Research On Few Shot Learning For No Reference Image Quality Assessment

No-reference image quality assessment(NR-IQA)attempts to design a reasonable computational model which can automatically estimate the perceptual quality of a given image without any reference image.For the current NR-IQA algorithm,it faces the dilemma of insufficient training data,which has caused the model capacity to be greatly limited and cannot further improve the prediction accuracy.This shows that finding an effective few-shot learning method has become the main goal of the current NR-IQA problem.In order to solve the above problems,this paper explores how to build a NR-IQA model based on the few shot learning.Specifically,this paper discussed the construction of NR-IQA models from three aspects,they are NR-IQA based on weakly supervised learning,NR-IQA based on Bayesian learning and NR-IQA based on meta learning respectively.The main content is summarized as follows.In the first part,we propose a NR-IQA model based on the weakly supervised learning.In detail,we pre-trained a novel encoder-decoder architecture by using the training data with weak quality annotations.The annotation is the error map between the distorted image and its undistorted version,which can roughly describe the distribution of distortion and can be easily acquired for training.Next,we fine-tuned the pre-trained encoder on the quality labeled data set.Moreover,we used the group convolution to reduce the parameters of the proposed metric and further reduce the risk of over-fitting.These training strategies,which reducing the risk of over-fitting,enable us to build a very deep neural network for BIQA to have a better performance.Experimental results showed that the proposed model had the state-of-art performance for various images with different distortion types.In the second part,we propose a NR-IQA model based on the probabilistic graph prior relationship modeling.This algorithm uses the phenomenon of context effects in human neuropsychological activity as a prior to guide the establishment of a NR-IQA model.At first,we use a graphical model to describe how the context effect influences human perception of image quality.Based on the established graph,we construct the context relation between the distorted image and the background environment by the Matchnet^[75].Then the context features are extracted from the constructed relation,and the quality-related features are extracted by the fine-tuned neural network from the distorted image in pixel-wise.Finally,these features are concatenated to quantify image quality degradations and then regress to quality scores.In addition,the proposed method is adaptive to various deep neural networks.Experimental results show that the proposed method not only has the state-of-art performance on the synthetic distorted images,but also has a great improvement on the authentic distorted images.In the third part,we propose a NR-IQA model based on the meta-learning.At first,we follow the general scheme of meta learning to construct several pairs of support sets and meta-test sets as the training data of proposed model.Then,we pre-trained a classic deep neural network as a quality-related feature extractor by using the public IQA data set.In each training iteration of the proposed model,the pre-trained quality feature extractor converts the images in the support set and meta-test set into corresponding quality-related features.Next,we use the LSTM to embed the obtained features into a context feature sequence,and use a distance measurement module to calculate the similarity between the embedded context features of support set and meta-test set.Based on the obtained similarity,we calculate a context quality score of the images in the meta-test set.At last,we directly map the original quality features of the image in the meta-test set to its pixel-level quality score,and then weights and sums the context quality score and pixel-level quality score as the final quality prediction.Experimental results show that the proposed model has high evaluation accuracy for various types of distorted images under the condition of limited labeled data.