Efficient binary segmentation through dense neural networks in a truncated frequency domain

This article presents a method for binary segmentation of any type of tensor given that a dataset of such tensors with ground truth segmentations is available. The proposed method compresses input tensors through the use of the Discrete Cosine Transform (DCT) followed by a truncation of the resulting spectrums. After the compression step, a shallow dense neural network performs the segmentation entirely in the frequency domain. The method is evaluated on a common robotics environment model known as an occupancy grid map. Results exhibit a correct segmentation for an especially small computational time of 2.16 ms for the largest neural network. Moreover, the computational requirements are freely configurable by the choice of the compression factor making such a method interesting for highly constrained hardware platforms found in embedded setups.