Badger Identification using Handcrafted Image Matching with Learned Convolutional Filter

A new image matching framework is introduced and applied to the challenging task of badger identification by matching facial characteristics of individual badgers. A novel filter design based on a shallow convolutional neural network for prefiltering images to improve the image matching accuracy is presented. Hill climbing, a commonly-used search optimization algorithm, is used to train this shallow and computationally efficient convolutional network to be deployed at the early stage of an image matching pipeline. The contributions of this work include a novel proposed technique for prefiltering the images using a shallow CNN (Convolutional Neural Network) and applying the filter to the fusion of two handcrafted descriptor algorithms, SIFT (Scale-Invariant Feature Transform) and BRISK (Binary Robust Invariant Scalable Keypoint). Our various combination of these two descriptors achieves a higher F-score than the respective baseline algorithms.