This paper presents an estimation method for emotion evoked by watching images focused on non-linear correlations of between gaze and visual features. The proposed method estimates the emotion using novel visual features obtained from original visual features projected to space that is maximized correlations of between gaze and visual features by Kernel Discriminative Locality Preserving Canonical Correlation Analysis (KDLPCCA). In the proposed method, the realization of accurate emotion estimation can be expected, since KDLPCCA performs the calculation of the novel visual features considering non-linear correlation, class information, and locality.