Research And Development Of Assistant Decision System For Mobile Phone Image Quality Assessment

Image Quality Assessment(IQA)is to evaluate the ability of camera system restoring the true color and texture of real subjects.After analyzing the evaluation results,the image effect engineer optimizes the image processing algorithm and its parameters to improve image quality.In recent years,most IQA methods based on Convolution Neural Networks(CNN)have obtained high accuracy on existing datasets,but these methods still have not been deployed in actual projects because of their uninterpretability.In the field of engineering,IQA project relys on image quality experts with years of IQA experience.However,their assessment results are inevitably influenced by subjective factors and the manual process is low efficiency.So it is necessary to design an automatic system based on digital image processing and machine learning to assistant image quality expert,which can improve the efficiency and ensure the effectiveness by a variety of objective parameters.In the development of assistant decision system for mobile phone image quality assessment,following work has been done:(1)Analyze the general process of image quality evaluation projects and summarize the problems in engineering projects: the interference of experts' subjective emotions,the inconsistency of the results of different experts,long evaluation cycle and lack of image quality experts;(2)The framework of assistant decision system for mobile phone image quality assessment is designed according to the pipeline of the IQA project.The scene description,attention area detection,image quality assessment and image quality scoring module are used to simulate the evaluation process of image quality experts.Taking the standardized images as input,system can output IQA results to assist expert;(3)Functional design,technology selection and algorithm implementation for each module of the auxiliary decision system.The method includes: Scene Description Module,which uses image classification network to assign a scene description to each group of images;Region Segmentation Module: which uses semantic segmentation,face detection,and edge information to segment the original image into several blocks;Image Quality Assessment Module,applying digital image processing technology to evaluate image quality for five evaluation items;Image Quality Scoring Module,designing a set of statistical models that canlearn the weights of each area of interest and the thresholds of qualitative problems judgment by themselves.The single-level test results show that each module of the system can achieve the corresponding functions.The overall experimental results of the engineering dataset show that the system using self-learning parameters can obtain the IQA results similar to the experts',and can complete the task of assisting the experts.