A deep learning approach to image quality assessment

Abstract

Deep convolutional neural networks (DCNNs) have an unchallengeable performance advantage over traditional machine learning in solving visual problems. However, DCNNs are vulnerable when the input signals are distorted or manipulated maliciously. We explore the computational modeling of image quality assessment (IQA), investigate the vulnerability of DCNNs, and utilize IQA principle to mitigate it. Firstly, works that pushing the performance limit of IQA modeling is discussed. Using transfer learning, we re-purpose off-the-shelf visual features for quality prediction. The recurrent neural network is added to distill global features. Experiments show that the proposed IQA models outperform or perform on par with counterparts in the literature. Subsequently, a fit-for-task detection framework is introduced. Through exploiting the correlation of visual characteristics between maliciously manipulated images and conventional quality degradation, the detector effectively protect DCNNs from producing wrong results in scenarios of benign quality degradation and malicious attack. Likewise, radiographs of inadequate quality could lead to false diagnosis by image analysis systems. The image quality in medical radiology includes many more aspects beyond pixels. Lastly, we present a method that helps diagnosis AI to recognize view types of chest radiographs.