Performance Optimization of Wildlife Recognition with Distributed Learning in Ecological Surveillance

Recently, deep learning technology has been widely used in ecological environment monitoring, but there are still huge challenges in specific tasks. For example, data sets have increased over the past few years, and in a CPU environment, training is quite time consuming. In order to solve these problems, we bring up a novel and effective method to improve training speed. In this paper, we make use of multi-GPUs to achieve distributed parallel acceleration. Small and shallow networks are not suitable for distributed training, because the computation of each parameter in this network is much higher than that of multi-layer perception or automatic encoder architecture, we use convolutional neural network with parameter sharing as the training network. In this paper, the main adopted method is to compare the performance of training wildlife recognition in single GPU and multi-GPUs environments by using distributed deep learning framework TensorFlow. The experimental results show that multi-GPUs which adopt distributed architecture can significantly accelerate training time consumption than single GPU. The results of this experiment also provides strong support for our follow-up work.