision but also contributes to model lightweighting, thereby offering tangible benefits for improving sorting efficiency and ensuring product quality in cherry tomato production lines.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: