Facial expression detection and recognition system
Abstract : In this paper, the integration of face feature detection and extraction, and facial expression recognition are discussed. In this paper, we propose an algorithm that utilizes multi-stage integral projection to extract facial features. Furthermore, in this project, we propose a statistical approach to process the optical flow data to obtain the overall value for the respective feature region in the face. This approach has eliminated the requirement of accurate identification of the feature boundary. Optical flow computations are utilized to identify the directions and the amount of motions in image sequences that are caused by human facial expressions. The optical flow computation results are processed using Kalman filtering. The filtered results are given to a neural network to realize a mapping into the facial expression space. This technique is used on a set of training and testing face images. Preliminary experiments indicate an accuracy between 60% - 80% on the Kalman filtered data when recognizing four types of expressions: anger, sad, happy and surprise. In an attempt to further improve the recognition results, we proposed a technique to process the optical flow results using a statistical approach instead of using Kalman filtering. The preliminary experiments on this proposal approach produced accuracy between 70% - 100% on the original optical flow results that is better than the Kalman filter technique.
? Therefore, for the separation of the datasets in training, validation and test data, the divisions were made in such a way that images of a specific person only existed in one of the datasets.
? This unintentional information is the primary way to know the transmitter's emotions in a non-invasive way.
? An algorithm would subsequently count the classifications within a time window.
? If the number of anger and disgust classifications exceeds a specific threshold, the driver would be considered under stress.
? The best classifier obtained was trained, not only with the images from the public datasets but also with images of the subjects posing for the stress classification.
? The integral projection technique was originally proposed by Kanade and modified by the authors to suit the problem of face position, size estimation and feature location.
? For certain types of problems, such as learning to interpret complex real-world sensor data, artificial neural networks are among the most effective learning method currently known.
? The back propagation algorithm that has proven successful in many practical problems is implemented here.
? Detection of face features such as eyes and mouth have been major issues of facial image processing which may be required for various areas such as emotion recognition and face identification.
• Our proposed method makes use of edges extracted from the facial image that are considered as general features during our feature detection process.
• Computer-based recognition of facial expressions goes a long way, and various methods have been proposed.
• In this work, we propose a method to combine feature detection and extraction and facial expression recognition into an integrated system so that the recognition results will not be influenced by subjective factors and the bound of areas are invariant during the whole sequence.
• We propose a method for facial expression recognition that uses integral projection, statistical computation, a neural network and kalman filtering.
? One advantage of optical flow is that optical flow information can be extracted easily even at low contrast edges.
? One advantage of this step is that the threshold is based on the statistics of the image data instead of manual value and is acceptable to small amount of noise.
? The main advantage is that this is a simple algorithm that is very fast.
? The main advantage of this approach is that it does not require any initial manual settings such as location of head.
? The performance of these approaches is affected by several external conditions such as illumination and skin color.
? The initial settings are predetermined using normalized coefficients obtained using a facial image database.
|