Investigating biological feature detectors in simple pattern recognition towards complex saliency prediction tasks

• Main Author: Cordel, Macario O., II
• Format: text
• Language: English
• Published: Animo Repository 2018
• Subjects: Computer vision; Pattern perception; Pattern recognition systems; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_doctoral/551
• Institution: De La Salle University

The conventional convolution filter in deep architectures has proven its capability to extract semantic information from the input images and to use these in different visual tasks. For many researchers in computer vision, this raises the question, have pattern recognition models begun to converge on human performance? This thesis explores a new biologically-inspired feature detector for pattern recognition which learns via competition. We describe and exhaustively characterize our proposed alternative feature detector and compare this with the traditional convolution filter feature detector. Our experiments show the potential of the proposed feature detector and that its performance is at par with the performance of the convolution filter. Using the feature detector with more desirable result, we then design and propose a computational model for one of the primitive pattern recognition tasks of the visual system, the saliency map generation. The study provides a methodology for quantifying the contribution of the convolution filter in simple pattern recognition tasks and use this to benchmark our proposed competition-based feature detectors. Towards achieving an improved computational model for a complex prediction task of visual systems, we further use the biological feature detectors in extracting and incorporating emotion-evoking objects in saliency prediction.