CMNet: an asymmetric dual-branch network for accurate cotton segmentation

Gengrong Zhang¹Halidanmu Abudukelimu^1*Mayilamu Musideke¹Shuqin Wu¹Abudukelimu Abulizi¹Cuiqin Guo¹Yajun Zhang²

• ¹Xinjiang University of Finance and Economics, Ürümqi, China
• ²Xinjiang University, Urumqi, China

In agricultural automation, precise cotton segmentation is a key step for tasks such as intelligent harvesting and yield estimation. However, in complex field environments, factors such as background interference and irregular target shapes severely affect segmentation accuracy. Existing deep learning methods offer certain advantages but still generally suffer from limitations including insufficient accuracy, over-segmentation, and misidentification. To address these challenges, this study proposes a novel dual-branch cotton segmentation network, Cotton-aware Mamba-enhanced UNet (CMNet), which optimizes the ParaTransCNN architecture by incorporating the 2D Selective Scan (SS2D) module to replace the original Transformer branch, effectively balancing the extraction of local details and global semantic information while reducing computational burden. To enhance the model's perception of irregularly shaped cotton, a Deformable Convolutional Networks v1 (DCNv1) module is integrated into the Vision Mamba (VMamba) branch, further improving the delineation of target boundaries. Additionally, an Atrous Spatial Pyramid Pooling (ASPP) module is introduced at the end of the Convolutional Neural Network (CNN) branch to strengthen multi-scale feature representation. To optimize the fusion of channel and spatial information, the Spatial and Channel Squeeze-and-Excitation (scSE) attention mechanism replaces the original module, enhancing feature modeling capability. Experimental results on an in-field cotton image dataset demonstrate that CMNet outperforms existing mainstream methods, achieving Dice, mIoU, and Accuracy of 91.06%, 84.18%, and 98.10%, respectively, while reducing parameter count and computational complexity, thus exhibiting excellent performance. Furthermore, generalization experiments on multiple other plant datasets also achieved outstanding results, validating the model's adaptability and potential for broader applications in multi-crop segmentation tasks, providing valuable insights for smart agriculture segmentation research. The source code and dataset of this work are publicly available at https://github.com/halidanmu/CMNet.git.