Abstract: Human facial emotion recognition (FER) has attracted the eye of the research network for its promising applications. Mapping one of a kind facial expressions to the respective emotional states are the primary task in FER. The classical FER consists of two most important steps: feature extraction and emotion recognition. presently, the Deep Neural Networks, particularly the Convolutional Neural network (CNN), is extensively used in FER with the aid of distinctive feature of its inherent feature extraction mechanism from pictures. numerous works were reported on CNN with only some layers to clear up FER issues. but, wellknown shallow CNNs with straightforward getting to know schemes have restricted characteristic extraction capability to seize emotion data from high-resolution pictures. A notable disadvantage of the most current techniques is that they consider only the frontal pictures (i.e., ignore profile perspectives for convenience), despite the fact that the profile perspectives taken from different angles are essential for a practical FER system. For growing a highly correct FER system, this study proposes a completely Deep CNN (DCNN) modeling thru transfer learning (TL) technique wherein a pre-skilled DCNN model is followed through changing its dense top layer(s) well suited with FER, and the model is great-tuned with facial emotion data. a novel pipeline strategy is brought, wherein the training of the dense layer(s) is accompanied via tuning each of the pre-skilled DCNN blocks successively that has brought about gradual improvement of the accuracy of FER to a better level.

Keywords: convolutional neural network (CNN); deep CNN; emotion recognition; transfer learning

| DOI: 10.17148/IARJSET.2021.8756