Authors

Abstract

To address edge blurring, soil-clod interference, and limited feature representation in potato image segmentation under sticky-soil conditions, this study proposes an improved MTFormer segmentation model. The model combines the complementary strengths of convolutional neural networks and Transformers, and further enhances segmentation performance through a multi-stage optimization strategy. For more informative and robust feature learning, a residual CNN-based extractor is introduced to strengthen multi-level feature representations. In addition, an MS-CAM attention mechanism is used to reduce channel redundancy, which helps mitigate adhesion-related target confusion in challenging scenes. Building on these features, the TBFE module promotes cross-channel feature fusion, while a Fourier-based FFCM structure compresses and reconstructs deep features in the frequency domain to improve feature compactness. Experiments on our self-built dataset show that MTFormer achieves an F1 score of 85.19%, an mIoU of 84.62%, and a pixel accuracy of 95.67%. Compared with the baseline model, U-Net, and DeepLabV3+, pixel accuracy increases by 1.75, 0.35, and 1.67 percentage points, respectively. Overall, the proposed approach improves segmentation reliability by strengthening feature representation while limiting unnecessary computation, providing practical support for accurate potato segmentation in sticky-soil environments.

Abstract in Chinese

为解决黏性土壤环境下马铃薯图像分割中存在的边缘模糊、土块干扰及特征表达不足等问题，本文提出一种改进的 MTFormer 分割模型。该模型融合了卷积神经网络与Transformer的优势，通过多阶段优化策略提升分割性能，利用残差CNN提取器增强特征层次表达，并嵌入MS-CAM注意力机制抑制通道冗余，以解决目标粘连与混淆问题，引入TBFE模块跨通道特征融合，采用基于傅里叶变换的FFCM结构对深层特征进行频域压缩与重构，显著提高了特征的紧凑性。试验表明，MTFormer在自建数据集上的F1 分数达到85.19% ，mIoU达到84.62%，像素准确率为95.67%。相较于原始模型、U-Net和DeepLabV3+，像素准确率分别提升1.75、0.35和1.67个百分点。该方法有效平衡了特征表达与计算冗余，可为黏性土壤环境下的马铃薯精准分割提供可靠技术支撑。