Resumen
Recognition of plant stems is vital to automating multiple processes in fruit and vegetable production. The colour similarity between stems and leaves of tomato plants presents a considerable challenge for recognising stems in colour images. With duality relation in edge pairs as a basis, we designed a recognition algorithm for stems of tomato plants based on a hybrid joint neural network, which was composed of the duality edge method and deep learning models. Pixel-level metrics were designed to evaluate the performance of the neural network. Tests showed that the proposed algorithm has performs well at detecting thin and long objects even if the objects have similar colour to backgrounds. Compared with other methods based on colour images, the hybrid joint neural network can recognise the main and lateral stems and has less false negatives and positives. The proposed method has low hardware cost and can be used in the automation of fruit and vegetable production, such as in automatic targeted fertilisation and spraying, deleafing, branch pruning, clustered fruit harvesting and harvesting with trunk shake, obstacle avoidance, and navigation.