Abstract: In response to the challenges faced by existing semantic segmentation networks, which were struggled with real‑time recognition in complex lighting conditions and tended to misclassify weed and crops in overlapping regions, a real‑time segmentation network called FA—BiSeNetV2 was proposed, which integrated frequency features for the recognition of sugar beets and weeds. Taking the semantic branch of the BiSeNetV2 model as the starting point, this approach achieved frequency feature extraction at different levels by adding 2D discrete cosine transform layers after each gather and expansion layer. The extracted frequency features were then subjected to an adaptive frequency processing module to analyze the distribution of scene data. This approach also involved combining the processed frequency features from various stages through weighted summation to obtain multi‑level frequency features. Finally, a multi‑scale spatial frequency fusion module was introduced to merge features from aggregation layers and multi‑level frequency features, achieving spatial feature reconstruction from global and local perspectives. The experiments results of RoniRob in the publicly available dataset showed that the mean intersection over of FA—BiSeNetV2 model was 87.12% and the mean pixel accuracy was 93.04%, which were higher than the BiSeNetV2 model by 4.71% and 6.87%, respectively, the FA—BiSeNetV2 model was highly efficient with only 3.138 M parameters.

Key words: sugar beet, weed recognition, frequency feature, semantic segmentation, lightweighting, real?time segmentaton

摘要： 针对现有的语义分割网络在复杂光照条件下识别实时性差、对杂草与作物重叠区域易误分类的问题，以甜菜和杂草作为识别对象，提出融合频率特征的实时分割网络FA—BiSeNetV2。首先，以BiSeNetV2模型中的语义分支为出发点，在各聚集扩展层后加入二维离散余弦变化层提取出不同层次的频率特征；其次，对频率特征采用自适应频率处理模块以解析场景数据分布，将处理后的各阶段频率特征加权求和得到多层次频率特征；最后，采用多尺度空间频率融合模块对聚合层特征和多层次频率特征从全局和局部两个方面进行融合，实现空间特征重构。在公开数据集RoniRob的试验结果表明，FA—BiSeNetV2模型的平均交并比达87.12%，平均像素精度为93.04%，相比BiSeNetV2模型，分别提高4.71%、6.87%，参数量仅为3.138 M。

关键词: 甜菜, 杂草识别, 频率特征, 语义分割, 轻量化, 实时分割