关键词: 智慧农业, 谷穗计数, 状态空间模型, 快速傅里叶变换, 频域增强

Abstract: Convolutional neural network (CNN) and Transformers, which are widely used for Earhead counting, continue to struggle with capturing long-range dependencies efficiently and often exhibit false positives or false negatives in densely populated regions. To overcome these limitations, a novel counting approach named PCLMamba is proposed in this paper. This lightweight network is built on a Mamba-based architecture and incorporates segmentation guidance and frequency domain enhancement. A Grouping Directional Mamba Model (GDMM) is introduced, which processes image sequences in groups and performs parallel selective scanning in different orders, thereby efficiently capturing long-range dependencies with low computational cost. Local features and globally scanned features are integrated via a Channel Attention Fusion Module (CAFM). To reduce false positives and negatives, a segmentation module (Segmenter) is designed to supply prior knowledge, which suppresses background interference while highlighting foreground regions. Additionally, a Frequency Spatial Enhancement module(FSE) is proposed to reinforce intermediate features in both spatial and frequency domains, further boosting counting accuracy. Experimental evaluations on multiple plant counting datasets including MTC, WED, and SHC perform well. In particular, the Mean Absolute Error (MAE) of the three datasets above reaches 4.4, 4.2, and 15.4 respectively, and other performance is also better than the comparison method as a whole.

Key words: smart agriculture, earhead counting, State-Space Model (SSM), Fast Fourier Transform (FFT), frequency enhancement