Sentiment Interaction Distillation Network for Image Sentiment Analysis

Sentiment is a high-level abstraction, and it challenging task to accurately extract sentimental features from visual contents due the “affective gap”. Previous works focus on extracting more concrete of individual objects by introducing saliency detection or instance segmentation into their models, neglecting interaction among objects. Inspired observation that can impact sentiment images, we propose Interaction Distillation (SID) Network, which utilizes object guide feature learning. Specifically, first utilize panoptic method obtain in images; then, sentiment-related edge generation employ Graph Convolution Network aggregate propagate relation representation. In addition, knowledge distillation framework information guiding global context learning, avoid noisy introduced error propagation varying number Experimental results show our outperforms state-of-the-art algorithm, e.g., about 1.2% improvement Flickr dataset 1.7% most subset Twitter I. It demonstrated reasonable use improve performance analysis.