基于多任务无参考图像质量评价模型研究

基于深度学习的无参考图像质量评价(NRIQA)模型常见2种结构，即单任务(Single-task)结构和多任务(Multi-task)结构。为了探讨在没有预训练情况下多任务结构对模型准确率影响，对比分析了基于MEON调整后的多任务模型及单任务模型在无参考图像质量评价任务上的性能优劣,其中多任务模型在图像质量评价数据库LIVE、TID2013上分别取得了0.882、0.871的准确率，表现出同等甚至优于单任务模型的性能。在此基础上，多任务模型的子任务输出维度实验表明在NRIQA研究中，子任务能够根据需求和目标在相关数据集上预训练，再结合质量评价任务微调，具有可迁移学习集成于其他任务中的优点。

Non-reference Image Quality Assessment Models Based on Multi-task

The Non-Reference Image Quality Assessment (NRIQA) models based on Deep Learning(DL) include two common structures, namely the single-task structure and the multi-task structure. In order to study the effect of the multi-task structure on model accuracy without pre-training, the performance of the MEON-adjusted multi-task model and the single-task model in the NRIQA task is compared and analyzed. The accuracy rate of the multi-task model reaches 0.882 and 0.871 respectively on LIVE and TID2013 image quality assessment databases. Experiments show that the multi-task model without pre-training still exhibits equal even better performance than the single-task model. On this basis, the sub-task output dimension experiment of the multi-task model shows that in NRIQA research, the sub-task can be pre-trained on the relevant datasets according to the demand and goal, and then fine-tuned in combination with the quality evaluation task, it has the advantage of being able to integrate into other tasks by transferring learning.