Facial expressions extraction from 3D sonography images

In this paper a novel approach for recognizing facial expressions from 3D sonography baby images is proposed. The posture of baby is uncertain in every image. It has been observed that sometimes face area is covered by some hand or leg portion. Identification of non-face part and removal of this portion is very essential. Preprocessing of image is essential to extract the face part from the image. A low intensity 3D sonography image is first preprocessed through various techniques such as histogram equalization, erosion and dilation. To extract the face part from the enhanced image, distance based clustering is applied. 3D sonography image may contain frontal or non-frontal face position because of the posture of baby during sonography. First level backpropagation neural network will classify the image in either frontal or non-frontal class with 97% of accuracy. Different feature extraction techniques have to be applied on classified frontal and non-frontal images. Local Binary Pattern (LBP) and template based matching techniques are used for feature extractions of frontal and non-frontal class of images respectively. The extracted features will be supplied to Minimum Redundancy and Maximum Relevance (MRMR) algorithm which will select the most promising N features. Extracted features are used to train the second level neural network and which classifies the image in any of the universal facial expression classes named: normal, happy and sad. Proposed approach is implemented in MATLAB and tested on about almost 500 images and significant amount of accuracy is obtained. Some anomalies like obesity, cleft lip or palate and birth-mark can also be found by the proposed approach.