Classifying a Limited Number of the Bamboo Species by the Transformation of Convolution Groups

Abstract

For agricultural special species, the labeled procedure of large-scale samples is costly, thus, the bamboo species only has a limited number for supervised learning. The fine-tuning strategy is important for deep neural network by transferring learning methods, which utilize the weight of the deep model of the source domain, and can solve the problem associated with insufficient samples to make the model more stability and robustness. In the manuscript, the novelty of the strategy, for images of bamboo species with low-shot classification, mainly proposed an idea that is the transfer of the convolutional group features of deep convolutional models. The deep models with a novel fine-tuning method and three optimizers that are stochastic gradient descent, Adaptive Moment estimation, and Adadelta respectively, are evaluated by the accuracy and the expected calibration error value for the analysis of deep model generalization. An analysis of the results showed that, based on the proportion of training dataset is only 30%, the innovative strategy for bamboo species classification achieved better performance that has an accuracy of 0.82, and the expected calibration error of 0.16, which were better stability and generalization than those of other fine-tuning strategies. Consequently, the novel fine-tuning strategy proposed in this manuscript transfers the features of deep convolutional groups, improves the accuracy and generalizability of the model, and resolves the problems associated with having insufficient samples of bamboo species for low-shot classification.