光谱成像技术研究室知识库

We aim to predict the sentiment related information reflected in images based on SentiBank, which is a library including Adjective Noun Pair (ANP) concept detectors for image sentiment analysis. Instead of using only ANP responses in images as mid-level features, we make full use of the ANPs' textual sentiment. We first give each ANP concept in SentiBank a sentiment value by adding together the textual sentiment value of the adjective and that of the noun. Having detected the presence of ANPs in an image, we define an image sentiment value by computing the weighted sum of the textual sentiment values of ANPs describing this image with corresponding ANP responses as weights. On the one hand, we adopt a one-dimension logistic regression model to predict the sentiment orientation according to the image sentiment value. On the other hand, we use the ANP responses detected in an image as mid-level representations to train a regularized logistic regression classifier for sentiment prediction. We finally employ a late fusion algorithm to combine the prediction results from the two schemes. Experimental results reveal that the proposed method which takes into account the textual sentiment of ANPs improves the performance of SentiBank based image sentiment prediction.