Probabilistic Neural Network On Image Emotion Classification

Affective Computing is a computer technology that builds a human-computer harmonious environment by giving the computer a higher,more comprehensive and more intelligent ability.At the same time,computer must have a complete set of computer vision system for better emotional interaction with human beings.As the important form of the visual perception,the image emotion computing has become the important part of the affective computing.As a high-level semantic classification,image emotion classification begins with content-based image retrieval.Early image classification only based on low-level feature to create feature vector index,although the retrieval efficiency and accuracy of the image are improved,this kind of retrieval technique neglects the emotional elements and ignores the "semantic gap" between the low-level features and the high-level semantics,so the results are not ideal.Therefore,how to make up the "semantic gap" effectively and establish a reasonable emotional mapping mechanism is the key to image emotion classification.This paper presents a new emotion mapping mechanism based on the micro-blog or We Chat circle of friends and depth studies of the relevant image feature extraction techniques and classification algorithms which is based on a large amount of literature reading.First of all,by comparing the different image features of human emotions will have what kind of impact,combined with psychology,cognitive theory,selected from the image to reflect the characteristics of the most emotional elements,and to construct eigenvector space.Secondly,based on the research of text emotion algorithm,an improved learning incremental bayesian algorithm is proposed.In the sample sequence selection problem of the incremental bayesian learning,the algorithm introduces the knowledge of paired test and class support,respectively from the horizontal and the vertical angles make full use of prior knowledge to select the best subset to amend the classifiers.The algorithm solves the problem that the training set is small in scale,can not make full use of the prior knowledge and the noise data is transmitted continuously,and finally improves the accuracy of text emotion classification.Finally,by using the emotional elements extracted from the text as the emotion labels of the images,the mapping mechanism between the low-level features and the high-level semantics is established.Then the image emotion classification model is established by probabilistic neural network.The main work of this paper is how to establish the mapping relationship between the lowlevel features and the emotional space,and the relationship is based on the following facts: in the micro-blog or micro-circle friends,text and pictures can always express the same mood of the user at the time,therefore the emotional elements extracted from the complex images cleverly extracted from the text,so that not only reduces the difficulty of extraction,but also learn from the existing text emotion technology.