Automatic Fetal Facial Expression Recognition by Hybridizing Saliency Maps with Recurrent Neural Network

Fetus facial expression analysis is a recent and upcoming field of study in the area of biomedical image processing. Fetus images are obtained using 3D ultra-sounds, and thus there is minimum clarity in terms of the fetus face alignment, the fetus face posture and the fetus face size. All these issues make it a challenging task to identify the location of fetus face, and thus the fetal expression or mood analysis becomes a complicated task. In this paper, a saliency map based method is proposed to segment out the fetus face with good level of accuracy, and then identify the fetus mood using a recurrent neural network based classifier. Our work shows more than 80% accuracy across various fetus aging images, and has moderate delay of classification. We also proposed techniques for improving the accuracy further and also improving the precision and recall rates for the classification process.