In this paper, we propose an efficient method for learning local image descriptors suitable for the use in image inpainting algorithms. We learn the descriptors using a convolutional autoencoder network that we design such that the network produces a computationally efficient extraction of patch descriptors through an intermediate image representation. This approach saves computational memory and time in comparison to existing methods when used with algorithms that require patch search and matching within a single image. We show these benefits by integrating our descriptor into an inpainting algorithm and comparing it to the existing autoencoder-based descriptor. We also show results indicating that our descriptor improves the robustness to missing areas of the patches.