Ensembling Convolutional Neural Networks for Perceptual Image Quality Assessment

Perceptual Image quality assessment is a challenging problem especially in the absence of reference information. No-reference quality assessment is required for a number of applications such as quality assessment of image acquisition, enhancement and communication scenarios. Conventionally the problem is addressed by extracting natural scene statistics but recent development of deep learning has paved the way of deep learning based methods. Convolutional Neural Networks (CNN) has shown surprising performance for the task of visual classification but they have some inherent limitations such as high computational requirements, limitations of scalability and model variance. Ensemble learning methods are used to improve the generalization performance of machine learning methods but their application to CNN is limited due to their already high computational requirements. We have proposed an approach to train a single CNN model with a learning rate scheduler and save its training states at regular intervals. These saved model states are treated as base models and some of them are selected to construct ensemble with weighted averaging. The proposed methods has provided promising results and indicate its utility for training of advanced architectures for ensemble learning.