Deep neural networks (DNN) have achieved tremendous development in the past decade. While many DNN-driven software applications have been deployed to solve various tasks, they could also produce incorrect behaviors and result in massive losses. To reveal the incorrect behaviors and improve the quality of DNN-driven applications, developers often need rich labeled data for the testing and optimization of DNN models. However, in practice, collecting diverse data from application scenarios and labeling them properly is often a highly expensive and time-consuming task. In this paper, we proposed an adaptive test selection method, namely ATS, for deep neural networks to alleviate this problem. ATS leverages the difference between the model outputs to measure the behavior diversity of DNN test data. And it aims at selecting a subset with diverse tests from a massive unlabelled dataset. We experiment ATS with four well-designed DNN models and four widely-used datasets in comparison with various kinds of neuron coverage (NC). The results demonstrate that ATS can significantly outperform all test selection methods in assessing both fault detection and model improvement capability of test suites. It is promising to save the data labeling and model retraining costs for deep neural networks.