Computer Vision News - January 2018

1. The PreProc (preprocessing) uses a bank of Gabor-like filters to extract edges from pixel values. It is structured as a pipeline consisting of 4 steps, which we will go into now. 1.1 2D Correlation - A predefined set of grayscale edge filters using 16 edge orientations, where the filters consist of a positive and a negative Gaussian. 1.2 Localization - A variant of the Canny edge detection is used to prevent replacement of the evidence corresponding to a single edge. 1.3 Cross-channel suppression - Cross-channel suppression is computed (using Fpooled - for the exact equation see article) to overcome the risk of minor noise causing an image edge halfway between two filter orientations being suppressed. 1.4 Conversion to log likelihoods - Normalizes the maximum brightness, such that edges with high contrast do not have a higher score, while edges with very low contrast will have a brightness proportional to the score of the filter. 2. Learning a hierarchical model from the preprocessed data. The algorithm constructs the RCN representation layer-by-layer, from the bottom up. The model needs to represent the learning of two components: a) features and (b) lateral connections. Computer Vision News Research 7 Research