CropRowSeg: A lightweight model for seedling crop row image segmentation combining multi-scale CNN and transformer

Crop row visual perception can provide navigation information for autonomous row-following equipment and its bottleneck is how to segment target crop rows from complex field images. Existing methods are often designed for specific scenarios or crops, suffering from complex hyperparameter tuning, large parameter sizes, high computational costs, poor convergence, and weak generalization. Departing from traditional row-by-row segmentation, we treat target crop rows aligned with the navigation path as foreground, while categorizing weeds, soil, and other rows as background, establishing an efficient annotation method. We propose CropRowSeg, a semantic segmentation model based on an encoder-decoder architecture. The encoder integrates parallel Convolutional Neural Network (CNN) and Vision Transformer (ViT) branches with hierarchical feature fusion, enabling local detail perception and global context modeling to address long-range dependencies. In the ViT branch, computational redundancy is reduced via feature dimensionality reduction and spatial-channel decoupled convolution, along with local feature enhancement. A hybrid attention adaptively calibrates feature importance across spatial and channel dimensions. The decoder employs a progressive architecture combining residual connections and coordinate attention to reconstruct details and mitigate information loss during upsampling. We introduce a special loss function to address foreground-background imbalance and spatial confusion between symmetrical rows. Experiments on a diverse crop row dataset demonstrate that CropRowSeg achieves accurate segmentation across environments, growth stages, and crop varieties, showing strong generalization. With only 6.91 M parameters and 12.06 Giga Floating-point Operations Per Second (GFLOPs), the model attains an mIoU of 82.55%, a Recall of 92.75%, and an F1-score of 90.26%. The improved model performance will support deployment on edge devices and promote practical applications.