This paper presents a method for reconstructing the original high dynamic range (HDR) image from a saturated low dynamic range (LDR) image with missing physical information, specifically for single dielectric objects. A deep neural network approach is employed to map an 8-bit LDR image directly to its corresponding HDR representation. We begin by analyzing the reflection and saturation characteristics of dielectric materials and then construct an HDR image database using a diverse set of dielectric objects. Each HDR image is clipped to generate a set of 8-bit LDR images. All HDR-LDR image pairs are normalized to a fixed resolution and used for training and validation. A deep convolutional neural network (CNN) is designed in the form of an autoencoder architecture with skip connections. The entire network is implemented using MATLAB’s machine learning toolbox, with the ADAM optimizer employed for training. The performance of the proposed method is evaluated using a separate validation set. Comparative experiments with existing methods demonstrate that our approach achieves significantly higher reconstruction accuracy and better histogram fitting.