Research On No-reference Image Quality Assessment Methods Via Deep Forest

No-reference image quality assessment(NR-IQA)is an important objective image quality assessment method.It is designed to assess the quality of the image without reference image information.How to propose an effective general-purpose NR-IQA method has been among the most challenging problems.In recent years,deep learning-based NR-IQA methods have performed well.Deep Forest is a decision tree-based deep learning model that can achieve the same performance as deep learning on many tasks.Based on this,we proposed two NR-IQA methods based on deep forest.1.We proposed a no-reference image quality assessment method for low-level feature distribution based on deep forest.This method first extracts the luminance and texture features under the normalization of the distorted image,and then uses the histogram distribution of the two types of features as quality perception features of the image,which can describe the luminance and texture distortion of the image.Second,the deep forest regression model is used to construct an NR-IQA model.Finally,the experiment is performed on five public databases.The experimental results show that the comprehensive performance of the proposed method is better than that of the comparative methods and has the best performance in LIVE and TID2013 databases with SROOC values of 0.974 and 0.890,respectively.2.We proposed a no-reference image quality assessment method for multi-level features distribution based on deep forest.Firstly,the pre-trained network is used to extract features of different levels,namely low-level,middle-level and high-level features,and then extract the fitting coefficient of the generalized Gaussian distribution of the three types of features as quality perception features.Secondly,deep forest is used to build regression model to evaluate the image quality.Finally,the experiment is compared with the state-of-the-art methods on six public databases.The results show that the proposed method performs well in the KADID-10k database,indicating that the proposed method is highly competitive for large image databases.