Robotic manipulation of deformable objects such as a towel is one of the most challenging tasks in the field of service robotics. Their unpredictable shape and pose makes it difficult to identify the most relevant parts than can be used for grasping. In this report we design a deep neural network that finds all the visible corners of a wrinkled towel. We also create a dataset to train the network. The results obtained show that it is possible to detect the corners of a cloth, but with some constraints due to the small number of images inside the dataset.