Depth data is a crucial source of information for applications like autonomous vehicles and 3D cinema. However, depth maps often contain holes or low-confidence regions, which must be addressed to ensure optimal performance in tasks like object avoidance and path planning. This paper proposes a convolutional neural network (CNN) to replace the feature extraction stage of a hybrid depth completion model. The hybrid model integrates convolutional stages with an infinity Laplacian-based interpolator to propagate sparse depth data into missing regions.