Anger and Stress Recognition from Spectrogram Images Using Pre-Trained Convolutional Neural Network

Abstract

The convolutional neural Network (CNN) has recently shown exceptionally good performance in image classification with a little amount of pre-processing compared to other classification algorithms. Building CNN from a scratch is not only an expensive and complex task but also requires expertise. Pre-trained CNN reduces efforts of creating models from scratch as they are earlier trained on the ImageNet dataset and could work well with the little training set. In the proposed work CNN is used for anger and stress classification through analyzing discriminatory patterns in spectrogram images. In the proposed work performance of pre-trained CNN ResNet50, SqueezeNet, WideResNet, and InceptionV3 are compared in identifying emotions (anger, stress, neutral). Pre- trained Network ResNet50 has shown the highest performance with an accuracy of 99.7%, followed by WideResNet, SqueezeNet, and InceptionV3 with accuracy 85.7%, 57.8%, and 38.5% respectively.