FINET: Facial Expression Recognition Based on Fusion Inherited Network

Abstract

Machine learning models built from complex hand-crafted features and classification processes are challenging to design and arenotrobustized. Due to this fact, convolutional architecture is incorporated to automatically extract and classify class labels with high levels of accuracy. In this work we propose a lightweight network FiNet: fusion inherited network for universal facial expression recognition. FiNet consist of two fusion bocks for contextual feature extraction from salient regions. Block-1 uses a single convolution filter to capture and preserve the domain features. Block-2 dispenses unwanted features from the enriched inherited features by employing a single filter. The inherited results of block one and two are fused into block 3 which merge all the comprehensive region features for classification by discriminating for higher adaptability. The two-stage fusion network significantly allows the network to only conserve spatial features which increases the discriminative power of FiNet. Space computation complexity is achieved by limiting the number of parameters and incorporating smaller database size which proves the reliability of FiNet. FiNet size is 3.6MB as compared to VGG16:500.5MB, and ResNet 88.4MB. Comparative analysis on ResNet and VGG16 proved that the results of our proposed model outperformed existing futuristic models. The recognition accuracy results of our model on JAFFE and CK+ database was 84.37% and 96.62% respectively. The proposed model provides higher adaptability to lower computational power and storage as compared to traditional systems. FiNet size is 3.6MB compared to VGG16:500.5MB, and ResNet 88.4MB. FiNet provides an opportunity to be deployed in smart gadgets.