RARITYNet: Rarity Guided Affective Emotion Learning Framework

TL;DR

RARITYNet is a novel framework that combines handcrafted RARITY features with deep learning to improve facial expression recognition under challenging conditions like pose, ethnicity, and illumination.

Contribution

It introduces a hybrid model integrating RARITY features with a deep network, enhancing emotion recognition from difficult facial images.

Findings

Improved accuracy in recognizing spontaneous expressions.

Robustness against pose and illumination variations.

Effective integration of handcrafted and deep features.

Abstract

Inspired from the assets of handcrafted and deep learning approaches, we proposed a RARITYNet: RARITY guided affective emotion learning framework to learn the appearance features and identify the emotion class of facial expressions. The RARITYNet framework is designed by combining the shallow (RARITY) and deep (AffEmoNet) features to recognize the facial expressions from challenging images as spontaneous expressions, pose variations, ethnicity changes, and illumination conditions. The RARITY is proposed to encode the inter-radial transitional patterns in the local neighbourhood. The AffEmoNet: affective emotion learning network is proposed by incorporating three feature streams: high boost edge filtering (HBSEF) stream, to extract the edge information of highly affected facial expressive regions, multi-scale sophisticated edge cumulative (MSSEC) stream is to learns the sophisticated edge information from multi-receptive fields and RARITY uplift complementary context feature (RUCCF) stream refines the RARITY-encoded features and aid the MSSEC stream features to enrich the learning ability of RARITYNet.