Relating orientation tuning and feature utilization during facial expression recognition.

Facial expression recognition (Huynh & Balas, 2014) correlates with horizontal information utilization, and it also correlates with local information utilization (e.g. mouth; Blais et al., 2012). However, the link between these two aspects of visual processing remains elusive. Our aim was to provide a first examination of this link, and refine current knowledge on orientation tuning. Twenty participants each performed 1,400 trials in a facial expression recognition task using 7 expressions. Seventy pictures of faces (10 identities; face width= 4deg) depicting the six basic emotions plus neutrality were used as stimuli. Images were randomly filtered in the orientation domain (orientation bubbles) and presented for 150ms. We developed this method to allow precise extraction of orientation utilization and exclude adaptation to predefined filters as a possible confound. Classification Images (CIs) were derived with a weighted sum of orientation samples, using Z-transformed accuracy scores as weights. A Pixel Test (Chauvin et al., 2005) was used on Z-transformed CIs (ZCIs) to establish statistical significance. Horizontal information significantly correlates with performance (Z-obs>Zcrit= 1.89) for anger [0°-6°; 173-180°], sadness [0°-3°; 176°-180°], disgust [0°-9°; 175°-180°], fear [0-10°; 176°-180°], happiness [0°-8°; 175°-180°], and neutrality [0°-10°; 175°-180°]. For surprise, vertical/oblique [64°-76°] information is significant. Participants also completed the same task, but with location randomly sampled (location bubbles; Gosselin & Schyns, 2001). The link between orientation tuning and location tuning was examined by regressing orientation ZCIs with location ZCIs. Interestingly, individuals more horizontally tuned rely significantly (Z-obs>Zcrit= 3.57) more on diagnostic locations (eyebrow junction for anger, left eye and mouth for fear, and mouth for all others) than those less tuned. Differences weren't reliable at other angles. Our results therefore imply that horizontally tuned individuals are more efficient at extracting local diagnostic information from faces. Implications for face processing will be discussed. Meeting abstract presented at VSS 2015.