Improving U-net network for semantic segmentation of corns and weeds during corn seedling stage in field

Weeds are one of the main factors affecting crop growth, making weed control a pressing global problem. In recent years, interest in intelligent mechanical weed-control equipment has been growing.

We propose a semantic segmentation network, RDS_Unet, based on corn seedling fields built upon an improved U-net network. This network accurately recognizes weeds even under complex environmental conditions, facilitating the use of mechanical weeding equipment for reducing weed density. Our research utilized field-grown maize seedlings and accompanying weeds in expansive fields. We integrated the U-net semantic segmentation network, employing ResNeXt-50 for feature extraction in the encoder stage. In the decoder phase, Layer 1 uses deformable convolution with adaptive offsets, replacing traditional convolution. Furthermore, concurrent spatial and channel squeeze and excitation is incorporated after ordinary convolutional layers in Layers 2, 3, and 4.

Compared with existing classical semantic segmentation models such as U-net, Pspnet, and DeeplabV3, our model demonstrated superior performance on our specially constructed seedling grass semantic segmentation dataset, CGSSD, during the maize seedling stage. The Q6mean intersection over union (MIoU), precision, and recall of this network are 82.36%, 91.36%, and 89.45%, respectively. Compared to those of the original network, the proposed network achieves improvements of 5.91, 3.50, and 5.49 percentage points in the MIoU, precision, and recall, respectively. The detection speed is 12.6 frames per second. In addition, ablation experiments further confirmed the impactful contribution of each improvement component on the overall semantic segmentation performance.

This study provides theoretical and technical support for the automated operation of intelligent mechanical weeding devices.