Research On Kansei Semantic Image Retrieval

Existing studies have suggested that different images induce different emotions, but the traditional technology of image retrieval and classification mostly retrieves images by analyzing the similarity of image visual features, however, it neglects the impact and function of image emotion, which can not meet the user's real needs, classifying images rationally using image kansei semantic will be an important and challenge topics in the field of image retrieval.This dissertation focuses on the kansei-based image retrieval, and take the natural scene and wood images as the data source, and have researched the important technologies and algorithms on the kansei semantic retrieval of these two data source respectively.Color is the most important feature of the image, and also the feature that can evoke the emotional changes. In this paper, we make the color as the feature of the natural scene image database, and established the mapping relations with the emotional characters. In accordance with the relationship between the properties of color and human visual perception, we take the appropriate way to quantify the color components respectively in HSV color space, and make those quantified components as the global color characteristics; according to the human's visual attention mechanism, in this paper, we genetrante the image saliency map based on the characteristics of brightness and color, and used the golden section theory to extract the color characteristic which in the main region of the human visual center in the saliency map, we make this as the local feature of the image. we take all the two parts as the emotional characteristics of the natural scene images.Texture, as the other part of most important visual feature, is also one of the part to the human visual understanding. So in this paper, we selected the texture-based object—wood images as the data source, we analysis the unique texture characteristics on the wood images, and extract texture's direction, coarseness, intensity and contrast form the wood images, according to the relationship between the color feature of the wood image and the emotion, we extract the color characteristic of the wood image through L*a*b* color space, through the wood's color and texture feature, we build the emotional feature space about "gorgeous"and "simple" for wood images.For the Back Propagation neural network has a strong ability of classification, in this paper, we take the Back Propagation neural network to build the mapping relation about natural image and wood image between low features and kansei semantics effectively, and the experiment proves that we get a better retrieval results.