Automated image quality assessment and its applications in computer vision

Abstract

The practical adoption of Convolutional Neural Networks (CNNs) in computer vision is widespread. However, CNN performance is heavily dependent on the perceptual quality of images. It is therefore necessary to monitor the quality of data input in order to verify that CNN predictions are reliable. While various algorithmic methods have been proposed to perform image quality prediction, the generalizability of these Image Quality Assessment (IQA) methods across diverse real-world scenarios is still the subject of ongoing research. Furthermore, few studies have used IQA to assess the robustness of CNNs to image quality degradations.

In this project, various learning-based IQA models from literature were implemented and benchmarked against multiple image domains. Among these models, we demonstrate the impressive ability of DeepBIQ, an approach that fuses the advantages of a CNN and a Support Vector Regressor (SVR), to predict the quality of unseen images following a mixed-dataset training approach. The model also shows promising results on real-world public surveillance footage. These results provide strong indications that IQA models can be implemented as part of practical image acquisition systems, to improve the reliability of downstream object detection and related pattern recognition pipelines.

In the second part of this project, we show that the quality predictions of DeepBIQ are closely correlated with object detector performance. This validates earlier studies investigating the relationship between perceptual quality and the accuracy of object detection systems. Our findings emphasize the importance of adapting object detectors to improve their robustness against image quality distortions.