We consider a method for reconstructing the original HDR image from a single LDR image suffering from saturation for metallic objects. A deep neural network approach is adopted for directly mapping from 8-bit LDR image to an HDR image. An HDR image database is first constructed using a large number of objects with different shapes and made of various metal materials. Each captured HDR image is clipped to create a set of 8-bit LDR images. The whole pairs of HDR and LDR images are separated and used to train and test the network. Next, we design a deep CNN in the form of a deep auto-encoder architecture. The network was also equipped with skip connections to keep high image resolution. The CNN algorithm is constructed using MATLAB's machine-learning functions. The entire network consists of 32 layers and 85,900 learnable parameters. The performances of the proposed method are examined in experiments using a test image set. We also compare our method with other methods. It is confirmed that our method is significantly superior in reconstruction accuracy and the good histogram fitting.