Study on predicting sentiment from images using categorical and sentimental keyword-based image retrieval

Visual stimuli are the most sensitive stimulus to affect human sentiments. Many researches have attempted to find the relationship between visual elements in images and sentimental elements using statistical approaches. In many cases, the range of sentiment that affects humans varies with image categories, such as landscapes, portraits, sports, and still life. Therefore, to enhance the performance of sentiment prediction, an individual prediction model must be established for each image category. However, collecting much ground truth sentiment data is one of the obstacles encountered by studies on this field. In this paper, we propose an approach that acquires a training data set for category classification and predicting sentiments from images. Using this approach, we collect a training data set and establish a predictor for sentiments from images. First, we estimate the image category from a given image, and then we predict the sentiment as coordinates on the arousal–valence space using the predictor of an estimated category. We show that the performance of our approach approximates performance using ground truth data. Based on our experiments, we argue that our approach, which utilizes big data on the web as the training set for predicting content sentiment, is useful for practical purposes.