Ghost imaging (GI) has been paid attention gradually because of its lens-less imaging capability, turbulence-free imaging and high detection sensitivity. However, low image quality and slow imaging speed restrict the application process of GI. In this paper, we propose a improved GI method based on Denoising Convolutional Neural Networks (DnCNN). Inspired by the corresponding between input (noisy image) and output (residual image) in DnCNN, we construct the mapping between speckles sequence and the corresponding noise distribution in GI through training. Then, the same speckles sequence is employed to illuminate unknown targets, and a de-noising target image will be obtained. The proposed method can be regarded as a general method for GI. Under two sampling rates, extensive experiments are carried out to compare with traditional GI method (basic correlation and compressed sensing) and DnCNN method on three data sets. Moreover, we set up a physical GI experiment system to verify the proposed method. The results show that the proposed method achieves promising performance.