A Unified Model for Real-Time Crop Recognition and Stem Localization Exploiting Cross-Task Feature Fusion

Robotic mechanical weed control is a promising solution to reduce the usage of herbicides. Efficient and accurate detection of crop stems is the premise of most robotic mechanical weeding machines. This paper proposes a unified convolutional neural network model, called UniStemNet, for real-time crop recognition and stem detection. The UniStemNet consists of a backbone network and two subnets to perform the two tasks simultaneously. According to the difference of targets in the two tasks, the varied-span feature fusion structure is established in the subnets. To improve the stem detection performance, a cross-task feature fusion strategy is devised which introduces a top-down guidance from the crop recognition subnet to the stem detection subnet. Experimental results demonstrate that the proposed UniStemNet can significantly outperform the state-of-the-art crop stem detection method, and perform comparably with leading-edge crop recognition methods. The results also validate the remarkable effect of the cross-task feature fusion strategy on improving the stem detection performance. The UniStemNet can process a 400x300 image within 6 ms. The code and dataset are available at https://github.com/ZhangXG001/Real-Time-Crop-Recognition-and-Stem-Localization.git.