Research On Image Emotion Categorization

The purpose of emotional image classification is that we hope the computer canexpress the emotion reaction when observing the image, and classify the images into thedifferent emotional categories automatically. Based on the key questions of emotionalimage semantic analysis, such as image feature extracting, generating visual dictionary,feature mapping, image description and so on. We should propose novel method andestablish the relationship between image feature and emotion mapping effectively. Soimprove the emotion image classification recognition rate and make the ability ofcomputer understand image close to people’s level. This must play a promoting role in thedeveloping of image classification and image retrieval research. In this paper, we will getdown to research follow:Firstly, considering the influence of the color information to emotion semanticanalysis. We introduced emotional color feature based on Scale Invariant FeatureTransform. Extract the SIFT feature from three color channels and string R-SIFT, G-SIFT,B-SIFT to form a348dimensions C-SIFT. Then draw Spatial Pyramid Matching, dividingthe image into more and more fine areas. Calculating the visual words’ distribution undermultiple space resolutions to form the spatial pyramid presentation. This method used thespace structure information of the images effective. Due to added the color informationand spatial information on the basis of classical algorithm, we have achieved betterexperiment results.Secondly, in view of the shortage of traditional visual vocabulary, which is formed byindependent local visual feature and overlook the relationship of adjacent image features.We build a visual characteristics of context information contains multiple directions.Because the big dimension and large number of contextual feature, we classified thefeature before image description. Only use edge feature to form visual dictionary.Thosemethods not only decrease the calculation, but also improve the distinguish of visualdictionary and image description.At last, to make the computer conform to human visual perception mechanism. Inthis paper, we use the image emotion classification based on visual attention mechanism. We combined the Itti-Koch model and Frequency-tuned model to form the image saliencymap. Then classified the feature and formed different types of visual dictionary. Eachimage is described by using LLC (Local-constrained Linear Coding) scheme, imagerepresentations were performed by the methodology of spatial pyramid. Finally, weperformed the emotional categorization by the training classifier.The improvements have been made on three aspects in this paper. We set IAPS(International Affective Picture System) and KDEF (Karolinska Directed Emotional Faces)as our experimental datasets and apply Linear-SVM with HIK(histogram intersectionkernel) to complete emotion categorization. The ideal experimental results illustrate thefine effectiveness of our approach.