The key to the image source identification algorithm based on sensor pattern noise is how to obtain high-quality PRNU noise. At present， most of the algorithms include to improve the quality of PRNU noise and compress them for purposes of practical application based on artificially assumed models. This paper proposes a purification scheme of image PRNU noise based on cross-matching sample training. The scheme uses a deep stacked auto-encoder network and designs a technology of sample cross-matching to train it. The use of this end-to-end deep neural network based on cross-matching sample training effectively avoids the limitations of manual design algorithms and purifies the image PRNU noise. Then the performance of PRNU noise is improved for image source identification. The results of the comparative experiment on the Dresden image dataset show that the purified PRNU noise has better performance.