EGF-Former: An efficient network for structural segmentation and phenotype extraction of sweet peppers in complex environments

The effective acquisition of crop phenotypic structure information is crucial for monitoring and analyzing crop growth. However, in high-density agricultural environments, the complexity and redundancy of the data, coupled with the need for sweet pepper plants to be supported by surrounding frames and wires, significantly complicate the extraction of phenotypic structures. To address these challenges, this paper proposes an EGF-Former-based method for extracting phenotypic structures of sweet peppers in complex environments. The model's ability to capture fine structures is enhanced through the EfficientMod Module (EMM), which extracts information from multiple feature spaces. A refined attention allocation strategy, Global Masked Attention (GMA), is introduced to construct separate masks for foreground and background regions, allowing for accurate segmentation of these areas. Additionally, we propose a module based on the Fast Fourier Transform (FFTM), which improves local contour extraction by obtaining high-frequency information in the frequency domain and performing cross-domain fusion with spatial domain features. This approach significantly enhances the segmentation of edge-blurred phenotypic structures. Experimental results demonstrate that our method exhibits strong robustness compared to state-of-the-art techniques, achieving 83.45 % and 90.37 % in the mIoU and mAcc metrics, respectively. Compared to the baseline model, the segmentation mIoU score improves by up to 3.1 %, while the model’s parameter count is reduced by 10 %, greatly enhancing the usability of EGF-Former in complex agricultural environments.