Image Quality Distortion Classification Using Vision Transformer

In this paper, we propose a method for classifying image quality distortions to identify common types of distortions typically present in images, utilizing a vision transformer. The method aims to enhance quality-related image processing approaches by identifying specific distortions as the initial step in distortion-based blind image quality assessment (BIQA). This simplifies the quality reconstruction process by tailoring it to the prior knowledge of distortion types, thereby aiding in improving image classification and potentially reducing biases caused by certain distortions. The proposed method is experimented on common benchmark image quality assessment (IQA) databases, including LIVE2008, TID2013, and KADID-10k. To generalize the performance with a larger database, we distorted images using four general distortion types: Gaussian noise, Gaussian blur, JPEG compression, and contrast degradation, applied to the ImageNet-1k database. The experimental results demonstrate that the proposed method outperforms other solutions in terms of accuracy