|
Visual sentiment classification is one of the popular research area in affective image content analysis (AICA). At present, most of the recent methods focus on utilizing only the last layer output of the convolutional neural network, which is suboptimal in this area. Various research work on psychology theory, art theory and hand-crafted feature design denote that low-level and mid-level features are also essential to visual sentiment classification. In this work, we focus instead on the low-level and mid-level features and propose a novel architectural unit, which we term the "Multilevel Feature Extracting Structure" (MFES) block, that adaptively utilizing the multilevel feature outputted by the base convolutional neural network. Based on the proposed module, our model achieves the accuracy of 67.39%, surpassing the original accuracy of 66.16% by 1.23%. |
|
Keywords:Visiual Sentiment Classification; Deep Learning; Neural Networks; Computer Vision. |
|