This study proposes a new deep neural network based camera calibration method that achieves flexible, high-precision calibration in complex environments, without having to classify or extract features from input data. By optimizing the network structure, hyperparameters, and training algorithms, the deep neural network can be quickly and effectively trained. The experimental results confirm that, compared with Zhang's calibration method and the shallow neural network, the proposed method can achieve high calibration accuracy under a wide range of imaging conditions involving multiple shooting angles or high distortion. For the images produced using a highly distorted lens, the proposed method achieves an average calibration error of only 0.1471 mm over the calibration range of 633 mm×763 mm.