A lightweight DSD-UNet-based method for clustered litchi segmentation

Automated litchi harvesting demands clustered fruit segmentation with both high accuracy and real-time performance, yet high computational complexity and parameter redundancy pose critical bottlenecks for practical deployment. To tackle these challenges, we propose a lightweight DSD-UNet method for fast and high-precision segmentation of litchi clusters. The DSD-UNet incorporates depthwise separable convolutions to replace traditional convolutional structures, significantly reducing parameters and computational complexity. To further enhance the model's performance, we introduce a dual cross-attention mechanism into the skip connections,which integrates spatial and channel attention to adaptively strengthen the representation of target regions in feature maps and improve sensitivity to slender stem features. Additionally, we develop a composite loss function (BL-Loss), combining binary cross-entropy with level-set loss, and optimize it using a staged training strategy to alleviate within-foreground pixel imbalance and preserve the boundary continuity of slender fruit-stem connections. Experimental results under varying illumination conditions demonstrate that DSD-UNet achieves a mean Intersection over Union (mIoU) of 91.40% and an F1-score of 95.42% for the segmentation of clustered litchi, with an inference speed of 15.41 frames per second and a parameter count of only 5.16M. Compared with the original U-Net, our DSD-UNet model improves mIoU and F1-score by 3.08 and 1.81 percentage points, respectively, increases inference speed by 138.9%, and achieves a 61.5% reduction in parameter count, confirming that the proposed method effectively enhances segmentation accuracy while maintaining model lightweightness. These results underscore the potential of our method to significantly improve automated harvesting processes in litchi production.