감정 인식을 위한 Saliency Map 기반 감정 우선 효과 통합 접근법

초록

Deep learning-based emotion recognition models, despite high performance, have limited reliability because the basis of their decisions is uninterpretable. To establish trust in these models, we must go beyond result interpretation and redesign their internal workings to be inherently explainable by emulating human cognitive processes. Human emotion recognition is closely linked to the visual system's attentional mechanisms. In particular, visual attention is preferentially drawn to emotional stimuli—an effect referred to as the emotion-prioritization effect. This suggests that key emotion-eliciting cues are concentrated in specific spatial regions that humans selectively focus on when recognizing emotions. By directly incorporating this human effect in the model's design, we can guide it to recognize emotions in a human-like manner. We propose the Emotion-Priority Saliency Module (EPSM), a saliency-map- based feature-map weighting module incorporating human visual attention. By integrating this module, we redesign the model to learn in a way consistent with the emotion-prioritization effect, thereby improving the explainability of the model's predictions.

키워드