The binarization step for old documents is still a challenging task even though many hand-engineered and deep learning algorithms have been offered. In this research work, we address foreground and background segmentation using a convolutional autoencoder network with 3 supporting components. The assessment of several hyper-parameters including the window size, the number of convolution layers, the kernel size, the number of filters as well as the number of encoder-decoder layers on the network is conducted. In addition, the skip connections approach is considered in the decoding procedure. Moreover, we evaluated the summation and concatenation function before the up-sampling process to reuse the previous low-level feature maps and to enrich the decoded representation. Based on several experiments, we determined that kernel size, the number of filters, and the number of encoder-decoder blocks have a little impact in term of binarization performance. While the window size and multiple convolutional layers are more impactful than other hyper-parameters. However, they require more storage and may increase computation costs. Moreover, a careful embedding of batch normalization and dropout layers also provides a contribution to handle overfitting in the deep learning model. Overall, the multiple convolutional autoencoder network with skip connection successfully enhances the binarization accuracy on old Sundanese palm leaf manuscripts compared to preceding state of the art methods.