Image Quality Assessment For Smartphone Based On Lightweight Neural Network

Images taken by smartphones are large and contain mixed distortions,which bring some challenges to the image quality assessment task on smartphone terminal.This thesis focuses on the quality assessment of images taken by smartphones.The main contents include image quality assessment model based on neural network,lightweight methods of the model and the image quality assessment system on smartphone.In view of the characteristics of images taken by smartphones,this thesis constructed a MultiScale and Multi-Task Image Quality Assessment Model(MSMT IQA Model)based on multi-scale feature fusion and multi-task learning.In this model,the backbone supports multi-scale feature fusion.Classification task and comparison task are used as auxiliary tasks to improve the learning ability of the main task.Results show that the proposed model has a Spearman’s Rank Correlation Coefficient(SRCC)of 91.23%,which is better than other traditional machine learning and deep learning methods.Ablation experiments further show that the feature fusion method and multi-task learning mechanism proposed in this thesis improve the performance of the model without increasing the network depth.SRCC is increased by 0.69%.Considering the limited computing resources and storage space of mobile phones,this thesis adopts a series of lightweight operations for the neural network model mentioned above.The operations include training lightweight network based on knowledge distillation,simplifying computing steps based on module fusion and compressing model based on INT8 quantization.Experiments show that the model size is reduced to 1/36,the throughput is increased by about 30 times,and the SRCC is 88.80%.In the stage of system realization,this thesis optimizes the strategy of region of interest selection based on rule of thirds in photography and represents the image quality with fewer image blocks to save the overall inference time.Experiments show that the throughput after the strategy optimization is 6 times that before the optimization,and the SRCC only decreases by 0.06%.