Key Technology Research About Micro-Expression Of Detection And Location

Micro-express ion is a kind of behavior that is not painstakingly controlled by human and spontaneously expresses human inner true feelings. Extracting effective non-verbal clues from this brief and detailed real reaction can accurately insight into the human psychological activity, which will require multiple research areas such as clinical psychology and computer science. It has practical application significance for clinical medicine, national security, judicial proceedings, and other fields. This paper focuses on the key technical research of micro-express ion detection and localization. The main research contents include the following three aspects:(1) Based on the research analysis of traditional local space-time interest detection algorithm, a novel approach named Mo SIFT algorithm is fist proposed to extract facial interest points under the word bag model, which does not need to locate interests in advance. It directly detects local space-time feature interest points in the video streams and at the same time, histogram of oriented gradient and histogram of oriented flow are adopted as feature descriptors that include space-time motion information.(2) In order to obtain discriminative representation, all expression features are clustered with the k-means method to learn a visual dictionary. Then, to overcome the shortage of bigger reconstruction error from traditional vector encoding method, simulation orthogonal matching pursuit is applied to map each feature vector into a certain code word. After obtaining the sparse representation of micro-expression features, the K-nearest neighbor and support vector regression are respectively used to test and analysis on two data sets.(3) On the CK+ database, K-nearest neighbor algorithm is used to classify four typical expressions. Contrast the macroscopic visual testing points and the recognition rate between STIP and Mo SIFT. Two kinds of detection methods and the results show that Mo SIFT steadily detects more key points on main facial action units, which will provide enough information for subsequent dictionary learning and character description information. At the same time, experiment has achieved an average of 86.35% correct recognition rate. Based on the results of the experiment in the database CK+ and at the perspective of the emotional state, the support vector regression method is used to build prediction model under four dimensions of emotional state values-arousal (how dynamic or lethargic the person appears to feel), expectation (how surprised the person appears to be), power (how much control the person when he affected by the outside world degree), valence (positive or negative features),and get the correct value changes of four emotional status in the prediction mode.