A Dual-Backbone Collaborative Feature Enhancement Network for Low-Contrast Small-Object Detection in Parallel Fruit-Sorting Robots

Detecting grape stems in clustered fruits is challenging because of their small size, low contrast with surrounding fruits, and frequent occlusion. To address these challenges, this study proposes ELCD-YOLO, a dual-backbone collaborative feature-enhancement network for grape-stem detection in the visual perception stage of parallel fruit-sorting robots. First, YOLOv12s and EfficientNetV2-N are employed as heterogeneous parallel backbones to extract complementary semantic and local texture features from the same input image. Cross-backbone feature fusion is combined with an adapted lightweight cross-layer reconstruction (LR) module to reorganize the fused P3, P4, and P5 features into higher-resolution P2, P3, and P4 detection representations without increasing the number of detection heads. DySample is employed within the cross-scale reconstruction pathways for content-adaptive spatial alignment, while CAFM is serially introduced after A2C2 for complementary contextual–local feature refinement. These components are coordinated within a unified architecture to address the coupled challenges of small target size, low contrast, occlusion, and complex background interference. Experiments were conducted on a grape-stem detection dataset constructed primarily from images acquired using the robotic platform, with data augmentation restricted to the training set and identified acquisition-group overlaps removed from the evaluation subsets. Across three independent training runs, ELCD-YOLO achieved an [email protected] of 68.00 ± 0.25%, Precision of 92.51 ± 0.25%, and Recall of 59.50 ± 0.35%, representing improvements of 2.89, 3.10, and 4.18 percentage points over YOLOv12s, respectively. The model requires 21.5 GFLOPs, occupies 23.6 MB, and achieves 81 FPS for detector-pipeline inference on an NVIDIA RTX 4070. These results demonstrate improved grape-stem detection performance while maintaining real-time inference capability, supporting further investigation of ELCD-YOLO as a visual perception component for parallel fruit-sorting robots.