Channel Coordination Attention for Crop and Weed Segmentation Neural Networks

Abstract

Convolutional neural network is widely used in the field of intelligent agricultural production, but in weed control, the complexity of the environment and the similarity between plants place higher demands on the task of crop and weed image recognition and segmentation. The attention mechanism, as a plug and play module, has been shown to significantly improve the network performance by dynamically weighting the input images to achieve the adjustment of the neural network attention. In this paper, a hybrid attention mechanism was proposed to establish global feature dependence by using channel autocorrelation matrix, named channel coordination attention. The attention mechanism established an autocorrelation matrix by obtaining the dependency relationships of feature channels, and used this matrix to aggregate spatial features in different directions to strengthen the location features. The resulting pair of spatial attention and previously obtained channel attention worked together on the input feature map to enhance feature representation capability. This study establishes an Unet network model to compare the performance of the proposed attention mechanism with other mainstream attention in crop and weed segmentation application scenarios. The experimental results showed that the proposed attention mechanism outperformed current mainstream attention mechanisms, such as Squeeze-and-Excitation attention (SE), Convolutional Block Attention Module (CBAM), Coordinated Attention (CA), etc., in terms of interaction ratio, accuracy, and average pixel accuracy. Finally, by comparing the performance in different recognition algorithms, it was proven that the proposed attention mechanism can be applied to different network models and significantly improve its segmentation performance.