No-reference image quality assessment based on automatic machine learning

Abstract

In different applications in deep learning, due to different required features, it is necessary to design specialized Neural Network structure. However, the design of the structure largely depends on the relevant subject knowledge of researchers and lots of experiments, resulting in huge waste of manpower. Therefore, in the field of Image Quality Assessment (IQA), the authors propose a method to apply Neural Architecture Search (NAS) to IQA. Mainly through the Differentiable Architecture Search algorithm, the structure of the modular Neural Network unit is searched by the stochastic gradient descent algorithm with better training performance by relaxing the operation features into a continuous space. Also, the idea of weight sharing is used to further save. The authors use the mainstream IQA database LIVE to search for Neural Network structures, and retrain and validate the searched structures in four datasets. A large number of experiments show that the model obtained by the search experiment achieves the effect of the best algorithm at this stage, and has a certain quality. The main contributions of this paper are: Transform the DARTS algorithm to adapt the regression problem, and introduce the Neural Architecture Search algorithm into the IQA field and conduct experimental verification.