Study of DNN Network Architecture Search for Robot Vision

Robot vision, which integrates measurement and perception, plays an important role in robot manipulation applications. Although the current deep learning-based visual neural network models have high perception capabilities, their network structures are usually too large to be implemented on embedded devices for robot vision. In this paper, we propose neural network architecture search (NAS) that combines pre-training and pruning operations to simplify deep neural network architectures. It not only solves this problem without losing network accuracy, but also significantly alleviates the difficulties of long network computation time and redundant search space in traditional NAS methods. Finally, the experimental results show that the neural network generated by the proposed algorithm outperforms the artificially designed neural network, which demonstrates the effectiveness of the method. At the end of the paper, the rationality of the method is proved by experiments and comparisons. The performance of the new algorithm and the generated neural network is better than that of the artificially designed neural network.