Precise recognition of two-dimensional crop imagery constitutes a foundational prerequisite for in-field digital perception and scientific agricultural decision-making. However, unstructured field environments present significant challenges, characterised by substantial variations in organ scale and frequent occlusion. These issues are particularly pronounced in pepper plants, which exhibit dense canopy architectures and ambiguous boundaries. Existing segmentation models often fail to adequately delineate slender morphological features due to restricted receptive fields and insufficient directional modelling. Furthermore, background textural similarity and occlusion interference frequently lead to boundary distinctness degradation and semantic confusion, thereby compromising overall recognition performance. To address these limitations, this study proposes CPO-SwinUnet, an enhanced Swin-Unet architecture optimised for the organ-level structural segmentation of pepper plants. The framework incorporates a Strip Pooling module at the encoder-decoder interface to augment long-range dependency representation through directional context aggregation, thereby enhancing the holistic perception and connectivity of slender structures. Additionally, a FreqFusion v2 module is engineered within the cross-scale feature fusion stage, employing difference-guided mechanisms to intensify edge response and reduce boundary blurring and class confusion. Experimental validation on a custom dataset spanning the full growth cycle demonstrates that the proposed method significantly outperforms the baseline Swin-Unet in complex field scenarios. With a 0.91% improvement in mean Intersection over Union (mIoU), the method achieves superior accuracy and stability in the segmentation of pepper organ regions.
Wang et al. (2026) studied this question.